[Bug] Vulkan Backend Produces Distorted / Gibrish Images with --vae-conv-direct in Higher Resolutions

Git commit

stable-diffusion.cpp version master-709-92a3b73, commit 92a3b73

Operating System & Version

Ubuntu 24.04

GGML backends

Vulkan

Command-line arguments used

./sd-cli --mode img_gen --diffusion-model /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf --vae /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors --llm /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf --lora-model-dir /home/tipu/AI/image/loras/ --lora-apply-mode immediately --cfg-scale 1 -H 1024 -W 1024 --steps 7 --seed 765675567 --rng cpu --scheduler smoothstep --sampling-method euler --flow-shift 2 -o output_$(date +%Y-%m-%d_%H-%M-%S).png --fa --vae-conv-direct -p "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere."

Steps to reproduce

First I produced image which was 512*512 with --vae-conv-direct all good.

./sd-cli --mode img_gen --diffusion-model /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf --vae /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors --llm /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf --lora-model-dir /home/tipu/AI/image/loras/ --lora-apply-mode immediately --cfg-scale 1 -H 512 -W 512 --steps 7 --seed 765675567 --rng cpu --scheduler smoothstep --sampling-method euler --flow-shift 2 -o output_$(date +%Y-%m-%d_%H-%M-%S).png --fa --vae-conv-direct -p "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere."

Increased the resolution to 1024*1024 with --vae-conv-direct image becomes all distorted

./sd-cli --mode img_gen --diffusion-model /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf --vae /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors --llm /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf --lora-model-dir /home/tipu/AI/image/loras/ --lora-apply-mode immediately --cfg-scale 1 -H 1024 -W 1024 --steps 7 --seed 765675567 --rng cpu --scheduler smoothstep --sampling-method euler --flow-shift 2 -o output_$(date +%Y-%m-%d_%H-%M-%S).png --fa --vae-conv-direct -p "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere."

Keep the resolution to 1024*1024. Remove the --vae-conv-direct and replace with --vae-on-cpu output image is as required

./sd-cli --mode img_gen --diffusion-model /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf --vae /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors --llm /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf --lora-model-dir /home/tipu/AI/image/loras/ --lora-apply-mode immediately --cfg-scale 1 -H 1024 -W 1024 --steps 7 --seed 765675567 --rng cpu --scheduler smoothstep --sampling-method euler --flow-shift 2 -o output_$(date +%Y-%m-%d_%H-%M-%S).png --fa --vae-on-cpu -p "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere."

What you expected to happen

Image should be normal as it is by replacing --vae-conv-direct and replace with --vae-on-cpu

What actually happened

gibberish / distorted images are produced. As earlier shared in #1666

Logs / error messages / stack trace

1024*1024 generation with distorted image.

logs

bash  ./sd-cli --mode img_gen --diffusion-model /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf --vae /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors --llm /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf --lora-model-dir /home/tipu/AI/image/loras/ --lora-apply-mode immediately --cfg-scale 1 -H 1024 -W 1024 --steps 7 --seed 765675567 --rng cpu --scheduler smoothstep --sampling-method euler --flow-shift 2 -o output_$(date +%Y-%m-%d_%H-%M-%S).png --fa --vae-conv-direct -p "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere."
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon Graphics (RADV RENOIR) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 0 | matrix cores: none
[INFO ] stable-diffusion.cpp:397  - loading diffusion model from '/home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf'
[INFO ] model_loader.cpp:235  - load /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf using gguf format
[INFO ] stable-diffusion.cpp:459  - loading llm from '/home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf'
[INFO ] model_loader.cpp:235  - load /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf using gguf format
[INFO ] stable-diffusion.cpp:473  - loading vae from '/home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors'
[INFO ] model_loader.cpp:238  - load /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors using safetensors format
[INFO ] stable-diffusion.cpp:524  - Version: Z-Image 
[INFO ] stable-diffusion.cpp:551  - Weight type stat:                      f32: 634  |    q8_0: 433  |    bf16: 28   
[INFO ] stable-diffusion.cpp:552  - Conditioner weight type stat:          f32: 145  |    q8_0: 253  
[INFO ] stable-diffusion.cpp:553  - Diffusion model weight type stat:      f32: 245  |    q8_0: 180  |    bf16: 28   
[INFO ] stable-diffusion.cpp:554  - VAE weight type stat:                  f32: 244  
[INFO ] stable-diffusion.cpp:955  - using VAE for encoding / decoding
[INFO ] auto_encoder_kl.hpp:527  - vae decoder: ch = 128
[INFO ] stable-diffusion.cpp:991  - Using Conv2d direct in the vae model
[INFO ] stable-diffusion.cpp:1059 - Using flash attention
[INFO ] stable-diffusion.cpp:1073 - Using flash attention in the diffusion model
[INFO ] stable-diffusion.cpp:1181 - total params memory size = 11131.89MB (VRAM 11131.89MB, RAM 0.00MB): text_encoders 4076.43MB(VRAM), diffusion_model 6891.46MB(VRAM), vae 164.00MB(VRAM), controlnet 0.00MB(N/A), extensions 0.00MB(N/A)
[INFO ] stable-diffusion.cpp:1281 - running in FLOW mode
[INFO ] stable-diffusion.cpp:4432 - generate_image 1024x1024
[INFO ] denoiser.hpp:607  - get_sigmas with SmoothStep scheduler
[INFO ] stable-diffusion.cpp:3456 - sampling using Euler method
  |##################################################| 386/386 - 2.42GB/s
[INFO ] model_loader.cpp:1224 - loading tensors completed, taking 1.60s (read: 0.15s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 1.17s)
[INFO ] stable-diffusion.cpp:4171 - get_learned_condition completed, taking 2.77s
[INFO ] stable-diffusion.cpp:4478 - generating image: 1/1 - seed 765675567
  |##################################################| 452/452 - 3.04GB/s
[INFO ] model_loader.cpp:1224 - loading tensors completed, taking 2.21s (read: 0.29s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 1.54s)
  |==================================================| 7/7 - 64.16s/it
[INFO ] stable-diffusion.cpp:4510 - sampling completed, taking 440.58s
[INFO ] stable-diffusion.cpp:4522 - generating 1 latent images completed, taking 440.58s
[INFO ] stable-diffusion.cpp:4195 - decoding 1 latents
  |##################################################| 138/138 - 940.30MB/s
[INFO ] model_loader.cpp:1224 - loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.05s, copy_to_backend: 0.10s)
[INFO ] stable-diffusion.cpp:4215 - latent 1 decoded, taking 25.17s
[INFO ] stable-diffusion.cpp:4219 - decode_first_stage completed, taking 25.17s
[INFO ] stable-diffusion.cpp:4659 - generate_image completed in 468.55s
[INFO ] main.cpp:463  - save result image 0 to 'output_2026-06-17_10-20-18.png' (success)
[INFO ] main.cpp:535  - 1/1 images saved

1024*1024 Working image

Details

 ./sd-cli --mode img_gen --diffusion-model /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf --vae /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors --llm /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf --lora-model-dir /home/tipu/AI/image/loras/ --lora-apply-mode immediately --cfg-scale 1 -H 1024 -W 1024 --steps 7 --seed 765675567 --rng cpu --scheduler smoothstep --sampling-method euler --flow-shift 2 -o output_$(date +%Y-%m-%d_%H-%M-%S).png --fa --vae-on-cpu -p "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere."
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon Graphics (RADV RENOIR) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 0 | matrix cores: none
[INFO ] stable-diffusion.cpp:397  - loading diffusion model from '/home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf'
[INFO ] model_loader.cpp:235  - load /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf using gguf format
[INFO ] stable-diffusion.cpp:459  - loading llm from '/home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf'
[INFO ] model_loader.cpp:235  - load /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf using gguf format
[INFO ] stable-diffusion.cpp:473  - loading vae from '/home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors'
[INFO ] model_loader.cpp:238  - load /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors using safetensors format
[INFO ] stable-diffusion.cpp:524  - Version: Z-Image 
[INFO ] stable-diffusion.cpp:551  - Weight type stat:                      f32: 634  |    q8_0: 433  |    bf16: 28   
[INFO ] stable-diffusion.cpp:552  - Conditioner weight type stat:          f32: 145  |    q8_0: 253  
[INFO ] stable-diffusion.cpp:553  - Diffusion model weight type stat:      f32: 245  |    q8_0: 180  |    bf16: 28   
[INFO ] stable-diffusion.cpp:554  - VAE weight type stat:                  f32: 244  
[INFO ] stable-diffusion.cpp:955  - using VAE for encoding / decoding
[INFO ] auto_encoder_kl.hpp:527  - vae decoder: ch = 128
[INFO ] stable-diffusion.cpp:1059 - Using flash attention
[INFO ] stable-diffusion.cpp:1073 - Using flash attention in the diffusion model
[INFO ] stable-diffusion.cpp:1181 - total params memory size = 11131.89MB (VRAM 10967.89MB, RAM 164.00MB): text_encoders 4076.43MB(VRAM), diffusion_model 6891.46MB(VRAM), vae 164.00MB(RAM), controlnet 0.00MB(N/A), extensions 0.00MB(N/A)
[INFO ] stable-diffusion.cpp:1281 - running in FLOW mode
[INFO ] stable-diffusion.cpp:4432 - generate_image 1024x1024
[INFO ] denoiser.hpp:607  - get_sigmas with SmoothStep scheduler
[INFO ] stable-diffusion.cpp:3456 - sampling using Euler method
  |##################################################| 386/386 - 2.76GB/s
[INFO ] model_loader.cpp:1224 - loading tensors completed, taking 1.41s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 1.07s)
[INFO ] stable-diffusion.cpp:4171 - get_learned_condition completed, taking 2.52s
[INFO ] stable-diffusion.cpp:4478 - generating image: 1/1 - seed 765675567
  |##################################################| 452/452 - 3.35GB/s
[INFO ] model_loader.cpp:1224 - loading tensors completed, taking 2.01s (read: 0.25s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 1.48s)
  |==================================================| 7/7 - 61.84s/it
[INFO ] stable-diffusion.cpp:4510 - sampling completed, taking 428.98s
[INFO ] stable-diffusion.cpp:4522 - generating 1 latent images completed, taking 428.98s
[INFO ] stable-diffusion.cpp:4195 - decoding 1 latents
  |##################################################| 138/138 - 866.98MB/s
[INFO ] model_loader.cpp:1224 - loading tensors completed, taking 0.22s (read: 0.02s, memcpy: 0.00s, convert: 0.11s, copy_to_backend: 0.00s)
[INFO ] stable-diffusion.cpp:4215 - latent 1 decoded, taking 64.52s
[INFO ] stable-diffusion.cpp:4219 - decode_first_stage completed, taking 64.52s
[INFO ] stable-diffusion.cpp:4659 - generate_image completed in 496.04s
[INFO ] main.cpp:463  - save result image 0 to 'output_2026-06-17_10-29-24.png' (success)
[INFO ] main.cpp:535  - 1/1 images saved

### Additional context / environment details

No response

You're using a modified encoder and VAE. That in itself could be the problem.

As far as I know, FA doesn't work well with Vega iGPUs. It just slows everything down. Disable it.

Also, don't use a fixed seed or GGUF Q8, the former causes bugs, and using a quantization higher than Q5 will push you up against the driver's limits, which can only allocate a maximum of about ~4GB.

You're using a modified encoder and VAE. That in itself could be the problem.

It was just fine couple of months back, I tested this combination before saving. Actually the image shared was same that I generated couple of months back and recorded cli command that is how I found out that a problem has happened now.
Earlier some images was generated by me and shared in #1666
--vae-conv-direct was working just fine.

As far as I know, FA doesn't work well with Vega iGPUs. It just slows everything down. Disable it.

It is iGPU and FA on it works just fine on llama.cpp that I have using for some time. I can generate upto 19 tps TG and 130 tps PP for Gemma 4 QAT through vulkan. I use Qwen3.6 and other models just fine through llama.cpp. Yes anything above A3B is slow though but this is the only option I have.

Also, don't use a fixed seed or GGUF Q8, the former causes bugs, and using a quantization higher than Q5 will push you up against the driver's limits, which can only allocate a maximum of about ~4GB.

I used fixed seed to compare with image that I generated some months back and recorded command for later. Hence now there is one to one comparison that it was working in past but not working now. As said earlier I downloaded is GGUF Q8 quite some time back and it was working fine.

I understand this driver's limits, which can only allocate a maximum of about ~4GB. In llama.cpp Jeff programmed something so that if 4GB plus buffer is required it falls back to CPU. To avoid that I was using --vae-conv-direct and now with -vae-on-cpu it is still working for generation.

If that limit is hit program should give OOM error rather then producing faulty images.

@GreenShadows

Also, don't use a fixed seed or GGUF Q8, the former causes bugs

@GreenShadows , could you point out an issue related to fixed seeds?

Also, current sd.cpp should be able to deal with the 4G allocation limit just fine, as long as the inference code doesn't have any huge individual op (e.g. non-tiled VAE for high resolutions).

And this is likely hardware dependent, since I use --vae-conv-direct with my card all the time (RX 7600 XT, radv driver). @engrtipusultan, could you try changing resolutions little by little, to find out what is the largest you can generate without the distortion? Also, does --params-backend cpu changes anything?

Using a fixed seed causes artifacts for me in some distilled models. But it's kinda random.

I have a laptop with this iGPU uArch, and I tested it: with a quantization level higher than Q5, it gets super slow and causes allocation errors in Vulkan.

Maybe this will improve a little now that AMD is working on adding support for Vega in ROCm. ROCm/rocm-libraries#5282

And this is likely hardware dependent, since I use --vae-conv-direct with my card all the time (RX 7600 XT, radv driver). @engrtipusultan, could you try changing resolutions little by little, to find out what is the largest you can generate without the distortion? Also, does --params-backend cpu changes anything?

I added --params-backend cpu it did not make any difference. I could not understand what it does?

I started increase resolution with a step of 64 from 512*512 so image is fine until 960 × 960 pixels on 1024 × 1024 pixels and 1088 × 1088 pixels image is garbage as shared above following are their logs.

960x960

bash  ./sd-cli --mode img_gen --diffusion-model /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf --vae /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors --llm /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf --lora-model-dir /home/tipu/AI/image/loras/ --lora-apply-mode immediately --cfg-scale 1 --steps 7 --seed 765675567 --rng cpu --scheduler smoothstep --sampling-method euler --flow-shift 2 -o output_$(date +%Y-%m-%d_%H-%M-%S).png --fa --vae-conv-direct -p "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere." -v -H 960 -W 960
[DEBUG] main.cpp:598  - version: stable-diffusion.cpp version master-709-92a3b73, commit 92a3b73
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon Graphics (RADV RENOIR) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 0 | matrix cores: none
[DEBUG] main.cpp:599  - System Info: 
   SSE3 = 1 |    SSSE3 = 1 |    AVX = 1 |    AVX2 = 1 |    F16C = 1 |    FMA = 1 |    BMI2 = 1 |    OPENMP = 1 |    REPACK = 1 | 
[DEBUG] main.cpp:600  - SDCliParams {
  mode: img_gen,
  output_path: "output_2026-06-18_10-18-48.png",
  image_path: "",
  metadata_format: "text",
  verbose: true,
  color: false,
  canny_preprocess: false,
  convert_name: false,
  preview_method: none,
  preview_interval: 1,
  preview_path: "preview.png",
  preview_fps: 16,
  taesd_preview: false,
  preview_noisy: false,
  metadata_raw: false,
  metadata_brief: false,
  metadata_all: false
}
[DEBUG] main.cpp:601  - SDContextParams {
  n_threads: 8,
  model_path: "",
  clip_l_path: "",
  clip_g_path: "",
  clip_vision_path: "",
  t5xxl_path: "",
  llm_path: "/home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf",
  llm_vision_path: "",
  diffusion_model_path: "/home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf",
  high_noise_diffusion_model_path: "",
  uncond_diffusion_model_path: "",
  embeddings_connectors_path: "",
  vae_path: "/home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors",
  vae_format: "auto",
  audio_vae_path: "",
  taesd_path: "",
  esrgan_path: "",
  control_net_path: "",
  embedding_dir: "",
  embeddings: {
  }
  wtype: NONE,
  tensor_type_rules: "",
  lora_model_dir: "/home/tipu/AI/image/loras/",
  hires_upscalers_dir: "",
  photo_maker_path: "",
  rng_type: cpu,
  sampler_rng_type: NONE,
  offload_params_to_cpu: false,
  max_vram: "0",
  stream_layers: false,
  backend: "",
  params_backend: "",
  enable_mmap: false,
  control_net_cpu: false,
  clip_on_cpu: false,
  vae_on_cpu: false,
  flash_attn: true,
  diffusion_flash_attn: false,
  diffusion_conv_direct: false,
  vae_conv_direct: true,
  circular: false,
  circular_x: false,
  circular_y: false,
  chroma_use_dit_mask: true,
  qwen_image_zero_cond_t: false,
  chroma_use_t5_mask: false,
  chroma_t5_mask_pad: 1,
  prediction: NONE,
  lora_apply_mode: immediately,
  force_sdxl_vae_conv_scale: false
}
[DEBUG] main.cpp:602  - SDGenerationParams {
  loras: "{
  }",
  high_noise_loras: "{
  }",
  prompt: "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere.",
  negative_prompt: "",
  clip_skip: -1,
  width: 960,
  height: 960,
  batch_count: 1,
  init_image_path: "",
  end_image_path: "",
  mask_image_path: "",
  control_image_path: "",
  ref_image_paths: [],
  control_video_path: "",
  auto_resize_ref_image: true,
  increase_ref_index: false,
  pm_id_images_dir: "",
  pm_id_embed_path: "",
  pm_style_strength: 20,
  skip_layers: [7, 8, 9],
  sample_params: (txt_cfg: 1.00, img_cfg: 1.00, distilled_guidance: 3.50, slg.layer_count: 0, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: smoothstep, sample_method: euler, sample_steps: 7, eta: inf, shifted_timestep: 0, flow_shift: 2.00, extra_sample_args: ),
  high_noise_skip_layers: [7, 8, 9],
  high_noise_sample_params: (txt_cfg: 7.00, img_cfg: 7.00, distilled_guidance: 3.50, slg.layer_count: 0, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: NONE, sample_steps: 20, eta: inf, shifted_timestep: 0, flow_shift: inf, extra_sample_args: ),
  custom_sigmas: [],
  cache_mode: "",
  cache_option: "",
  cache: disabled (threshold=inf, start=0.15, end=0.95),
  moe_boundary: 0.875,
  video_frames: 1,
  fps: 16,
  vace_strength: 1,
  strength: 0.75,
  control_strength: 0.9,
  seed: 765675567,
  upscale_repeats: 1,
  upscale_tile_size: 128,
  hires: { enabled: false, upscaler: "Latent", model_path: "", scale: 2, target_width: 0, target_height: 0, steps: 0, denoising_strength: 0.7, custom_sigmas: [], upscale_tile_size: 128 },
  vae_tiling_params: { 0, 0, 0, 0, 0.5, 0, 0, "" },
}
[DEBUG] ggml_extend_backend.cpp:401  - Found 3 backend devices:
[DEBUG] ggml_extend_backend.cpp:404  - #0: Vulkan0
[DEBUG] ggml_extend_backend.cpp:404  - #1: BLAS
[DEBUG] ggml_extend_backend.cpp:404  - #2: CPU
[DEBUG] ggml_extend_backend.cpp:326  - Initializing backend: Vulkan0
[DEBUG] model_loader.cpp:227  - using 8 threads for model loading
[INFO ] stable-diffusion.cpp:397  - loading diffusion model from '/home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf'
[INFO ] model_loader.cpp:235  - load /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf using gguf format
[DEBUG] model_loader.cpp:284  - init from '/home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf'
[INFO ] stable-diffusion.cpp:459  - loading llm from '/home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf'
[INFO ] model_loader.cpp:235  - load /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf using gguf format
[DEBUG] model_loader.cpp:284  - init from '/home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf'
[INFO ] stable-diffusion.cpp:473  - loading vae from '/home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors'
[INFO ] model_loader.cpp:238  - load /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors using safetensors format
[DEBUG] model_loader.cpp:312  - init from '/home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors', prefix = 'vae.'
[INFO ] stable-diffusion.cpp:524  - Version: Z-Image 
[INFO ] stable-diffusion.cpp:551  - Weight type stat:                      f32: 634  |    q8_0: 433  |    bf16: 28   
[INFO ] stable-diffusion.cpp:552  - Conditioner weight type stat:          f32: 145  |    q8_0: 253  
[INFO ] stable-diffusion.cpp:553  - Diffusion model weight type stat:      f32: 245  |    q8_0: 180  |    bf16: 28   
[INFO ] stable-diffusion.cpp:554  - VAE weight type stat:                  f32: 244  
[DEBUG] stable-diffusion.cpp:556  - ggml tensor size = 400 bytes
[DEBUG] qwen2_tokenizer.cpp:14   - merges size 151387
[DEBUG] qwen2_tokenizer.cpp:39   - vocab size: 151674
[DEBUG] llm.hpp:222  - llm: num_layers = 36, vocab_size = 151936, hidden_size = 2560, intermediate_size = 9728
[DEBUG] z_image.hpp:110  - z_image: num_layers = 30, num_refiner_layers = 2, hidden_size = 3840, num_heads = 30, num_kv_heads = 30, in_channels = 16, out_channels = 16
[INFO ] stable-diffusion.cpp:955  - using VAE for encoding / decoding
[INFO ] auto_encoder_kl.hpp:527  - vae decoder: ch = 128
[INFO ] stable-diffusion.cpp:991  - Using Conv2d direct in the vae model
[INFO ] stable-diffusion.cpp:1059 - Using flash attention
[INFO ] stable-diffusion.cpp:1073 - Using flash attention in the diffusion model
[DEBUG] stable-diffusion.cpp:1091 - validating model metadata
[DEBUG] stable-diffusion.cpp:1141 - model metadata validated; weights will be prepared lazily
[INFO ] stable-diffusion.cpp:1181 - total params memory size = 11131.89MB (VRAM 11131.89MB, RAM 0.00MB): text_encoders 4076.43MB(VRAM), diffusion_model 6891.46MB(VRAM), vae 164.00MB(VRAM), controlnet 0.00MB(N/A), extensions 0.00MB(N/A)
[INFO ] stable-diffusion.cpp:1281 - running in FLOW mode
[INFO ] stable-diffusion.cpp:4432 - generate_image 960x960
[INFO ] denoiser.hpp:607  - get_sigmas with SmoothStep scheduler
[INFO ] stable-diffusion.cpp:3456 - sampling using Euler method
[DEBUG] conditioner.hpp:1607 - parse '<|im_start|>user
A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere.<|im_end|>
<|im_start|>assistant
' to [['<|im_start|>user
', 1], ['A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere.', 1], ['<|im_end|>
<|im_start|>assistant
', 1], ]
[DEBUG] bpe_tokenizer.cpp:207  - split prompt "<|im_start|>user
" to tokens ["<|im_start|>", "user", "Ċ", ]
[DEBUG] bpe_tokenizer.cpp:207  - split prompt "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere." to tokens ["A", "Ġcozy", "Ġstudy", "Ġwith", "Ġwalls", "Ġcovered", "Ġin", "Ġflowing", "Ġink", "Ġsketches", "Ġand", "Ġsoft", "Ġwater", "color", "Ġwash", "es", ".", "ĠAt", "Ġthe", "Ġcenter", ",", "Ġa", "Ġcircular", "Ġtear", "Ġin", "Ġthe", "Ġpaper", "Ġwall", "Ġopens", "Ġupward", "Ġlike", "Ġa", "Ġmagical", "Ġportal", ",", "Ġrevealing", "Ġa", "Ġlush", "Ġ", "3", "D", "Ġme", "adow", "Ġbath", "ed", "Ġin", "Ġgentle", "Ġmorning", "Ġsunlight", ".", "ĠThe", "Ġcurled", "Ġedge", "Ġof", "Ġthe", "Ġtorn", "Ġpaper", "Ġcasts", "Ġsoft", "Ġshadows", "Ġonto", "Ġthe", "Ġdrawn", "Ġwalls", ".", "ĠA", "Ġplayful", "Ġfox", "Ġleaps", "Ġout", "Ġfrom", "Ġthe", "Ġportal", ":", "Ġits", "Ġfront", "Ġlegs", "Ġand", "Ġhead", "Ġfully", "Ġin", "Ġthe", "Ġ", "3", "D", "Ġme", "adow", ",", "Ġmid", "-air", "Ġbody", "Ġpartially", "Ġtransitioning", "Ġinto", "Ġsketch", "Ġlines", ",", "Ġtail", "Ġstill", "Ġa", "Ġdelicate", "Ġink", "Ġdrawing", ".", "ĠEveryday", "Ġobjects", "Ġin", "Ġthe", "Ġroom", "âĢĶ", "like", "Ġa", "Ġdrawn", "Ġte", "ac", "up", "Ġand", "Ġfloating", "Ġbooks", "âĢĶ", "cast", "Ġsubtle", "Ġrealistic", "Ġshadows", ",", "Ġblending", "Ġthe", "Ġdimensions", ".", "ĠSoft", "Ġpast", "el", "Ġpalette", ",", "Ġvisible", "Ġpaper", "Ġtexture", ",", "Ġhigh", "Ġdetail", ",", "Ġwhims", "ical", "Ġyet", "Ġbelievable", "Ġ'", "2", "D", "-to", "-", "3", "D", "Ġtransition", "'", "Ġrealism", ",", "Ġdynamic", "Ġmotion", ",", "Ġenchant", "ing", "Ġand", "Ġtactile", "Ġatmosphere", ".", ]
[DEBUG] bpe_tokenizer.cpp:207  - split prompt "<|im_end|>
<|im_start|>assistant
" to tokens ["<|im_end|>", "Ċ", "<|im_start|>", "assistant", "Ċ", ]
[DEBUG] model_loader.cpp:987  - loading 386/398 tensors from /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf
  |##################################################| 386/386 - 2.75GB/s
[INFO ] model_loader.cpp:1224 - loading tensors completed, taking 1.41s (read: 0.12s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 1.07s)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1020.00 MB, 64 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1020.00 MB, 64 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1022.57 MB, 96 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (911.57 MB, 162 tensors, VRAM)
[DEBUG] ggml_extend.hpp:1995 - qwen3 compute buffer size: 19.04 MB(VRAM)
[DEBUG] conditioner.hpp:2021 - computing condition graph completed, taking 2526 ms
[INFO ] stable-diffusion.cpp:4171 - get_learned_condition completed, taking 2.53s
[INFO ] stable-diffusion.cpp:4478 - generating image: 1/1 - seed 765675567
[DEBUG] model_loader.cpp:987  - loading 453/453 tensors from /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf
  |##################################################| 453/453 - 3.33GB/s
[INFO ] model_loader.cpp:1224 - loading tensors completed, taking 2.02s (read: 0.25s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 1.49s)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1016.95 MB, 30 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1016.02 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (991.11 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1021.00 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1011.49 MB, 33 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1002.11 MB, 136 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (832.78 MB, 167 tensors, VRAM)
[DEBUG] ggml_extend.hpp:1995 - z_image compute buffer size: 619.88 MB(VRAM)
  |==================================================| 7/7 - 54.30s/it
[INFO ] stable-diffusion.cpp:4510 - sampling completed, taking 378.19s
[INFO ] stable-diffusion.cpp:4522 - generating 1 latent images completed, taking 378.19s
[INFO ] stable-diffusion.cpp:4195 - decoding 1 latents
[DEBUG] model_loader.cpp:987  - loading 138/244 tensors from /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors
  |##################################################| 138/138 - 940.30MB/s
[INFO ] model_loader.cpp:1224 - loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.04s, copy_to_backend: 0.10s)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer ( 96.57 MB, 138 tensors, VRAM)
[DEBUG] ggml_extend.hpp:1995 - vae compute buffer size: 2363.38 MB(VRAM)
[DEBUG] vae.hpp:215  - computing vae decode graph completed, taking 17.92s
[INFO ] stable-diffusion.cpp:4215 - latent 1 decoded, taking 17.92s
[INFO ] stable-diffusion.cpp:4219 - decode_first_stage completed, taking 17.92s
[INFO ] stable-diffusion.cpp:4659 - generate_image completed in 398.66s
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1020.00 MB, 64 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1020.00 MB, 64 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1022.57 MB, 96 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (911.57 MB, 162 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1016.95 MB, 30 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1016.02 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (991.11 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1021.00 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1011.49 MB, 33 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1002.11 MB, 136 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (832.78 MB, 167 tensors, VRAM)

1024x1024

bash  ./sd-cli --mode img_gen --diffusion-model /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf --vae /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors --llm /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf --lora-model-dir /home/tipu/AI/image/loras/ --lora-apply-mode immediately --cfg-scale 1 --steps 7 --seed 765675567 --rng cpu --scheduler smoothstep --sampling-method euler --flow-shift 2 -o output_$(date +%Y-%m-%d_%H-%M-%S).png --fa --vae-conv-direct -p "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere." -v -H 1024 -W 1024
[DEBUG] main.cpp:598  - version: stable-diffusion.cpp version master-709-92a3b73, commit 92a3b73
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon Graphics (RADV RENOIR) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 0 | matrix cores: none
[DEBUG] main.cpp:599  - System Info: 
   SSE3 = 1 |    SSSE3 = 1 |    AVX = 1 |    AVX2 = 1 |    F16C = 1 |    FMA = 1 |    BMI2 = 1 |    OPENMP = 1 |    REPACK = 1 | 
[DEBUG] main.cpp:600  - SDCliParams {
  mode: img_gen,
  output_path: "output_2026-06-18_10-27-27.png",
  image_path: "",
  metadata_format: "text",
  verbose: true,
  color: false,
  canny_preprocess: false,
  convert_name: false,
  preview_method: none,
  preview_interval: 1,
  preview_path: "preview.png",
  preview_fps: 16,
  taesd_preview: false,
  preview_noisy: false,
  metadata_raw: false,
  metadata_brief: false,
  metadata_all: false
}
[DEBUG] main.cpp:601  - SDContextParams {
  n_threads: 8,
  model_path: "",
  clip_l_path: "",
  clip_g_path: "",
  clip_vision_path: "",
  t5xxl_path: "",
  llm_path: "/home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf",
  llm_vision_path: "",
  diffusion_model_path: "/home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf",
  high_noise_diffusion_model_path: "",
  uncond_diffusion_model_path: "",
  embeddings_connectors_path: "",
  vae_path: "/home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors",
  vae_format: "auto",
  audio_vae_path: "",
  taesd_path: "",
  esrgan_path: "",
  control_net_path: "",
  embedding_dir: "",
  embeddings: {
  }
  wtype: NONE,
  tensor_type_rules: "",
  lora_model_dir: "/home/tipu/AI/image/loras/",
  hires_upscalers_dir: "",
  photo_maker_path: "",
  rng_type: cpu,
  sampler_rng_type: NONE,
  offload_params_to_cpu: false,
  max_vram: "0",
  stream_layers: false,
  backend: "",
  params_backend: "",
  enable_mmap: false,
  control_net_cpu: false,
  clip_on_cpu: false,
  vae_on_cpu: false,
  flash_attn: true,
  diffusion_flash_attn: false,
  diffusion_conv_direct: false,
  vae_conv_direct: true,
  circular: false,
  circular_x: false,
  circular_y: false,
  chroma_use_dit_mask: true,
  qwen_image_zero_cond_t: false,
  chroma_use_t5_mask: false,
  chroma_t5_mask_pad: 1,
  prediction: NONE,
  lora_apply_mode: immediately,
  force_sdxl_vae_conv_scale: false
}
[DEBUG] main.cpp:602  - SDGenerationParams {
  loras: "{
  }",
  high_noise_loras: "{
  }",
  prompt: "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere.",
  negative_prompt: "",
  clip_skip: -1,
  width: 1024,
  height: 1024,
  batch_count: 1,
  init_image_path: "",
  end_image_path: "",
  mask_image_path: "",
  control_image_path: "",
  ref_image_paths: [],
  control_video_path: "",
  auto_resize_ref_image: true,
  increase_ref_index: false,
  pm_id_images_dir: "",
  pm_id_embed_path: "",
  pm_style_strength: 20,
  skip_layers: [7, 8, 9],
  sample_params: (txt_cfg: 1.00, img_cfg: 1.00, distilled_guidance: 3.50, slg.layer_count: 0, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: smoothstep, sample_method: euler, sample_steps: 7, eta: inf, shifted_timestep: 0, flow_shift: 2.00, extra_sample_args: ),
  high_noise_skip_layers: [7, 8, 9],
  high_noise_sample_params: (txt_cfg: 7.00, img_cfg: 7.00, distilled_guidance: 3.50, slg.layer_count: 0, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: NONE, sample_steps: 20, eta: inf, shifted_timestep: 0, flow_shift: inf, extra_sample_args: ),
  custom_sigmas: [],
  cache_mode: "",
  cache_option: "",
  cache: disabled (threshold=inf, start=0.15, end=0.95),
  moe_boundary: 0.875,
  video_frames: 1,
  fps: 16,
  vace_strength: 1,
  strength: 0.75,
  control_strength: 0.9,
  seed: 765675567,
  upscale_repeats: 1,
  upscale_tile_size: 128,
  hires: { enabled: false, upscaler: "Latent", model_path: "", scale: 2, target_width: 0, target_height: 0, steps: 0, denoising_strength: 0.7, custom_sigmas: [], upscale_tile_size: 128 },
  vae_tiling_params: { 0, 0, 0, 0, 0.5, 0, 0, "" },
}
[DEBUG] ggml_extend_backend.cpp:401  - Found 3 backend devices:
[DEBUG] ggml_extend_backend.cpp:404  - #0: Vulkan0
[DEBUG] ggml_extend_backend.cpp:404  - #1: BLAS
[DEBUG] ggml_extend_backend.cpp:404  - #2: CPU
[DEBUG] ggml_extend_backend.cpp:326  - Initializing backend: Vulkan0
[DEBUG] model_loader.cpp:227  - using 8 threads for model loading
[INFO ] stable-diffusion.cpp:397  - loading diffusion model from '/home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf'
[INFO ] model_loader.cpp:235  - load /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf using gguf format
[DEBUG] model_loader.cpp:284  - init from '/home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf'
[INFO ] stable-diffusion.cpp:459  - loading llm from '/home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf'
[INFO ] model_loader.cpp:235  - load /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf using gguf format
[DEBUG] model_loader.cpp:284  - init from '/home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf'
[INFO ] stable-diffusion.cpp:473  - loading vae from '/home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors'
[INFO ] model_loader.cpp:238  - load /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors using safetensors format
[DEBUG] model_loader.cpp:312  - init from '/home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors', prefix = 'vae.'
[INFO ] stable-diffusion.cpp:524  - Version: Z-Image 
[INFO ] stable-diffusion.cpp:551  - Weight type stat:                      f32: 634  |    q8_0: 433  |    bf16: 28   
[INFO ] stable-diffusion.cpp:552  - Conditioner weight type stat:          f32: 145  |    q8_0: 253  
[INFO ] stable-diffusion.cpp:553  - Diffusion model weight type stat:      f32: 245  |    q8_0: 180  |    bf16: 28   
[INFO ] stable-diffusion.cpp:554  - VAE weight type stat:                  f32: 244  
[DEBUG] stable-diffusion.cpp:556  - ggml tensor size = 400 bytes
[DEBUG] qwen2_tokenizer.cpp:14   - merges size 151387
[DEBUG] qwen2_tokenizer.cpp:39   - vocab size: 151674
[DEBUG] llm.hpp:222  - llm: num_layers = 36, vocab_size = 151936, hidden_size = 2560, intermediate_size = 9728
[DEBUG] z_image.hpp:110  - z_image: num_layers = 30, num_refiner_layers = 2, hidden_size = 3840, num_heads = 30, num_kv_heads = 30, in_channels = 16, out_channels = 16
[INFO ] stable-diffusion.cpp:955  - using VAE for encoding / decoding
[INFO ] auto_encoder_kl.hpp:527  - vae decoder: ch = 128
[INFO ] stable-diffusion.cpp:991  - Using Conv2d direct in the vae model
[INFO ] stable-diffusion.cpp:1059 - Using flash attention
[INFO ] stable-diffusion.cpp:1073 - Using flash attention in the diffusion model
[DEBUG] stable-diffusion.cpp:1091 - validating model metadata
[DEBUG] stable-diffusion.cpp:1141 - model metadata validated; weights will be prepared lazily
[INFO ] stable-diffusion.cpp:1181 - total params memory size = 11131.89MB (VRAM 11131.89MB, RAM 0.00MB): text_encoders 4076.43MB(VRAM), diffusion_model 6891.46MB(VRAM), vae 164.00MB(VRAM), controlnet 0.00MB(N/A), extensions 0.00MB(N/A)
[INFO ] stable-diffusion.cpp:1281 - running in FLOW mode
[INFO ] stable-diffusion.cpp:4432 - generate_image 1024x1024
[INFO ] denoiser.hpp:607  - get_sigmas with SmoothStep scheduler
[INFO ] stable-diffusion.cpp:3456 - sampling using Euler method
[DEBUG] conditioner.hpp:1607 - parse '<|im_start|>user
A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere.<|im_end|>
<|im_start|>assistant
' to [['<|im_start|>user
', 1], ['A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere.', 1], ['<|im_end|>
<|im_start|>assistant
', 1], ]
[DEBUG] bpe_tokenizer.cpp:207  - split prompt "<|im_start|>user
" to tokens ["<|im_start|>", "user", "Ċ", ]
[DEBUG] bpe_tokenizer.cpp:207  - split prompt "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere." to tokens ["A", "Ġcozy", "Ġstudy", "Ġwith", "Ġwalls", "Ġcovered", "Ġin", "Ġflowing", "Ġink", "Ġsketches", "Ġand", "Ġsoft", "Ġwater", "color", "Ġwash", "es", ".", "ĠAt", "Ġthe", "Ġcenter", ",", "Ġa", "Ġcircular", "Ġtear", "Ġin", "Ġthe", "Ġpaper", "Ġwall", "Ġopens", "Ġupward", "Ġlike", "Ġa", "Ġmagical", "Ġportal", ",", "Ġrevealing", "Ġa", "Ġlush", "Ġ", "3", "D", "Ġme", "adow", "Ġbath", "ed", "Ġin", "Ġgentle", "Ġmorning", "Ġsunlight", ".", "ĠThe", "Ġcurled", "Ġedge", "Ġof", "Ġthe", "Ġtorn", "Ġpaper", "Ġcasts", "Ġsoft", "Ġshadows", "Ġonto", "Ġthe", "Ġdrawn", "Ġwalls", ".", "ĠA", "Ġplayful", "Ġfox", "Ġleaps", "Ġout", "Ġfrom", "Ġthe", "Ġportal", ":", "Ġits", "Ġfront", "Ġlegs", "Ġand", "Ġhead", "Ġfully", "Ġin", "Ġthe", "Ġ", "3", "D", "Ġme", "adow", ",", "Ġmid", "-air", "Ġbody", "Ġpartially", "Ġtransitioning", "Ġinto", "Ġsketch", "Ġlines", ",", "Ġtail", "Ġstill", "Ġa", "Ġdelicate", "Ġink", "Ġdrawing", ".", "ĠEveryday", "Ġobjects", "Ġin", "Ġthe", "Ġroom", "âĢĶ", "like", "Ġa", "Ġdrawn", "Ġte", "ac", "up", "Ġand", "Ġfloating", "Ġbooks", "âĢĶ", "cast", "Ġsubtle", "Ġrealistic", "Ġshadows", ",", "Ġblending", "Ġthe", "Ġdimensions", ".", "ĠSoft", "Ġpast", "el", "Ġpalette", ",", "Ġvisible", "Ġpaper", "Ġtexture", ",", "Ġhigh", "Ġdetail", ",", "Ġwhims", "ical", "Ġyet", "Ġbelievable", "Ġ'", "2", "D", "-to", "-", "3", "D", "Ġtransition", "'", "Ġrealism", ",", "Ġdynamic", "Ġmotion", ",", "Ġenchant", "ing", "Ġand", "Ġtactile", "Ġatmosphere", ".", ]
[DEBUG] bpe_tokenizer.cpp:207  - split prompt "<|im_end|>
<|im_start|>assistant
" to tokens ["<|im_end|>", "Ċ", "<|im_start|>", "assistant", "Ċ", ]
[DEBUG] model_loader.cpp:987  - loading 386/398 tensors from /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf
  |##################################################| 386/386 - 2.75GB/s
[INFO ] model_loader.cpp:1224 - loading tensors completed, taking 1.41s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 1.09s)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1020.00 MB, 64 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1020.00 MB, 64 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1022.57 MB, 96 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (911.57 MB, 162 tensors, VRAM)
[DEBUG] ggml_extend.hpp:1995 - qwen3 compute buffer size: 19.04 MB(VRAM)
[DEBUG] conditioner.hpp:2021 - computing condition graph completed, taking 2522 ms
[INFO ] stable-diffusion.cpp:4171 - get_learned_condition completed, taking 2.52s
[INFO ] stable-diffusion.cpp:4478 - generating image: 1/1 - seed 765675567
[DEBUG] model_loader.cpp:987  - loading 452/453 tensors from /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf
  |##################################################| 452/452 - 3.35GB/s
[INFO ] model_loader.cpp:1224 - loading tensors completed, taking 2.01s (read: 0.26s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 1.48s)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1016.95 MB, 30 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1016.02 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (991.11 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1021.00 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1011.49 MB, 33 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1002.10 MB, 135 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (832.78 MB, 167 tensors, VRAM)
[DEBUG] ggml_extend.hpp:1995 - z_image compute buffer size: 697.81 MB(VRAM)
  |==================================================| 7/7 - 62.05s/it
[INFO ] stable-diffusion.cpp:4510 - sampling completed, taking 432.47s
[INFO ] stable-diffusion.cpp:4522 - generating 1 latent images completed, taking 432.47s
[INFO ] stable-diffusion.cpp:4195 - decoding 1 latents
[DEBUG] model_loader.cpp:987  - loading 138/244 tensors from /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors
  |##################################################| 138/138 - 940.30MB/s
[INFO ] model_loader.cpp:1224 - loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.05s, copy_to_backend: 0.08s)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer ( 96.57 MB, 138 tensors, VRAM)
[DEBUG] ggml_extend.hpp:1995 - vae compute buffer size: 2817.00 MB(VRAM)
[DEBUG] vae.hpp:215  - computing vae decode graph completed, taking 25.27s
[INFO ] stable-diffusion.cpp:4215 - latent 1 decoded, taking 25.27s
[INFO ] stable-diffusion.cpp:4219 - decode_first_stage completed, taking 25.27s
[INFO ] stable-diffusion.cpp:4659 - generate_image completed in 460.30s
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1020.00 MB, 64 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1020.00 MB, 64 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1022.57 MB, 96 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (911.57 MB, 162 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1016.95 MB, 30 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1016.02 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (991.11 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1021.00 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1011.49 MB, 33 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1002.10 MB, 135 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (832.78 MB, 167 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer ( 96.57 MB, 138 tensors, VRAM)
[INFO ] main.cpp:463  - save result image 0 to 'output_2026-06-18_10-27-27.png' (success)
[INFO ] main.cpp:535  - 1/1 images saved

1088x1088

 tipu-dev-machine   ~/Applications/sdcpp                                                                                          10:35:09 
 bash  ./sd-cli --mode img_gen --diffusion-model /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf --vae /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors --llm /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf --lora-model-dir /home/tipu/AI/image/loras/ --lora-apply-mode immediately --cfg-scale 1 --steps 7 --seed 765675567 --rng cpu --scheduler smoothstep --sampling-method euler --flow-shift 2 -o output_$(date +%Y-%m-%d_%H-%M-%S).png --fa --vae-conv-direct -p "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere." -v -H 1088 -W 1088
[DEBUG] main.cpp:598  - version: stable-diffusion.cpp version master-709-92a3b73, commit 92a3b73
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon Graphics (RADV RENOIR) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 0 | matrix cores: none
[DEBUG] main.cpp:599  - System Info: 
   SSE3 = 1 |    SSSE3 = 1 |    AVX = 1 |    AVX2 = 1 |    F16C = 1 |    FMA = 1 |    BMI2 = 1 |    OPENMP = 1 |    REPACK = 1 | 
[DEBUG] main.cpp:600  - SDCliParams {
  mode: img_gen,
  output_path: "output_2026-06-18_10-53-03.png",
  image_path: "",
  metadata_format: "text",
  verbose: true,
  color: false,
  canny_preprocess: false,
  convert_name: false,
  preview_method: none,
  preview_interval: 1,
  preview_path: "preview.png",
  preview_fps: 16,
  taesd_preview: false,
  preview_noisy: false,
  metadata_raw: false,
  metadata_brief: false,
  metadata_all: false
}
[DEBUG] main.cpp:601  - SDContextParams {
  n_threads: 8,
  model_path: "",
  clip_l_path: "",
  clip_g_path: "",
  clip_vision_path: "",
  t5xxl_path: "",
  llm_path: "/home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf",
  llm_vision_path: "",
  diffusion_model_path: "/home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf",
  high_noise_diffusion_model_path: "",
  uncond_diffusion_model_path: "",
  embeddings_connectors_path: "",
  vae_path: "/home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors",
  vae_format: "auto",
  audio_vae_path: "",
  taesd_path: "",
  esrgan_path: "",
  control_net_path: "",
  embedding_dir: "",
  embeddings: {
  }
  wtype: NONE,
  tensor_type_rules: "",
  lora_model_dir: "/home/tipu/AI/image/loras/",
  hires_upscalers_dir: "",
  photo_maker_path: "",
  rng_type: cpu,
  sampler_rng_type: NONE,
  offload_params_to_cpu: false,
  max_vram: "0",
  stream_layers: false,
  backend: "",
  params_backend: "",
  enable_mmap: false,
  control_net_cpu: false,
  clip_on_cpu: false,
  vae_on_cpu: false,
  flash_attn: true,
  diffusion_flash_attn: false,
  diffusion_conv_direct: false,
  vae_conv_direct: true,
  circular: false,
  circular_x: false,
  circular_y: false,
  chroma_use_dit_mask: true,
  qwen_image_zero_cond_t: false,
  chroma_use_t5_mask: false,
  chroma_t5_mask_pad: 1,
  prediction: NONE,
  lora_apply_mode: immediately,
  force_sdxl_vae_conv_scale: false
}
[DEBUG] main.cpp:602  - SDGenerationParams {
  loras: "{
  }",
  high_noise_loras: "{
  }",
  prompt: "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere.",
  negative_prompt: "",
  clip_skip: -1,
  width: 1088,
  height: 1088,
  batch_count: 1,
  init_image_path: "",
  end_image_path: "",
  mask_image_path: "",
  control_image_path: "",
  ref_image_paths: [],
  control_video_path: "",
  auto_resize_ref_image: true,
  increase_ref_index: false,
  pm_id_images_dir: "",
  pm_id_embed_path: "",
  pm_style_strength: 20,
  skip_layers: [7, 8, 9],
  sample_params: (txt_cfg: 1.00, img_cfg: 1.00, distilled_guidance: 3.50, slg.layer_count: 0, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: smoothstep, sample_method: euler, sample_steps: 7, eta: inf, shifted_timestep: 0, flow_shift: 2.00, extra_sample_args: ),
  high_noise_skip_layers: [7, 8, 9],
  high_noise_sample_params: (txt_cfg: 7.00, img_cfg: 7.00, distilled_guidance: 3.50, slg.layer_count: 0, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: NONE, sample_steps: 20, eta: inf, shifted_timestep: 0, flow_shift: inf, extra_sample_args: ),
  custom_sigmas: [],
  cache_mode: "",
  cache_option: "",
  cache: disabled (threshold=inf, start=0.15, end=0.95),
  moe_boundary: 0.875,
  video_frames: 1,
  fps: 16,
  vace_strength: 1,
  strength: 0.75,
  control_strength: 0.9,
  seed: 765675567,
  upscale_repeats: 1,
  upscale_tile_size: 128,
  hires: { enabled: false, upscaler: "Latent", model_path: "", scale: 2, target_width: 0, target_height: 0, steps: 0, denoising_strength: 0.7, custom_sigmas: [], upscale_tile_size: 128 },
  vae_tiling_params: { 0, 0, 0, 0, 0.5, 0, 0, "" },
}
[DEBUG] ggml_extend_backend.cpp:401  - Found 3 backend devices:
[DEBUG] ggml_extend_backend.cpp:404  - #0: Vulkan0
[DEBUG] ggml_extend_backend.cpp:404  - #1: BLAS
[DEBUG] ggml_extend_backend.cpp:404  - #2: CPU
[DEBUG] ggml_extend_backend.cpp:326  - Initializing backend: Vulkan0
[DEBUG] model_loader.cpp:227  - using 8 threads for model loading
[INFO ] stable-diffusion.cpp:397  - loading diffusion model from '/home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf'
[INFO ] model_loader.cpp:235  - load /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf using gguf format
[DEBUG] model_loader.cpp:284  - init from '/home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf'
[INFO ] stable-diffusion.cpp:459  - loading llm from '/home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf'
[INFO ] model_loader.cpp:235  - load /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf using gguf format
[DEBUG] model_loader.cpp:284  - init from '/home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf'
[INFO ] stable-diffusion.cpp:473  - loading vae from '/home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors'
[INFO ] model_loader.cpp:238  - load /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors using safetensors format
[DEBUG] model_loader.cpp:312  - init from '/home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors', prefix = 'vae.'
[INFO ] stable-diffusion.cpp:524  - Version: Z-Image 
[INFO ] stable-diffusion.cpp:551  - Weight type stat:                      f32: 634  |    q8_0: 433  |    bf16: 28   
[INFO ] stable-diffusion.cpp:552  - Conditioner weight type stat:          f32: 145  |    q8_0: 253  
[INFO ] stable-diffusion.cpp:553  - Diffusion model weight type stat:      f32: 245  |    q8_0: 180  |    bf16: 28   
[INFO ] stable-diffusion.cpp:554  - VAE weight type stat:                  f32: 244  
[DEBUG] stable-diffusion.cpp:556  - ggml tensor size = 400 bytes
[DEBUG] qwen2_tokenizer.cpp:14   - merges size 151387
[DEBUG] qwen2_tokenizer.cpp:39   - vocab size: 151674
[DEBUG] llm.hpp:222  - llm: num_layers = 36, vocab_size = 151936, hidden_size = 2560, intermediate_size = 9728
[DEBUG] z_image.hpp:110  - z_image: num_layers = 30, num_refiner_layers = 2, hidden_size = 3840, num_heads = 30, num_kv_heads = 30, in_channels = 16, out_channels = 16
[INFO ] stable-diffusion.cpp:955  - using VAE for encoding / decoding
[INFO ] auto_encoder_kl.hpp:527  - vae decoder: ch = 128
[INFO ] stable-diffusion.cpp:991  - Using Conv2d direct in the vae model
[INFO ] stable-diffusion.cpp:1059 - Using flash attention
[INFO ] stable-diffusion.cpp:1073 - Using flash attention in the diffusion model
[DEBUG] stable-diffusion.cpp:1091 - validating model metadata
[DEBUG] stable-diffusion.cpp:1141 - model metadata validated; weights will be prepared lazily
[INFO ] stable-diffusion.cpp:1181 - total params memory size = 11131.89MB (VRAM 11131.89MB, RAM 0.00MB): text_encoders 4076.43MB(VRAM), diffusion_model 6891.46MB(VRAM), vae 164.00MB(VRAM), controlnet 0.00MB(N/A), extensions 0.00MB(N/A)
[INFO ] stable-diffusion.cpp:1281 - running in FLOW mode
[INFO ] stable-diffusion.cpp:4432 - generate_image 1088x1088
[INFO ] denoiser.hpp:607  - get_sigmas with SmoothStep scheduler
[INFO ] stable-diffusion.cpp:3456 - sampling using Euler method
[DEBUG] conditioner.hpp:1607 - parse '<|im_start|>user
A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere.<|im_end|>
<|im_start|>assistant
' to [['<|im_start|>user
', 1], ['A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere.', 1], ['<|im_end|>
<|im_start|>assistant
', 1], ]
[DEBUG] bpe_tokenizer.cpp:207  - split prompt "<|im_start|>user
" to tokens ["<|im_start|>", "user", "Ċ", ]
[DEBUG] bpe_tokenizer.cpp:207  - split prompt "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere." to tokens ["A", "Ġcozy", "Ġstudy", "Ġwith", "Ġwalls", "Ġcovered", "Ġin", "Ġflowing", "Ġink", "Ġsketches", "Ġand", "Ġsoft", "Ġwater", "color", "Ġwash", "es", ".", "ĠAt", "Ġthe", "Ġcenter", ",", "Ġa", "Ġcircular", "Ġtear", "Ġin", "Ġthe", "Ġpaper", "Ġwall", "Ġopens", "Ġupward", "Ġlike", "Ġa", "Ġmagical", "Ġportal", ",", "Ġrevealing", "Ġa", "Ġlush", "Ġ", "3", "D", "Ġme", "adow", "Ġbath", "ed", "Ġin", "Ġgentle", "Ġmorning", "Ġsunlight", ".", "ĠThe", "Ġcurled", "Ġedge", "Ġof", "Ġthe", "Ġtorn", "Ġpaper", "Ġcasts", "Ġsoft", "Ġshadows", "Ġonto", "Ġthe", "Ġdrawn", "Ġwalls", ".", "ĠA", "Ġplayful", "Ġfox", "Ġleaps", "Ġout", "Ġfrom", "Ġthe", "Ġportal", ":", "Ġits", "Ġfront", "Ġlegs", "Ġand", "Ġhead", "Ġfully", "Ġin", "Ġthe", "Ġ", "3", "D", "Ġme", "adow", ",", "Ġmid", "-air", "Ġbody", "Ġpartially", "Ġtransitioning", "Ġinto", "Ġsketch", "Ġlines", ",", "Ġtail", "Ġstill", "Ġa", "Ġdelicate", "Ġink", "Ġdrawing", ".", "ĠEveryday", "Ġobjects", "Ġin", "Ġthe", "Ġroom", "âĢĶ", "like", "Ġa", "Ġdrawn", "Ġte", "ac", "up", "Ġand", "Ġfloating", "Ġbooks", "âĢĶ", "cast", "Ġsubtle", "Ġrealistic", "Ġshadows", ",", "Ġblending", "Ġthe", "Ġdimensions", ".", "ĠSoft", "Ġpast", "el", "Ġpalette", ",", "Ġvisible", "Ġpaper", "Ġtexture", ",", "Ġhigh", "Ġdetail", ",", "Ġwhims", "ical", "Ġyet", "Ġbelievable", "Ġ'", "2", "D", "-to", "-", "3", "D", "Ġtransition", "'", "Ġrealism", ",", "Ġdynamic", "Ġmotion", ",", "Ġenchant", "ing", "Ġand", "Ġtactile", "Ġatmosphere", ".", ]
[DEBUG] bpe_tokenizer.cpp:207  - split prompt "<|im_end|>
<|im_start|>assistant
" to tokens ["<|im_end|>", "Ċ", "<|im_start|>", "assistant", "Ċ", ]
[DEBUG] model_loader.cpp:987  - loading 386/398 tensors from /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf
  |##################################################| 386/386 - 2.76GB/s
[INFO ] model_loader.cpp:1224 - loading tensors completed, taking 1.41s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 1.09s)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1020.00 MB, 64 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1020.00 MB, 64 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1022.57 MB, 96 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (911.57 MB, 162 tensors, VRAM)
[DEBUG] ggml_extend.hpp:1995 - qwen3 compute buffer size: 19.04 MB(VRAM)
[DEBUG] conditioner.hpp:2021 - computing condition graph completed, taking 2518 ms
[INFO ] stable-diffusion.cpp:4171 - get_learned_condition completed, taking 2.52s
[INFO ] stable-diffusion.cpp:4478 - generating image: 1/1 - seed 765675567
[DEBUG] model_loader.cpp:987  - loading 453/453 tensors from /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf
  |##################################################| 453/453 - 3.32GB/s
[INFO ] model_loader.cpp:1224 - loading tensors completed, taking 2.02s (read: 0.25s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 1.49s)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1016.95 MB, 30 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1016.02 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (991.11 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1021.00 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1011.49 MB, 33 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (1002.11 MB, 136 tensors, VRAM)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer (832.78 MB, 167 tensors, VRAM)
[DEBUG] ggml_extend.hpp:1995 - z_image compute buffer size: 786.13 MB(VRAM)
  |==================================================| 7/7 - 71.57s/it
[INFO ] stable-diffusion.cpp:4510 - sampling completed, taking 493.51s
[INFO ] stable-diffusion.cpp:4522 - generating 1 latent images completed, taking 493.51s
[INFO ] stable-diffusion.cpp:4195 - decoding 1 latents
[DEBUG] model_loader.cpp:987  - loading 138/244 tensors from /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors
  |##################################################| 138/138 - 940.30MB/s
[INFO ] model_loader.cpp:1224 - loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.06s, copy_to_backend: 0.08s)
[DEBUG] model_manager.cpp:218  - model manager prepared params backend buffer ( 96.57 MB, 138 tensors, VRAM)
[DEBUG] ggml_extend.hpp:1995 - vae compute buffer size: 3758.13 MB(VRAM)
[DEBUG] vae.hpp:215  - computing vae decode graph completed, taking 26.39s
[INFO ] stable-diffusion.cpp:4215 - latent 1 decoded, taking 26.39s
[INFO ] stable-diffusion.cpp:4219 - decode_first_stage completed, taking 26.39s
[INFO ] stable-diffusion.cpp:4659 - generate_image completed in 522.44s
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1020.00 MB, 64 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1020.00 MB, 64 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1022.57 MB, 96 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (911.57 MB, 162 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1016.95 MB, 30 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1016.02 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (991.11 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1021.00 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1011.49 MB, 33 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (1002.11 MB, 136 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer (832.78 MB, 167 tensors, VRAM)
[DEBUG] model_manager.cpp:759  - model manager releasing params backend buffer ( 96.57 MB, 138 tensors, VRAM)
[INFO ] main.cpp:463  - save result image 0 to 'output_2026-06-18_10-53-03.png' (success)
[INFO ] main.cpp:535  - 1/1 images saved

The garbage image looks similar to the ones mentioned in #1589 . One interesting aspect is that, on some of those examples, --vae-conv-direct avoids the issue, instead of triggering it.

I added --params-backend cpu it did not make any difference. I could not understand what it does?

It keeps the weights in RAM, moving them to VRAM on demand. It was a test to see if total VRAM usage could change the result.

I've reported this as a backend issue on: ggml-org/llama.cpp#24943 .

@engrtipusultan

@engrtipusultan , could you confirm if this issue still happens in the latest release (master-737-3b6c9ca)? I cannot reproduce this anymore on my iGPU.

@engrtipusultan , could you confirm if this issue still happens in the latest release (master-737-3b6c9ca)? I cannot reproduce this anymore on my iGPU.

I am getting coredump now with same command as opening post.

Details

 bash  ./sd-cli --mode img_gen --diffusion-model /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf --vae /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors --llm /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf --lora-model-dir /home/tipu/AI/image/loras/ --lora-apply-mode immediately --cfg-scale 1 -H 1024 -W 1024 --steps 7 --seed 765675567 --rng cpu --scheduler smoothstep --sampling-method euler --flow-shift 2 -o output_$(date +%Y-%m-%d_%H-%M-%S).png --fa --vae-conv-direct -p "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere." -v
[DEBUG] main.cpp:602  - version: stable-diffusion.cpp version master-737-3b6c9ca, commit 3b6c9ca
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon Graphics (RADV RENOIR) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 0 | matrix cores: none
[DEBUG] main.cpp:603  - System Info: 
   SSE3 = 1 |    SSSE3 = 1 |    AVX = 1 |    AVX2 = 1 |    F16C = 1 |    FMA = 1 |    BMI2 = 1 |    OPENMP = 1 |    REPACK = 1 | 
[DEBUG] main.cpp:604  - SDCliParams {
  mode: img_gen,
  output_path: "output_2026-06-30_09-14-02.png",
  image_path: "",
  metadata_format: "text",
  verbose: true,
  color: false,
  canny_preprocess: false,
  convert_name: false,
  preview_method: none,
  preview_interval: 1,
  preview_path: "preview.png",
  preview_fps: 16,
  taesd_preview: false,
  preview_noisy: false,
  metadata_raw: false,
  metadata_brief: false,
  metadata_all: false
}
[DEBUG] main.cpp:605  - SDContextParams {
  n_threads: 8,
  model_path: "",
  clip_l_path: "",
  clip_g_path: "",
  clip_vision_path: "",
  t5xxl_path: "",
  llm_path: "/home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf",
  llm_vision_path: "",
  diffusion_model_path: "/home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf",
  high_noise_diffusion_model_path: "",
  uncond_diffusion_model_path: "",
  embeddings_connectors_path: "",
  vae_path: "/home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors",
  vae_format: "auto",
  audio_vae_path: "",
  taesd_path: "",
  esrgan_path: "",
  control_net_path: "",
  embedding_dir: "",
  embeddings: {
  }
  wtype: NONE,
  tensor_type_rules: "",
  lora_model_dir: "/home/tipu/AI/image/loras/",
  hires_upscalers_dir: "",
  photo_maker_path: "",
  rng_type: cpu,
  sampler_rng_type: NONE,
  offload_params_to_cpu: false,
  max_vram: "0",
  stream_layers: false,
  eager_load: false,
  backend: "",
  params_backend: "",
  enable_mmap: false,
  control_net_cpu: false,
  clip_on_cpu: false,
  vae_on_cpu: false,
  flash_attn: true,
  diffusion_flash_attn: false,
  diffusion_conv_direct: false,
  vae_conv_direct: true,
  circular: false,
  circular_x: false,
  circular_y: false,
  chroma_use_dit_mask: true,
  qwen_image_zero_cond_t: false,
  chroma_use_t5_mask: false,
  chroma_t5_mask_pad: 1,
  prediction: NONE,
  lora_apply_mode: immediately,
  force_sdxl_vae_conv_scale: false
}
[DEBUG] main.cpp:606  - SDGenerationParams {
  loras: "{
  }",
  high_noise_loras: "{
  }",
  prompt: "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere.",
  negative_prompt: "",
  clip_skip: -1,
  width: 1024,
  height: 1024,
  batch_count: 1,
  init_image_path: "",
  end_image_path: "",
  mask_image_path: "",
  control_image_path: "",
  ref_image_paths: [],
  control_video_path: "",
  auto_resize_ref_image: true,
  increase_ref_index: false,
  pm_id_images_dir: "",
  pm_id_embed_path: "",
  pm_style_strength: 20,
  skip_layers: [7, 8, 9],
  sample_params: (txt_cfg: 1.00, img_cfg: 1.00, distilled_guidance: 3.50, slg.layer_count: 0, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: smoothstep, sample_method: euler, sample_steps: 7, eta: inf, shifted_timestep: 0, flow_shift: 2.00, extra_sample_args: ),
  high_noise_skip_layers: [7, 8, 9],
  high_noise_sample_params: (txt_cfg: 7.00, img_cfg: 7.00, distilled_guidance: 3.50, slg.layer_count: 0, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: NONE, sample_steps: 20, eta: inf, shifted_timestep: 0, flow_shift: inf, extra_sample_args: ),
  custom_sigmas: [],
  cache_mode: "",
  cache_option: "",
  cache: disabled (threshold=inf, start=0.15, end=0.95),
  moe_boundary: 0.875,
  video_frames: 1,
  fps: 16,
  vace_strength: 1,
  strength: 0.75,
  control_strength: 0.9,
  seed: 765675567,
  upscale_repeats: 1,
  upscale_tile_size: 128,
  hires: { enabled: false, upscaler: "Latent", model_path: "", scale: 2, target_width: 0, target_height: 0, steps: 0, denoising_strength: 0.7, custom_sigmas: [], upscale_tile_size: 128 },
  vae_tiling_params: { 0, 0, 0, 0, 0.5, 0, 0, "" },
}
[DEBUG] ggml_extend_backend.cpp:401  - Found 3 backend devices:
[DEBUG] ggml_extend_backend.cpp:404  - #0: Vulkan0
[DEBUG] ggml_extend_backend.cpp:404  - #1: BLAS
[DEBUG] ggml_extend_backend.cpp:404  - #2: CPU
[DEBUG] ggml_extend_backend.cpp:326  - Initializing backend: Vulkan0
[DEBUG] model_loader.cpp:226  - using 8 threads for model loading
[INFO ] stable-diffusion.cpp:405  - loading diffusion model from '/home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf'
[INFO ] model_loader.cpp:234  - load /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf using gguf format
[DEBUG] model_loader.cpp:283  - init from '/home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf'
[INFO ] stable-diffusion.cpp:467  - loading llm from '/home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf'
[INFO ] model_loader.cpp:234  - load /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf using gguf format
[DEBUG] model_loader.cpp:283  - init from '/home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf'
[INFO ] stable-diffusion.cpp:481  - loading vae from '/home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors'
[INFO ] model_loader.cpp:237  - load /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors using safetensors format
[DEBUG] model_loader.cpp:311  - init from '/home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors', prefix = 'vae.'
[INFO ] stable-diffusion.cpp:532  - Version: Z-Image 
[INFO ] stable-diffusion.cpp:558  - Weight type stat:                      f32: 634  |    q8_0: 433  |    bf16: 28   
[INFO ] stable-diffusion.cpp:559  - Conditioner weight type stat:          f32: 145  |    q8_0: 253  
[INFO ] stable-diffusion.cpp:560  - Diffusion model weight type stat:      f32: 245  |    q8_0: 180  |    bf16: 28   
[INFO ] stable-diffusion.cpp:561  - VAE weight type stat:                  f32: 244  
[DEBUG] stable-diffusion.cpp:563  - ggml tensor size = 400 bytes
[DEBUG] qwen2_tokenizer.cpp:14   - merges size 151387
[DEBUG] qwen2_tokenizer.cpp:39   - vocab size: 151674
[DEBUG] llm.hpp:259  - llm: num_layers = 36, vocab_size = 151936, hidden_size = 2560, intermediate_size = 9728
[DEBUG] z_image.hpp:110  - z_image: num_layers = 30, num_refiner_layers = 2, hidden_size = 3840, num_heads = 30, num_kv_heads = 30, in_channels = 16, out_channels = 16
[INFO ] stable-diffusion.cpp:983  - using VAE for encoding / decoding
[INFO ] auto_encoder_kl.hpp:527  - vae decoder: ch = 128
[INFO ] stable-diffusion.cpp:1019 - Using Conv2d direct in the vae model
[INFO ] stable-diffusion.cpp:1087 - Using flash attention
[INFO ] stable-diffusion.cpp:1101 - Using flash attention in the diffusion model
[DEBUG] stable-diffusion.cpp:1119 - validating model metadata
[DEBUG] stable-diffusion.cpp:1176 - model metadata validated; weights will be prepared lazily
[INFO ] stable-diffusion.cpp:1217 - total params memory size = 11131.89MB (VRAM 11131.89MB, RAM 0.00MB): text_encoders 4076.43MB(VRAM), diffusion_model 6891.46MB(VRAM), vae 164.00MB(VRAM), controlnet 0.00MB(N/A), extensions 0.00MB(N/A)
[INFO ] stable-diffusion.cpp:1324 - running in FLOW mode
[INFO ] stable-diffusion.cpp:4530 - generate_image 1024x1024
[INFO ] denoiser.hpp:919  - get_sigmas with SmoothStep scheduler
[INFO ] stable-diffusion.cpp:3554 - sampling using Euler method
[DEBUG] conditioner.hpp:1604 - parse '<|im_start|>user
A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere.<|im_end|>
<|im_start|>assistant
' to [['<|im_start|>user
', 1], ['A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere.', 1], ['<|im_end|>
<|im_start|>assistant
', 1], ]
[DEBUG] bpe_tokenizer.cpp:208  - split prompt "<|im_start|>user
" to tokens ["<|im_start|>", "user", "Ċ", ]
[DEBUG] bpe_tokenizer.cpp:208  - split prompt "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere." to tokens ["A", "Ġcozy", "Ġstudy", "Ġwith", "Ġwalls", "Ġcovered", "Ġin", "Ġflowing", "Ġink", "Ġsketches", "Ġand", "Ġsoft", "Ġwater", "color", "Ġwash", "es", ".", "ĠAt", "Ġthe", "Ġcenter", ",", "Ġa", "Ġcircular", "Ġtear", "Ġin", "Ġthe", "Ġpaper", "Ġwall", "Ġopens", "Ġupward", "Ġlike", "Ġa", "Ġmagical", "Ġportal", ",", "Ġrevealing", "Ġa", "Ġlush", "Ġ", "3", "D", "Ġme", "adow", "Ġbath", "ed", "Ġin", "Ġgentle", "Ġmorning", "Ġsunlight", ".", "ĠThe", "Ġcurled", "Ġedge", "Ġof", "Ġthe", "Ġtorn", "Ġpaper", "Ġcasts", "Ġsoft", "Ġshadows", "Ġonto", "Ġthe", "Ġdrawn", "Ġwalls", ".", "ĠA", "Ġplayful", "Ġfox", "Ġleaps", "Ġout", "Ġfrom", "Ġthe", "Ġportal", ":", "Ġits", "Ġfront", "Ġlegs", "Ġand", "Ġhead", "Ġfully", "Ġin", "Ġthe", "Ġ", "3", "D", "Ġme", "adow", ",", "Ġmid", "-air", "Ġbody", "Ġpartially", "Ġtransitioning", "Ġinto", "Ġsketch", "Ġlines", ",", "Ġtail", "Ġstill", "Ġa", "Ġdelicate", "Ġink", "Ġdrawing", ".", "ĠEveryday", "Ġobjects", "Ġin", "Ġthe", "Ġroom", "âĢĶ", "like", "Ġa", "Ġdrawn", "Ġte", "ac", "up", "Ġand", "Ġfloating", "Ġbooks", "âĢĶ", "cast", "Ġsubtle", "Ġrealistic", "Ġshadows", ",", "Ġblending", "Ġthe", "Ġdimensions", ".", "ĠSoft", "Ġpast", "el", "Ġpalette", ",", "Ġvisible", "Ġpaper", "Ġtexture", ",", "Ġhigh", "Ġdetail", ",", "Ġwhims", "ical", "Ġyet", "Ġbelievable", "Ġ'", "2", "D", "-to", "-", "3", "D", "Ġtransition", "'", "Ġrealism", ",", "Ġdynamic", "Ġmotion", ",", "Ġenchant", "ing", "Ġand", "Ġtactile", "Ġatmosphere", ".", ]
[DEBUG] bpe_tokenizer.cpp:208  - split prompt "<|im_end|>
<|im_start|>assistant
" to tokens ["<|im_end|>", "Ċ", "<|im_start|>", "assistant", "Ċ", ]
[DEBUG] model_loader.cpp:996  - loading 386/398 tensors from /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf
  |##################################################| 386/386 - 3.19GB/s
[INFO ] model_loader.cpp:1233 - loading tensors completed, taking 1.22s (read: 0.39s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.49s)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (1020.00 MB, 64 tensors, VRAM)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (1020.00 MB, 64 tensors, VRAM)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (1022.57 MB, 96 tensors, VRAM)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (911.57 MB, 162 tensors, VRAM)
[DEBUG] ggml_extend.hpp:2004 - qwen3 compute buffer size: 19.04 MB(VRAM)
[DEBUG] conditioner.hpp:2100 - computing condition graph completed, taking 2350 ms
[INFO ] stable-diffusion.cpp:4269 - get_learned_condition completed, taking 2.35s
[INFO ] stable-diffusion.cpp:4576 - generating image: 1/1 - seed 765675567
[DEBUG] model_loader.cpp:996  - loading 452/453 tensors from /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf
  |##################################################| 452/452 - 3.69GB/s
[INFO ] model_loader.cpp:1233 - loading tensors completed, taking 1.82s (read: 0.59s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.83s)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (1016.95 MB, 30 tensors, VRAM)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (1016.02 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (991.11 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (1021.00 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (1011.49 MB, 33 tensors, VRAM)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (1002.10 MB, 135 tensors, VRAM)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (832.78 MB, 167 tensors, VRAM)
[DEBUG] ggml_extend.hpp:2004 - z_image compute buffer size: 697.81 MB(VRAM)
  |==================================================| 7/7 - 67.35s/it
[INFO ] stable-diffusion.cpp:4608 - sampling completed, taking 469.57s
[INFO ] stable-diffusion.cpp:4620 - generating 1 latent images completed, taking 469.57s
[INFO ] stable-diffusion.cpp:4293 - decoding 1 latents
[DEBUG] model_loader.cpp:996  - loading 138/244 tensors from /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors
  |##################################################| 138/138 - 945.00MB/s
[INFO ] model_loader.cpp:1233 - loading tensors completed, taking 0.20s (read: 0.02s, memcpy: 0.00s, convert: 0.08s, copy_to_backend: 0.03s)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer ( 96.57 MB, 138 tensors, VRAM)
[DEBUG] ggml_extend.hpp:2004 - vae compute buffer size: 2817.00 MB(VRAM)
radv/amdgpu: The CS has been cancelled because the context is lost. This context is guilty of a hard recovery.
terminate called after throwing an instance of 'vk::DeviceLostError'
  what():  vk::Queue::submit: ErrorDeviceLost
Aborted (core dumped)

Interesting: the exact same error happened with my iGPU. Please try setting the env var GGML_VK_MAX_NODES_PER_SUBMIT to a low value, e.g. GGML_VK_MAX_NODES_PER_SUBMIT=1.

It works and image is also perfectly fine.

Can you kindly breifly tell what does GGML_VK_MAX_NODES_PER_SUBMIT=1 do ?

Details

bash  GGML_VK_MAX_NODES_PER_SUBMIT=1 ./sd-cli --mode img_gen --diffusion-model /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf --vae /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors --llm /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf --lora-model-dir /home/tipu/AI/image/loras/ --lora-apply-mode immediately --cfg-scale 1 -H 1024 -W 1024 --steps 7 --seed 765675567 --rng cpu --scheduler smoothstep --sampling-method euler --flow-shift 2 -o output_$(date +%Y-%m-%d_%H-%M-%S).png --fa --vae-conv-direct -p "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere." -v
[DEBUG] main.cpp:602  - version: stable-diffusion.cpp version master-737-3b6c9ca, commit 3b6c9ca
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon Graphics (RADV RENOIR) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 0 | matrix cores: none
[DEBUG] main.cpp:603  - System Info: 
   SSE3 = 1 |    SSSE3 = 1 |    AVX = 1 |    AVX2 = 1 |    F16C = 1 |    FMA = 1 |    BMI2 = 1 |    OPENMP = 1 |    REPACK = 1 | 
[DEBUG] main.cpp:604  - SDCliParams {
  mode: img_gen,
  output_path: "output_2026-06-30_15-18-11.png",
  image_path: "",
  metadata_format: "text",
  verbose: true,
  color: false,
  canny_preprocess: false,
  convert_name: false,
  preview_method: none,
  preview_interval: 1,
  preview_path: "preview.png",
  preview_fps: 16,
  taesd_preview: false,
  preview_noisy: false,
  metadata_raw: false,
  metadata_brief: false,
  metadata_all: false
}
[DEBUG] main.cpp:605  - SDContextParams {
  n_threads: 8,
  model_path: "",
  clip_l_path: "",
  clip_g_path: "",
  clip_vision_path: "",
  t5xxl_path: "",
  llm_path: "/home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf",
  llm_vision_path: "",
  diffusion_model_path: "/home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf",
  high_noise_diffusion_model_path: "",
  uncond_diffusion_model_path: "",
  embeddings_connectors_path: "",
  vae_path: "/home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors",
  vae_format: "auto",
  audio_vae_path: "",
  taesd_path: "",
  esrgan_path: "",
  control_net_path: "",
  embedding_dir: "",
  embeddings: {
  }
  wtype: NONE,
  tensor_type_rules: "",
  lora_model_dir: "/home/tipu/AI/image/loras/",
  hires_upscalers_dir: "",
  photo_maker_path: "",
  rng_type: cpu,
  sampler_rng_type: NONE,
  offload_params_to_cpu: false,
  max_vram: "0",
  stream_layers: false,
  eager_load: false,
  backend: "",
  params_backend: "",
  enable_mmap: false,
  control_net_cpu: false,
  clip_on_cpu: false,
  vae_on_cpu: false,
  flash_attn: true,
  diffusion_flash_attn: false,
  diffusion_conv_direct: false,
  vae_conv_direct: true,
  circular: false,
  circular_x: false,
  circular_y: false,
  chroma_use_dit_mask: true,
  qwen_image_zero_cond_t: false,
  chroma_use_t5_mask: false,
  chroma_t5_mask_pad: 1,
  prediction: NONE,
  lora_apply_mode: immediately,
  force_sdxl_vae_conv_scale: false
}
[DEBUG] main.cpp:606  - SDGenerationParams {
  loras: "{
  }",
  high_noise_loras: "{
  }",
  prompt: "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere.",
  negative_prompt: "",
  clip_skip: -1,
  width: 1024,
  height: 1024,
  batch_count: 1,
  init_image_path: "",
  end_image_path: "",
  mask_image_path: "",
  control_image_path: "",
  ref_image_paths: [],
  control_video_path: "",
  auto_resize_ref_image: true,
  increase_ref_index: false,
  pm_id_images_dir: "",
  pm_id_embed_path: "",
  pm_style_strength: 20,
  skip_layers: [7, 8, 9],
  sample_params: (txt_cfg: 1.00, img_cfg: 1.00, distilled_guidance: 3.50, slg.layer_count: 0, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: smoothstep, sample_method: euler, sample_steps: 7, eta: inf, shifted_timestep: 0, flow_shift: 2.00, extra_sample_args: ),
  high_noise_skip_layers: [7, 8, 9],
  high_noise_sample_params: (txt_cfg: 7.00, img_cfg: 7.00, distilled_guidance: 3.50, slg.layer_count: 0, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: NONE, sample_steps: 20, eta: inf, shifted_timestep: 0, flow_shift: inf, extra_sample_args: ),
  custom_sigmas: [],
  cache_mode: "",
  cache_option: "",
  cache: disabled (threshold=inf, start=0.15, end=0.95),
  moe_boundary: 0.875,
  video_frames: 1,
  fps: 16,
  vace_strength: 1,
  strength: 0.75,
  control_strength: 0.9,
  seed: 765675567,
  upscale_repeats: 1,
  upscale_tile_size: 128,
  hires: { enabled: false, upscaler: "Latent", model_path: "", scale: 2, target_width: 0, target_height: 0, steps: 0, denoising_strength: 0.7, custom_sigmas: [], upscale_tile_size: 128 },
  vae_tiling_params: { 0, 0, 0, 0, 0.5, 0, 0, "" },
}
[DEBUG] ggml_extend_backend.cpp:401  - Found 3 backend devices:
[DEBUG] ggml_extend_backend.cpp:404  - #0: Vulkan0
[DEBUG] ggml_extend_backend.cpp:404  - #1: BLAS
[DEBUG] ggml_extend_backend.cpp:404  - #2: CPU
[DEBUG] ggml_extend_backend.cpp:326  - Initializing backend: Vulkan0
[DEBUG] model_loader.cpp:226  - using 8 threads for model loading
[INFO ] stable-diffusion.cpp:405  - loading diffusion model from '/home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf'
[INFO ] model_loader.cpp:234  - load /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf using gguf format
[DEBUG] model_loader.cpp:283  - init from '/home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf'
[INFO ] stable-diffusion.cpp:467  - loading llm from '/home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf'
[INFO ] model_loader.cpp:234  - load /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf using gguf format
[DEBUG] model_loader.cpp:283  - init from '/home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf'
[INFO ] stable-diffusion.cpp:481  - loading vae from '/home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors'
[INFO ] model_loader.cpp:237  - load /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors using safetensors format
[DEBUG] model_loader.cpp:311  - init from '/home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors', prefix = 'vae.'
[INFO ] stable-diffusion.cpp:532  - Version: Z-Image 
[INFO ] stable-diffusion.cpp:558  - Weight type stat:                      f32: 634  |    q8_0: 433  |    bf16: 28   
[INFO ] stable-diffusion.cpp:559  - Conditioner weight type stat:          f32: 145  |    q8_0: 253  
[INFO ] stable-diffusion.cpp:560  - Diffusion model weight type stat:      f32: 245  |    q8_0: 180  |    bf16: 28   
[INFO ] stable-diffusion.cpp:561  - VAE weight type stat:                  f32: 244  
[DEBUG] stable-diffusion.cpp:563  - ggml tensor size = 400 bytes
[DEBUG] qwen2_tokenizer.cpp:14   - merges size 151387
[DEBUG] qwen2_tokenizer.cpp:39   - vocab size: 151674
[DEBUG] llm.hpp:259  - llm: num_layers = 36, vocab_size = 151936, hidden_size = 2560, intermediate_size = 9728
[DEBUG] z_image.hpp:110  - z_image: num_layers = 30, num_refiner_layers = 2, hidden_size = 3840, num_heads = 30, num_kv_heads = 30, in_channels = 16, out_channels = 16
[INFO ] stable-diffusion.cpp:983  - using VAE for encoding / decoding
[INFO ] auto_encoder_kl.hpp:527  - vae decoder: ch = 128
[INFO ] stable-diffusion.cpp:1019 - Using Conv2d direct in the vae model
[INFO ] stable-diffusion.cpp:1087 - Using flash attention
[INFO ] stable-diffusion.cpp:1101 - Using flash attention in the diffusion model
[DEBUG] stable-diffusion.cpp:1119 - validating model metadata
[DEBUG] stable-diffusion.cpp:1176 - model metadata validated; weights will be prepared lazily
[INFO ] stable-diffusion.cpp:1217 - total params memory size = 11131.89MB (VRAM 11131.89MB, RAM 0.00MB): text_encoders 4076.43MB(VRAM), diffusion_model 6891.46MB(VRAM), vae 164.00MB(VRAM), controlnet 0.00MB(N/A), extensions 0.00MB(N/A)
[INFO ] stable-diffusion.cpp:1324 - running in FLOW mode
[INFO ] stable-diffusion.cpp:4530 - generate_image 1024x1024
[INFO ] denoiser.hpp:919  - get_sigmas with SmoothStep scheduler
[INFO ] stable-diffusion.cpp:3554 - sampling using Euler method
[DEBUG] conditioner.hpp:1604 - parse '<|im_start|>user
A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere.<|im_end|>
<|im_start|>assistant
' to [['<|im_start|>user
', 1], ['A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere.', 1], ['<|im_end|>
<|im_start|>assistant
', 1], ]
[DEBUG] bpe_tokenizer.cpp:208  - split prompt "<|im_start|>user
" to tokens ["<|im_start|>", "user", "Ċ", ]
[DEBUG] bpe_tokenizer.cpp:208  - split prompt "A cozy study with walls covered in flowing ink sketches and soft watercolor washes. At the center, a circular tear in the paper wall opens upward like a magical portal, revealing a lush 3D meadow bathed in gentle morning sunlight. The curled edge of the torn paper casts soft shadows onto the drawn walls. A playful fox leaps out from the portal: its front legs and head fully in the 3D meadow, mid-air body partially transitioning into sketch lines, tail still a delicate ink drawing. Everyday objects in the room—like a drawn teacup and floating books—cast subtle realistic shadows, blending the dimensions. Soft pastel palette, visible paper texture, high detail, whimsical yet believable '2D-to-3D transition' realism, dynamic motion, enchanting and tactile atmosphere." to tokens ["A", "Ġcozy", "Ġstudy", "Ġwith", "Ġwalls", "Ġcovered", "Ġin", "Ġflowing", "Ġink", "Ġsketches", "Ġand", "Ġsoft", "Ġwater", "color", "Ġwash", "es", ".", "ĠAt", "Ġthe", "Ġcenter", ",", "Ġa", "Ġcircular", "Ġtear", "Ġin", "Ġthe", "Ġpaper", "Ġwall", "Ġopens", "Ġupward", "Ġlike", "Ġa", "Ġmagical", "Ġportal", ",", "Ġrevealing", "Ġa", "Ġlush", "Ġ", "3", "D", "Ġme", "adow", "Ġbath", "ed", "Ġin", "Ġgentle", "Ġmorning", "Ġsunlight", ".", "ĠThe", "Ġcurled", "Ġedge", "Ġof", "Ġthe", "Ġtorn", "Ġpaper", "Ġcasts", "Ġsoft", "Ġshadows", "Ġonto", "Ġthe", "Ġdrawn", "Ġwalls", ".", "ĠA", "Ġplayful", "Ġfox", "Ġleaps", "Ġout", "Ġfrom", "Ġthe", "Ġportal", ":", "Ġits", "Ġfront", "Ġlegs", "Ġand", "Ġhead", "Ġfully", "Ġin", "Ġthe", "Ġ", "3", "D", "Ġme", "adow", ",", "Ġmid", "-air", "Ġbody", "Ġpartially", "Ġtransitioning", "Ġinto", "Ġsketch", "Ġlines", ",", "Ġtail", "Ġstill", "Ġa", "Ġdelicate", "Ġink", "Ġdrawing", ".", "ĠEveryday", "Ġobjects", "Ġin", "Ġthe", "Ġroom", "âĢĶ", "like", "Ġa", "Ġdrawn", "Ġte", "ac", "up", "Ġand", "Ġfloating", "Ġbooks", "âĢĶ", "cast", "Ġsubtle", "Ġrealistic", "Ġshadows", ",", "Ġblending", "Ġthe", "Ġdimensions", ".", "ĠSoft", "Ġpast", "el", "Ġpalette", ",", "Ġvisible", "Ġpaper", "Ġtexture", ",", "Ġhigh", "Ġdetail", ",", "Ġwhims", "ical", "Ġyet", "Ġbelievable", "Ġ'", "2", "D", "-to", "-", "3", "D", "Ġtransition", "'", "Ġrealism", ",", "Ġdynamic", "Ġmotion", ",", "Ġenchant", "ing", "Ġand", "Ġtactile", "Ġatmosphere", ".", ]
[DEBUG] bpe_tokenizer.cpp:208  - split prompt "<|im_end|>
<|im_start|>assistant
" to tokens ["<|im_end|>", "Ċ", "<|im_start|>", "assistant", "Ċ", ]
[DEBUG] model_loader.cpp:996  - loading 386/398 tensors from /home/tipu/AI/models/other/Qwen3-4b-z-engineer/Qwen3-4b-Z-Engineer-V2-Q8_0.gguf
  |##################################################| 386/386 - 1.72GB/s
[INFO ] model_loader.cpp:1233 - loading tensors completed, taking 2.26s (read: 1.68s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.44s)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (1020.00 MB, 64 tensors, VRAM)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (1020.00 MB, 64 tensors, VRAM)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (1022.57 MB, 96 tensors, VRAM)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (911.57 MB, 162 tensors, VRAM)
[DEBUG] ggml_extend.hpp:2004 - qwen3 compute buffer size: 19.04 MB(VRAM)
[DEBUG] conditioner.hpp:2100 - computing condition graph completed, taking 3438 ms
[INFO ] stable-diffusion.cpp:4269 - get_learned_condition completed, taking 3.44s
[INFO ] stable-diffusion.cpp:4576 - generating image: 1/1 - seed 765675567
[DEBUG] model_loader.cpp:996  - loading 452/453 tensors from /home/tipu/AI/image/unet/Z-Image-Turbo-Art-Q8_0.gguf
  |##################################################| 452/452 - 2.15GB/s
[INFO ] model_loader.cpp:1233 - loading tensors completed, taking 3.13s (read: 2.19s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.67s)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (1016.95 MB, 30 tensors, VRAM)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (1016.02 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (991.11 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (1021.00 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (1011.49 MB, 33 tensors, VRAM)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (1002.10 MB, 135 tensors, VRAM)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer (832.78 MB, 167 tensors, VRAM)
[DEBUG] ggml_extend.hpp:2004 - z_image compute buffer size: 697.81 MB(VRAM)
  |==================================================| 7/7 - 64.76s/it
[INFO ] stable-diffusion.cpp:4608 - sampling completed, taking 456.83s
[INFO ] stable-diffusion.cpp:4620 - generating 1 latent images completed, taking 456.83s
[INFO ] stable-diffusion.cpp:4293 - decoding 1 latents
[DEBUG] model_loader.cpp:996  - loading 138/244 tensors from /home/tipu/AI/image/vae/UltraFlux-v1_vae.safetensors
  |##################################################| 138/138 - 465.52MB/s
[INFO ] model_loader.cpp:1233 - loading tensors completed, taking 0.41s (read: 0.10s, memcpy: 0.00s, convert: 0.06s, copy_to_backend: 0.04s)
[DEBUG] model_manager.cpp:229  - model manager prepared params backend buffer ( 96.57 MB, 138 tensors, VRAM)
[DEBUG] ggml_extend.hpp:2004 - vae compute buffer size: 2817.00 MB(VRAM)
[DEBUG] vae.hpp:215  - computing vae decode graph completed, taking 22.43s
[INFO ] stable-diffusion.cpp:4313 - latent 1 decoded, taking 22.43s
[INFO ] stable-diffusion.cpp:4317 - decode_first_stage completed, taking 22.43s
[INFO ] stable-diffusion.cpp:4757 - generate_image completed in 482.73s
[DEBUG] model_manager.cpp:770  - model manager releasing params backend buffer (1020.00 MB, 64 tensors, VRAM)
[DEBUG] model_manager.cpp:770  - model manager releasing params backend buffer (1020.00 MB, 64 tensors, VRAM)
[DEBUG] model_manager.cpp:770  - model manager releasing params backend buffer (1022.57 MB, 96 tensors, VRAM)
[DEBUG] model_manager.cpp:770  - model manager releasing params backend buffer (911.57 MB, 162 tensors, VRAM)
[DEBUG] model_manager.cpp:770  - model manager releasing params backend buffer (1016.95 MB, 30 tensors, VRAM)
[DEBUG] model_manager.cpp:770  - model manager releasing params backend buffer (1016.02 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:770  - model manager releasing params backend buffer (991.11 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:770  - model manager releasing params backend buffer (1021.00 MB, 29 tensors, VRAM)
[DEBUG] model_manager.cpp:770  - model manager releasing params backend buffer (1011.49 MB, 33 tensors, VRAM)
[DEBUG] model_manager.cpp:770  - model manager releasing params backend buffer (1002.10 MB, 135 tensors, VRAM)
[DEBUG] model_manager.cpp:770  - model manager releasing params backend buffer (832.78 MB, 167 tensors, VRAM)
[DEBUG] model_manager.cpp:770  - model manager releasing params backend buffer ( 96.57 MB, 138 tensors, VRAM)
[INFO ] main.cpp:467  - save result image 0 to 'output_2026-06-30_15-18-11.png' (success)
[INFO ] main.cpp:539  - 1/1 images saved

I see you had #24872 in llama.cpp.

I see you had #24872 in llama.cpp.

Yeah, ggml-org/llama.cpp#24872 is a workaround for an issue (ggml-org/llama.cpp#21724) that started to happen some time ago on llama.cpp, mainly on low-powered GPUs.

Could you let me know what's the minimum value that works for you?

I have tried 1 to 10 values all work on my vega8 gfx90c. I have set value to 5 in my bashrc

Maybe wbruna can shed some light how to find best suitable value.

Navigation Menu

Repository navigation

[Bug] Vulkan Backend Produces Distorted / Gibrish Images with `--vae-conv-direct` in Higher Resolutions #1673

Git commit

Operating System & Version

GGML backends

Command-line arguments used

Steps to reproduce

What you expected to happen

What actually happened

Logs / error messages / stack trace

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

[Bug] Vulkan Backend Produces Distorted / Gibrish Images with --vae-conv-direct in Higher Resolutions #1673

Description

Git commit

Operating System & Version

GGML backends

Command-line arguments used

Steps to reproduce

What you expected to happen

What actually happened

Logs / error messages / stack trace

Activity

GreenShadows commented on Jun 17, 2026

engrtipusultan commented on Jun 17, 2026

engrtipusultan commented on Jun 17, 2026

wbruna commented on Jun 17, 2026

GreenShadows commented on Jun 17, 2026

engrtipusultan commented on Jun 18, 2026

wbruna commented on Jun 29, 2026

wbruna commented on Jun 29, 2026

engrtipusultan commented on Jun 30, 2026

wbruna commented on Jun 30, 2026

engrtipusultan commented on Jun 30, 2026

engrtipusultan commented on Jun 30, 2026

wbruna commented on Jun 30, 2026

LostRuins commented on Jun 30, 2026

engrtipusultan commented on Jun 30, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions

[Bug] Vulkan Backend Produces Distorted / Gibrish Images with `--vae-conv-direct` in Higher Resolutions #1673