**Keywords: v0.3.68**
---
ComfyUI v0.3.68 is released (https://github.com/comfyanonymous/ComfyUI/releases/tag/v0.3.68, https://github.com/comfyanonymous/ComfyUI/releases/tag/v0.3.67, https://github.com/comfyanonymous/ComfyUI/releases/tag/v0.3.66, https://github.com/comfyanonymous/ComfyUI/releases/tag/v0.3.65, https://github.com/comfyanonymous/ComfyUI/releases/tag/v0.3.64, https://github.com/comfyanonymous/ComfyUI/releases/tag/v0.3.63)
## v0.3.68
- Bump stable portable to cu130 python 3.13.9
- Speed up offloading using pinned memory with race condition fixes
- Add RAM Pressure cache mode for better resource management
- Small speed improvements to --async-offload
- Optimizations for fp8 torch.compile operations
- Multiple API node conversions to new client standards (Luma, Minimax, Pixverse, Ideogram, StabilityAI, Pika)
- Convert nodes_recraft.py to V3 schema and hypernetwork/OpenAI node conversions
- Remove comfy api key from queue api for improved security
- Add a ScaleROPE node (works on WAN and Lumina models)
- Support for 12-20 second durations in LTXV API nodes
- Race condition fixes in async-offload preventing corruption
- Cache signature hashing fixes for bytes data
- Torch compile regression fixes
- And more...
---
## v0.3.67
- Only disable cudnn on newer AMD GPUs to improve compatibility
- Add endpoint support for published subgraphs from custom nodes
- Integrated dependency-aware caching and resolved issues with --cache-none when using loops and lazy evaluation
- Implemented preliminary support for handling multi-dimensional latent configurations
- Upgraded network client to version 2 featuring async operations, cancellation support, and download capabilities
- Converted Tripo and Gemini API nodes to V3 schema, plus added LTXV API nodes
- Bumped portable dependencies to PyTorch cu130 with Python 3.13.9 support
- Added batch script option to run ComfyUI without API nodes
- Updated frontend to version 1.28.8
- Resolved Windows-specific retry issues and torch-directml usage warnings
- Minor README improvements and template updates (0.2.2 → 0.2.4)
- And more...
---
## v0.3.66
- Faster workflow cancelling implemented
- Chroma radiance speedup and batch size handling corrections
- PyTorch compiler disabled for cast_bias_weight function
- CUDA malloc turned off by default with --fast autotune
- Python 3.14 installation instructions added
- gfx942 GPU no longer supports fp8 operations
- PyTorch stable version updated to cu130
- CuDNN workaround for VAE memory issues on torch 2.9
- Veo3.1 model added to api-nodes
- TemporalScoreRescaling node introduced
- Dynamic pricing format fixed in api-nodes
- Deprecated API alert feature implemented
- Manual patches merging refactored with merge_nested_dicts
- Frontend bumped to version 1.28.7
- EasyCache apply_cache_diff batch slicing corrected
- Chroma radiance batch size >1 output issues resolved
- And more...
---
## v0.3.65
- Multiple node files converted to V3 schema (compositing, latent operations, SD3, Flux, upscaling, Hunyuan models)
- Better memory estimation for the SD/Flux VAE on AMD GPUs
- RDNA4 pytorch attention support added for ROCm 7.0+
- Fixed loading old stable diffusion ckpt files on newer numpy
- Audio node stereo/mono handling corrections
- Fixed fp8 scaled LoRA application
- VAE cache VRAM leak fixes
- Enum class support for Combo options
- Lazy formatting in logging
- Price extractor feature for API nodes
- Aspect ratio parameters for GeminiImage nodes
- mmaudio 16k VAE implemented
- Diffusion models now always set to eval() mode
- Documentation updated to version 0.3.0
- And more...
---
## v0.3.64
- Enhanced pylint rules for API nodes
- Support for multiline negative prompts in PixVerse API nodes
- Converted nodes_pika.py and nodes_kling.py to V3 schema format
- Implemented Gemma 3 as a text encoder option
- Fixed custom multipart parser in ReCraft-API node to properly return FormData
- Added Sora2 API node for video generation capabilities
- Temporary fix applied for LTXV custom nodes compatibility
- Bumped frontend version to 1.27.10
- Updated template to version 0.1.94
- And more...
---
## v0.3.63
- Bump frontend to 1.27.7
- Multiple V3 schema conversions for nodes (audio encoder, differential diffusion, morphology, torch compile)
- VAE tiled fallback VRAM leak resolution for both SD and WAN VAE
- Remove soundfile dependency. No more torchaudio load or save
- Turn on TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL by default for AMD ROCm
- Epsilon Scaling node for exposure bias correction
- Support for new Hunyuan VAE
- Add kling-2-5-turbo to txt2video and img2video nodes
- Enhanced linting with pylint for API nodes folder
- Various indentation and import fixes across API nodes
- And more...
---
Issue News:
Features/Updates News:
See #2276