🔬MicroVerse

[ICLR 2026] MicroVerse: A Preliminary Exploration Toward a Micro-World Simulation

MicroVerse is the first AI framework designed to simulate microscopic biological processes at organ, cellular, and subcellular levels. Unlike existing video generation models that mostly mimic visual textures, MicroVerse integrates physical and biological constraints to faithfully reproduce the mechanisms of life. Leveraging the MicroSim-10K dataset and evaluated against MicroWorldBench, it sets a new benchmark for microscale AI simulations, enabling applications in biomedical research, education, and interactive visualization.

This work represents a leap from “macro-scale visual imitation” to mechanism-aware micro-world modeling, making AI a tool for understanding life at its most fundamental scales.

✨ Highlights

• 🌱 True Microscale Simulation: Accurately models organ, cellular, and subcellular dynamics, capturing DNA replication, cell division, and apoptosis with scientific fidelity.
• 📊 Benchmarking Science: Evaluated with MicroWorldBench, the first expert-annotated rubric-based test for microscopic simulations.
• 💾 Expert-Verified Dataset: MicroSim-10K provides 9,601 high-resolution, semantically rich video clips guiding AI toward mechanism-level understanding.
• 🧠 Mechanism over Visuals: Moves beyond realistic textures to faithfully reproduce underlying biological laws and temporal dynamics.
• 🎓 Research & Education Ready: Enables cost-effective, interactive exploration of life’s hidden processes, bridging AI, biology, and learning.

📣 Latest News!!

• [2026.04.01] 🎉 MicroWorldBench evaluation and MicroVerse training code released.
• [2026.03.03] 🚀 Paper released.
• [2026.01.26] 🎉 MicroVerse has been officially accepted as a Poster at ICLR 2026.

🚀 Training

We provide fine-tuning scripts for MicroVerse under the train/ directory, supporting LoRA and full parameter training on the MicroSim-10K dataset.

Quick Start:

cd train
pip install -r requirements.txt

# Prepare your dataset (see train/README.md for metadata format)
# then run LoRA fine-tuning:
bash scripts/train_lora_1.3B.sh   # 1x 24GB GPU
bash scripts/train_lora_14B.sh    # 2x 24GB GPUs

# Inference after training:
python inference.py \
  --model_size 1.3B --mode lora \
  --checkpoint ./outputs/lora_1.3B \
  --prompt "DNA replication inside a eukaryotic cell nucleus" \
  --output ./results/output.mp4

See train/README.md for dataset format, GPU requirements, and all training arguments.

---

🧪 Evaluation (MicroWorldBench)

We provide the evaluation pipeline for MicroWorldBench under the eval/ directory. It assesses generated videos across three dimensions: Scientific Accuracy, Visual Quality, and Instruction Following.

Quick Start:

# 1. Install dependencies
pip install requests opencv-python numpy tqdm

# 2. Configure API credentials
cp eval/.env.example eval/.env  # then fill in EVAL_API_KEY

# 3. Place your model's videos
mkdir eval/<ModelName>
# Add videos named by task index: 1.mp4, 2.mp4, ...

# 4. Run evaluation
cd eval
python video_evaluator.py   # Step 1: extract frames & judge
python calculate_scores.py  # Step 2: aggregate scores & report

Results are saved to eval/eval_result/. See eval/README.md for full details.

⚙️ Acknowledgments

This project builds upon the Wan2.1 video generation model, with its codebase based on DiffSynth-Studio.

✏️ Citation

If you find our paper and code useful in your research, please consider giving a star ⭐ and citation 📝:

@article{wang2026microverse,
  title={MicroVerse: A Preliminary Exploration Toward a Micro-World Simulation},
  author={Wang, Rongsheng and Wu, Minghao and Zhou, Hongru and Yu, Zhihan and Cai, Zhenyang and Chen, Junying and Wang, Benyou},
  journal={arXiv preprint arXiv:2603.00585},
  year={2026}
}