MIGA is a novel train-free method for infinite-frame video generation that addresses two key limitations of existing frame-level autoregressive frameworks (e.g., FIFO-Diffusion):
-
Two-Stage Training-Inference Alignment (TTA): Reduces the noise span of latents fed to the model during inference through zigzag iterative denoising (Stage 1) and unified noise-level denoising (Stage 2), effectively bridging the training-inference gap.
-
Dual Consistency Enhancement (DCE): Promotes long-term temporal consistency through:
- Self-Reflection: Evaluates and corrects early high-noise frames via latent-space similarity analysis
- Long-Range Frame Guidance: Leverages distant low-noise frames to steer generation
MIGA achieves state-of-the-art performance on VBench and NarrLV benchmarks while maintaining constant memory consumption.
- [2025/05] MIGA paper is accepted by ICML 2026~.
git clone https://github.com/AMAP-ML/MIGA.git
cd MIGAMIGA is implemented on two foundation models: Wan2.1 and VideoCrafter2. Instructions for each are provided below.
☀️ Start with Wan2.1
conda create -n miga_wan python=3.10 -y
conda activate miga_wan
pip install torch==2.4.0 torchvision==0.19.0 --index-url https://download.pytorch.org/whl/cu121
pip install diffusers==0.31.0 transformers accelerate
pip install easydict einops imageio imageio-ffmpeg
pip install opencv-python pillow tqdm thop pyyaml| Model | Resolution | Checkpoint |
|---|---|---|
| Wan2.1-1.3B | 480x832 | Hugging Face |
Store them as the following structure:
Wan2.1-T2V-1.3B/
├── Wan2.1_VAE.pth
├── models_t5_umt5-xxl-enc-bf16.pth
└── ...
cd wan_based
python generate.py \
--ckpt_dir /path/to/Wan2.1-T2V-1.3B \
--miga_config ../configs/wan2.1_1.3B.yaml \
--prompt "A fluffy Corgi dog trots happily across a lush green lawn. Its short legs and wagging tail convey pure delight as it explores the grassy expanse, a charming and energetic display of canine joy." \
--save_dir ./outputs \
--exp_name corgi_demo☀️ Start with VideoCrafter2
conda create -n miga_vc2 python=3.10 -y
conda activate miga_vc2
pip install torch==2.1.0 torchvision==0.16.0 --index-url https://download.pytorch.org/whl/cu121
pip install pytorch-lightning==1.9.0
pip install omegaconf einops decord imageio imageio-ffmpeg
pip install opencv-python pillow tqdm open-clip-torch
pip install transformers kornia pyyaml| Model | Resolution | Checkpoint |
|---|---|---|
| VideoCrafter2 (Text2Video) | 320x512 | Hugging Face |
Store them as the following structure:
videocrafter_models/
└── base_512_v2/
└── model.ckpt
cd videocraft_based
python generate.py \
--ckpt_path /path/to/videocrafter_models/base_512_v2/model.ckpt \
--miga_config ../configs/videocrafter2.yaml \
--prompt "An astronaut floating in space, high quality, 4K resolution." \
--save_dir ./outputs \
--exp_name astronaut_demoMIGA hyperparameters are managed via YAML config files in configs/. You can modify these files or override individual parameters via CLI arguments.
| Parameter | Wan2.1 | VC2 | Description |
|---|---|---|---|
sampling_steps |
54 | 64 | Total denoising steps T |
saw_width |
7 | 4 | Zigzag width L_zig in Stage 1 |
long_iter_nums |
20 | 30 | Number of generated latent chunks |
temporal_memory_len |
4 | 4 | Long-range guidance frames m_guid |
involve_resample |
true | false | Enable self-reflection (DCE) |
resample_threshold |
0.001 | 0.05 | Consistency drop threshold delta_adju |
Generated video length: N = long_iter_nums x saw_width latent frames
Override example:
python generate.py \
--ckpt_dir /path/to/model \
--miga_config ../configs/wan2.1_1.3B.yaml \
--prompt "Your prompt here" \
--long_iter_nums 40 \
--involve_resample trueIf you find this work useful, please cite our paper:
@inproceedings{feng2025miga,
title={Enhancing Train-Free Infinite-Frame Generation for Consistent Long Videos},
author={Xiaokun Feng and Mingze Wu and Hao Yu and Jitan Tan and Jie Xiao and Fangyuan Zhao and Jiaxu Miao and Kaiwen Hu and Jie Wu and Xiangyang Chu and Ke Huang},
booktitle={ICML},
year={2025}
}Our codebase builds on VideoCrafter2, Wan2.1, and FIFO-Diffusion. We appreciate their excellent work!