FACM: Flow-Anchored Consistency Models

🔥 FACM outperforms 2×250-step Lightning-DiT on ImageNet 256 with only 2 steps

FID=1.70 (1-step)      FID=1.32 (2-step)

📄 This is the official implementation of the paper: Flow-Anchored Consistency Models.

Code: https://github.com/ali-vilab/FACM

Abstract

Continuous-time Consistency Models (CMs) promise efficient few-step generation but face significant challenges with training instability. We argue this instability stems from a fundamental conflict: by training a network to learn only a shortcut across a probability flow, the model loses its grasp on the instantaneous velocity field that defines the flow. Our solution is to explicitly anchor the model in the underlying flow during training. We introduce the Flow-Anchored Consistency Model (FACM), a simple but effective training strategy that uses a Flow Matching (FM) task as an anchor for the primary CM shortcut objective. This Flow-Anchoring approach requires no architectural modifications and is broadly compatible with standard model architectures. By distilling a pre-trained LightningDiT model, our method achieves a state-of-the-art FID of 1.32 with two steps (NFE=2) and 1.76 with just one step (NFE=1) on ImageNet 256x256, significantly outperforming previous methods. This provides a general and effective recipe for building high-performance, few-step generative models. Our code and pretrained models: this https URL .

ImageNet 256 Performance on 8 × A100 GPUs

Quick Start

Prerequisites:

Download the required model weights and statistics files from HuggingFace or ModelScope to ./cache

Including: fid-50k-256.npz, latents_stats.pt, vavae-imagenet256-f16d32-dinov2.pt

Data Preparation

export DATA_PATH="/path/to/imagenet"
export OUTPUT_PATH="/path/to/latents"
bash scripts/extract.sh

*Note: You can also download pre-extracted ImageNet latents following Lightning-DiT.

Inference

Download pretrained FACM model checkpoint 100ep-stg2.pt or 400ep-stg2.pt to ./cache

bash scripts/test.sh --ckpt-path cache/100ep-stg2.pt --sampling-steps 2

bash scripts/test.sh --ckpt-path cache/400ep-stg2.pt --sampling-steps 1

Training

Download pretrained FM model checkpoint 800ep-stg1.pt to ./cache

export DATA_PATH="/path/to/latents"
bash scripts/train.sh

Pretraining (Optional)

Replace configs/lightningdit_xl_vavae_f16d32.yaml and transport/transport.py of Lightning-DiT with our ldit/lightningdit_xl_vavae_f16d32.yaml and ldit/transport.py, then follow the instructions.

Reproductions

reproductions

We include reproductions of MeanFlow and sCM. Switch methods by changing the loss function in train.py line 81:

facm_loss = FACMLoss()      # FACM (default)
facm_loss = MeanFlowLoss()  # MeanFlow  
facm_loss = sCMLoss()       # sCM

Citation

If you use FACM or its methods in your work, please cite the following BibTeX entries:

bibtex

@misc{peng2025facm,
      title={Flow-Anchored Consistency Models}, 
      author={Yansong Peng and Kai Zhu and Yu Liu and Pingyu Wu and Hebei Li and Xiaoyan Sun and Feng Wu},
      year={2025},
      eprint={2507.03738},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Acknowledgements

The model architecture part is based on the Lightning-DiT repository.

✨ Feel free to contribute and reach out if you have any questions! ✨