[CVPR 2026] Denoising as Path Planning: Training-Free Acceleration of Diffusion Models with DPCache

wan_720.mp4

DPCache accelerates diffusion models via prediction-based caching, using dynamic programming with Path-Aware Cost Tensor (PACT) to optimize the caching schedule.

✨ Features

• Efficient Caching: Dynamic programming scheduling with Taylor expansion predictor
• Path-Aware Cost Tensor: 3D cost matrix for optimal step selection
• Easy Adaptation: Modular design with CacheHelper for quick integration
• Reference Examples: FLUX (text-to-image) and Wan2.1 (image-to-video) implementations

📰 News

• [2025-03] 🚀 Released with support for FLUX (text-to-image) and Wan2.1 (image-to-video).
• [2025-02] 🎉 DPCache has been accepted to CVPR 2026!

🤖 Supported Models

Model	Type	Reference Implementation
FLUX.1-dev	Text-to-Image	dpcache_flux.py
Wan2.1-I2V-14B-720P	Image-to-Video	dpcache_wan.py

📋 Requirements

• Python 3.10+
• CUDA 12.0+ (recommended)
• Additional GPU memory is required during calibration

Install dependencies:

pip install -r requirements.txt

🚀 Quick Start

Example 1: FLUX (Text-to-Image)

Inference with pre-calibrated cost matrix:

python dpcache_flux_infer.py --mode infer \
    --k 13 \
    --first_full_steps 3 \
    --dataset drawbench \
    --sample_size 100 \
    --output_path "flux_output"

Calibration (generate custom cost matrix):

python dpcache_flux_infer.py --mode calibrate \
    --cali_prefix "flux_calibration" \
    --dataset drawbench \
    --sample_rule fix \
    --sample_size 10 \
    --output_path "calibration_results"

This will generate a cost matrix file named final_cost_matrix_flux_calibration.pkl. Use it in inference via --cost_matrix_path "final_cost_matrix_flux_calibration.pkl".

Example 2: Wan2.1 (Image-to-Video)

Inference with pre-calibrated cost matrix:

python dpcache_wan_infer.py --mode infer \
    --image_path test.jpg \
    --prompt "a blue car driving down a dirt road near train tracks" \
    --k 12 

Disable cache (baseline):

python dpcache_wan_infer.py --mode infer --no_cache

🔧 Adapt to Your Own Model

DPCache can be integrated into other diffusion models. Follow these steps:

1. Import CacheHelper from dpcache/cache_utils.py
2. Initialize cache using init_cache() with your model's transformer blocks
3. Wrap forward pass using CacheHelper to manage caching logic (customize config as needed)
4. Run calibration to generate cost matrix for your specific model/config

By default, this repo uses Taylor expansion-based predictors (as in TaylorSeer), but any other predictor can be plugged in, as long as the same predictor is used consistently in both calibration and inference.

Reference implementations:

• dpcache_flux.py - Integration with FLUX transformer blocks
• dpcache_wan.py - Integration with Wan2.1 transformer blocks

For adapting other diffusion models, please refer directly to the above files for how to:

• initialize cache with init_cache
• manage caching logic via CacheHelper in the main transformer forward
• cache or reuse block outputs inside each transformer block

📖 Citation

If you use this code, please cite:

@article{DPCache2026,
  title={Denoising as Path Planning: Training-Free Acceleration of Diffusion Models with DPCache},
  author={Cui, Bowen and Wang, Yuanbin and Xu, Huajiang and Chen, Biaolong and Zhang, Aixi and Jiang, Hao and Jin, Zhengzheng and Liu, Xu and Huang, Pipei},
  journal={arXiv preprint arXiv:2602.22654},
  year={2026}
}

🙏 Acknowledgements

This repository is built upon Diffusers. We thank the following projects for their great work and contributions:

• Models: FLUX by Black Forest Labs, Wan2.1 by Wan-AI
• Methods: TaylorSeer for Taylor expansion-based feature prediction
• Datasets: DrawBench, PartiPrompts, VBench for quality evaluation