ProfilingDiT

Official Implementation of ["Model Reveals What to Cache: Profiling-Based Feature Reuse for Video Diffusion Models"]

📄Paper 🔗arXiv

This repository contains the official implementation of our paper: Model Reveals What to Cache: Profiling-Based Feature Reuse for Video Diffusion Models.
Please follow the official link for setting up the environment.

---

📌 Table of Contents

• 🔥 Latest News
• 📀 Installation
• 🚀 Running the Code
• 📊 Quantitative Comparison
• ⚡ Scale to Multi-GPU
• 📝 To-Do List

---

🔥 Latest News

🔔 Latest News
• If you like our project, please give us a star ⭐ on GitHub for the latest update.
• [2025/04/04] 🎉 Submitted to arXiv for review.
• [2025/04/04] 🔥 Released open-source code for the latest model.

---

📀 Installation

Follow the official HunyuanVideo and WAN 2.1 environment setup guide.

pip install -r requirements.txt

---

🚀 Running the Code

HunyuanVideo

cd HunyuanVideo
python3 sample_video.py \
    --video-size 360 720 \
    --video-length 129 \
    --infer-steps 50 \
    --prompt "cat walk on grass" \
    --flow-reverse \
    --use-cpu-offload \
    --save-path ./results \
    --seed 42 \
    --model-base "ckpts" \
    --dit-weight "ckpts/hunyuan-video-t2v-720p/transformers/mp_rank_00_model_states.pt" \
    --delta_cache

WAN 2.1

cd Wan2.1
python generate.py \
    --task t2v-14B \
    --size 832*480 \
    --frame_num 81 \
    --prompt "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage." \
    --delta_cache

---

📊 Quantitative Comparison

HunyuanVideo Baseline

Method	VBench ↑	LPIPS ↓	PSNR ↑	SSIM ↑	FID ↓	Latency (ms) ↓	Speedup ↑
HunyuanVideo (720P, 129 frames)	0.7703	--	--	--	--	1745	--
TeaCache (slow) Tea	0.7700	0.1720	21.91	0.7456	77.67	1052	1.66×
TeaCache (fast) Tea	0.7677	0.1830	21.60	0.7323	83.85	753	2.31×
Ours (HunyuanVideo)	0.7642	0.1203	26.44	0.8445	41.10	932	1.87×

Wan2.1 Baseline

Method	VBench ↑	LPIPS ↓	PSNR ↑	SSIM ↑	FID ↓	Latency (ms) ↓	Speedup ↑
Wan2.1 (480P, 81 frames)	0.7582	--	--	--	--	497	--
TeaCache (0.2thres) Tea	0.7604	0.2913	16.17	0.5685	117.61	249	2.00×
Ours (Wan2.1)	0.7615	0.1256	22.02	0.7899	62.56	247	2.01×

Tables: Quantitative comparison with prior methods under HunyuanVideo and Wan2.1 baselines.
🔺 Higher is better for VBench, PSNR, SSIM, and Speedup.
🔻 Lower is better for LPIPS, FID, and Latency.

---

⚡ Scale to Multi-GPU

Our method efficiently scales across multiple GPUs to accelerate inference and training.
By leveraging model parallelism, NCCL communication, and optimized memory management, we achieve significant speedup without compromising quality.

🔑 Key Features:

• Increased Throughput 🚀: Distributes computation across multiple GPUs to process more frames in parallel.
• Optimized Memory Usage 🔧: Dynamically allocates memory to prevent bottlenecks.
• Flexible Deployment 💡: Works seamlessly on both single-node and distributed setups.
• NCCL Optimization 🔄: Uses efficient GPU-GPU communication to minimize overhead.

For detailed setup and configurations, please refer to our Multi-GPU Guide. 🚀

---

📝 To-Do List:

• OpenSora2 🏗️ (Upcoming Support)
• Optimize Caching for CogVideoX ⚙️

---

📚 Citation

@misc{ma2025modelrevealscacheprofilingbased,
      title={Model Reveals What to Cache: Profiling-Based Feature Reuse for Video Diffusion Models}, 
      author={Xuran Ma and Yexin Liu and Yaofu Liu and Xianfeng Wu and Mingzhe Zheng and Zihao Wang and Ser-Nam Lim and Harry Yang},
      year={2025},
      eprint={2504.03140},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2504.03140}, 
}

📜 License

This project is licensed under the Apache 2.0 License.