(CVPR 2025, Highlight) ShortDF: Optimizing for the Shortest Path in Denoising Diffusion Model

🧭 Description

This repository is the official implementation of ShortDF (CVPR 2025).

ShortDF acts as an "intelligent navigation system" for diffusion models. Instead of blindly following a fixed trajectory, it solves for the optimal path via Implicit Graph Modeling and Shortest-Path Relaxation. This allows a single step to achieve the efficacy of multiple steps.

Core Mechanism

• Path Optimization: We treat diffusion steps as nodes in a graph. If a multi-step path (e.g., $10 \to 2 \to 0$) yields better quality than a direct step ($10 \to 0$), the model optimizes the direct step to match that higher quality.
• Error Propagation: Through iterative training, long paths (e.g., $100 \to 0$) absorb the refined information from intermediate steps, achieving fewer-step convergence comparable to the original multi-step process.

Highlights

• 5× Speedup: Achieves quality comparable to 10-step DDIM on CIFAR-10 in just 2 steps.
• Higher Fidelity: Improves FID by 18.5% on CIFAR-10.
• Robustness: Demonstrates superior performance on CelebA and LSUN-Church datasets across various sampling steps.

Figure 1: Extreme Speed Test on CIFAR-10. ShortDF achieves comparable quality at 2 steps, demonstrating a 5× speedup.

Figure 2: Multi-Dataset Performance and Sampling Trajectory (CelebA, Church). Note ShortDF's clear quality advantage across the sampling sequence (Step i).

For more details, please refer to our CVPR 2025 paper.

---

🚀 Running the Experiments

Training

Training follows the standard DDPM protocol.

python main.py --config {DATASET}.yml --exp {PROJECT_PATH} --doc {MODEL_NAME} --ni

Loss Design & Strategy

The ShortDF-specific loss is implemented in ./functions/losses.py as shortdf_relax_loss. We recommend the following training strategies:

1. Two-stage training (Recommended):

   • Phase 1: Train using standard noise loss (or load a pretrained DDPM checkpoint) to stabilize the model.
   • Phase 2: Fine-tune with shortdf_relax_loss to optimize for shortest-path residuals.
   • Benefit: Reduces training complexity and ensures stable convergence.

2. One-stage training (Optional):

   • Train with both standard noise loss and shortdf_relax_loss from scratch.
   • Configuration: Adjust noise_weight and relax_weight in the config file to balance the contributions based on your dataset and model size.

Sampling

1. Download Pretrained Models

We provide pretrained models for the CIFAR-10, CelebA, and LSUN-Church datasets.

• Download Link: Google Drive
• Setup: After downloading, please place the model file in the following directory structure: logs/{DATASET}/ckpt.pth

2. General Sampling (FID Evaluation)

To generate samples and evaluate the Fréchet Inception Distance (FID):

python main.py --config {DATASET}.yml --exp {PROJECT_PATH} --doc {MODEL_NAME} --sample --fid --timesteps {STEPS} --eta {ETA} --ni

• --eta: Controls the variance scale ($\eta=0$ for DDIM, $\eta=1$ for DDPM).
• --timesteps: Specifies the number of diffusion steps ($T$).
• --doc: Identifies the folder name containing the checkpoint.

Example (CIFAR-10):

python main.py --config cifar10.yml --exp ./ --doc cifar10 --sample --fid --timesteps 2 --eta 0 --ni --skip_type quad
python main.py --config cifar10.yml --exp ./ --doc cifar10 --sample --fid --timesteps 10 --eta 1 --ni --skip_type quad

Example (LSUN-Church):

python main.py --config church.yml --exp ./ --doc church --sample --fid --timesteps 20 --eta 1 --ni --skip_type uniform

---

Note:

FID scores are computed using the provided reference statistics in the stats/ directory, and are intended for relative comparison under a unified evaluation setting.

When the number of steps increases, it poses a common risk of over-denoising, which is similar to other distillation schemes. In such cases, it is recommended to decrease the $\eta$ parameter to achieve better results.

---

⚙️ Requirements

• Python ≥ 3.9
• PyTorch ≥ 1.6
• Dependencies: torchvision, numpy, tqdm

---

📖 References and Acknowledgements

@inproceedings{chen2025optimizing,

title={Optimizing for the Shortest Path in Denoising Diffusion Model},

author={Chen, Ping and Zhang, Xingpeng and Liu, Zhaoxiang and Hu, Huan and Liu, Xiang and Wang, Kai and Wang, Min and Qian, Yanlin and Lian, Shiguo},

booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},

pages={18021--18030},

year={2025}
}

This implementation is based on / inspired by:

• DDIM PyTorch repo (code structure).
• PyTorch-DDPM repo (accelerated FID evaluation).

---

🔮 Future Directions

We are extending ShortDF to text-to-image and multi-modal tasks. We encourage the community to explore more efficient training strategies based on this shortest-path paradigm.