TC-Padé: Trajectory-Consistent Padé Approximation for Diffusion Acceleration

Official implementation of TC-Padé (CVPR 2026) 🎉

To address the computational inefficiency of diffusion models and the limitations of existing polynomial-based feature caching methods that suffer from error accumulation in practical low-step regimes, this paper proposes Trajectory-Consistent Padé approximation (TC-Padé). By modeling feature evolution through rational functions rather than Taylor series, TC-Padé captures complex asymptotic behaviors more accurately and incorporates adaptive coefficient modulation alongside step-aware prediction strategies to handle distinct denoising phases.

🔧 Installation

git clone https://github.com/Alibaba-Yufeng/TC_Pade.git
cd TC_Pade
pip install -r requirements.txt

Requirements: Python >= 3.9, PyTorch >= 2.6, CUDA-capable GPU.

🚀 Usage

Baseline

python run.py --model_path /path/to/flux.1-dev --num_inference_steps 50

TC-Padé Accelerated Inference

python run.py \
    --model_path /path/to/flux.1-dev \
    --use_predict \
    --num_inference_steps 50 \
    --N 1.4 \
    --interval 8

Argument List

Argument	Default	Description
--model_path	path_to_flux.1-dev	Path to the pretrained FLUX model
--prompts_file	./example_prompts.json	Path to the prompts JSON file
--output_dir	auto-generated	Output directory for generated images
--num_inference_steps	50	Number of denoising steps
--seed	42	Random seed for reproducibility
--use_predict	False	Enable TC-Padé acceleration
--start_step	4	Step to begin prediction
--interval	8	Prediction interval
--N	1.4	Curvature threshold (larger = faster, more aggressive skipping)
--predictor_order	3	Padé predictor order
--predictor_history_size	6	Residual history buffer size

📄 Citation

If you find this work useful, please cite:

@article{cui2026tc,
  title={TC-Pad$\backslash$'e: Trajectory-Consistent Pad$\backslash$'e Approximation for Diffusion Acceleration},
  author={Cui, Benlei and He, Shaoxuan and Huang, Bukun and Ye, Zhizeng and Sun, Yunyun and Huang, Longtao and Xue, Hui and Yang, Yang and Tang, Jingqun and Zhao, Zhou and others},
  journal={arXiv preprint arXiv:2603.02943},
  year={2026}
}