Research map / Image generation / Fast diffusion sampling
TC-Pade: Trajectory-Consistent Pade Approximation for Diffusion Acceleration
Shaoxuan He, Benlei Cui, Bukun Huang, Zhizeng Ye, Yunyun Sun, Longtao Huang, Hui Xue, Yang Yang, Haiwen Hong, Jingqun Tang, Zhou Zhao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| cache residuals across the entire block instead of double-stream blocks only (ablation) | TC-Padé · DrawBench, FLUX.1-dev | ImageReward: 0.7921 -> 0.9184 | Improved |
| cache residuals across the entire block instead of single-stream blocks only (ablation) | TC-Padé · DrawBench, FLUX.1-dev | ImageReward: 0.8717 -> 0.9184 | Improved |
| cache residuals across the entire block instead of double-stream only (ablation) | TC-Padé · DrawBench, FLUX.1-dev | latency speedup: 1.36× -> 2.88× | Improved |
| lower TSI threshold from 1.3 to 0.7 (more aggressive skipping) (ablation) | TC-Padé · DrawBench, FLUX.1-dev | speedup: 1.63× -> 2.88× | Improved |
| lower TSI threshold from 1.3 to 0.7 (more aggressive skipping) (ablation) | TC-Padé · DrawBench, FLUX.1-dev | ImageReward: 0.9562 -> 0.9184 | Hurt |
| predict residuals with Padé instead of raw features with Taylor series (TC-Padé vs TaylorSeer) | TC-Padé · COCO 2017, FLUX.1-dev, 20 steps | PSNR: 16.567 -> 21.962 | Improved |
| predict residuals with Padé instead of Taylor extrapolation on video | TC-Padé · Wan2.1-1.3B, VBench-2.0 | VBench-2.0 score (%): 54.50% -> 60.38% | Improved |
| apply TC-Padé to class-conditional DiT-XL/2 | DiT-XL/2 · ImageNet 256x256, 20 steps | FID: 3.56 -> 6.93 | Hurt |
| apply TC-Padé acceleration to FLUX.1-dev | FLUX.1-dev · COCO 2017, 20 steps | FID: 23.38 -> 24.14 | Hurt |
| add quantization on top of TC-Padé | TC-Padé on FLUX.1-dev · COCO 2017, 20 steps | CLIP score: 31.82 -> 31.08 | Hurt |
| add quantization on top of TC-Padé | TC-Padé on FLUX.1-dev · COCO 2017, 20 steps | FID: 24.14 -> 24.31 | No clear effect |
Closest papers by meaning
- HiCache: A Plug-in Scaled-Hermite Upgrade for Taylor-Style Cache-then-Forecast Diffusion Acceleration (ICLR 2026 · similarity 0.90)
- IdealCache: Rethinking Cache Scheduling in Diffusion Transformers via Ideal Trajectories (NeurIPS 2026 · similarity 0.90)
- Denoising as Path Planning: Training-Free Acceleration of Diffusion Models with DPCache (CVPR 2026 · similarity 0.90)
- Map-Guided Caching: A Global Perspective for Efficient Diffusion Transformer (NeurIPS 2026 · similarity 0.90)
- Beyond Fixed Formulas: Data-Driven Linear Predictor for Efficient Diffusion Models (CVPR 2026 · similarity 0.89)
- Accelerating Rectified Flow Models via Trajectory-Aware Caching (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
