Research map / Image generation / Fast diffusion sampling
From Reusing to Forecasting: Accelerating Diffusion Models with TaylorSeers
Jiacheng Liu, Chang Zou, Yuanhuiyi Lyu, Junjie Chen, Linfeng Zhang
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 |
|---|---|---|---|
| add first-order Taylor forecasting (O=1) instead of direct reuse (O=0) (ablation) | TaylorSeer N=7 · DiT-XL/2 ImageNet | FID: 12.67 -> 4.29 | Improved |
| raise Taylor order from 1 to 2 (ablation) | TaylorSeer N=7 · DiT-XL/2 ImageNet | FID: 4.29 -> 3.66 | Improved |
| raise Taylor order from 2 to 3 (ablation) | TaylorSeer N=7 · DiT-XL/2 ImageNet | FID: 3.66 -> 3.63 | Improved |
| raise Taylor order from 3 to 4 (ablation) | TaylorSeer N=3 · DiT-XL/2 ImageNet | FID: 2.34 -> 2.35 | No clear effect |
| add Taylor forecasting (O=1) instead of direct reuse (ablation) | TaylorSeer N=3 · DiT-XL/2 ImageNet | FID: 3.55 -> 2.49 | Improved |
| add Taylor forecasting (O=1) instead of direct reuse (ablation) | TaylorSeer N=5 · DiT-XL/2 ImageNet | FID: 6.58 -> 3.05 | Improved |
| lengthen caching interval from N=3 to N=7 with O=3 (ablation) | TaylorSeer · DiT-XL/2 ImageNet | FID: 2.34 -> 3.63 | Hurt |
| replace direct reuse caching with TaylorSeer at about 4.5x speedup | FORA · DiT-XL/2 ImageNet | FID: 6.58 -> 2.65 | Improved |
| replace TeaCache with TaylorSeer at about 5x speedup | TeaCache · HunyuanVideo VBench | VBench score: 79.36 -> 79.93 | Improved |
| use TaylorSeer at low speedup ratio | FLUX.1-dev · DrawBench | ImageReward: 0.9898 -> 1.0181 | Improved |
| use TaylorSeer at low speedup ratio | HunyuanVideo · VBench | VBench score: 80.66 -> 80.74 | Improved |
| replace TeaCache with TaylorSeer at about 4x speedup | TeaCache · HiDream ImageReward | ImageReward: 0.9849 -> 1.0833 | Improved |
Closest papers by meaning
- Forecast Then Calibrate: Feature Caching as ODE for Efficient Diffusion Transformers (AAAI 2026 · similarity 0.93)
- Accelerating Diffusion Transformers with Token-wise Feature Caching (ICLR 2025 · similarity 0.92)
- HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025 · similarity 0.91)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.90)
- BlockDance: Reuse Structurally Similar Spatio-Temporal Features to Accelerate Diffusion Transformers (CVPR 2025 · similarity 0.89)
- FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality (ICLR 2025 · similarity 0.89)
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.
