Research map / Image generation / Fast diffusion sampling
TAP: A Token-Adaptive Predictor Framework for Training-Free Diffusion Acceleration
Haowei Zhu, Tingxuan Huang, Xing Wang, Tianyu Zhao, Jiexi Wang, Weifeng Chen, Xurui Peng, Fangmin Chen, Junhai Yong, Bin Wang
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 |
|---|---|---|---|
| make predictor selection token-adaptive instead of a single global Taylor predictor | TaylorSeer · FLUX.1-dev DrawBench (~6.24x) | ImageReward: 0.91 -> 0.99 | Improved |
| enlarge predictor pool in both distance and order (ablation) | TAP · FLUX.1-dev DrawBench (N=7) | ImageReward: 0.95 -> 0.99 | Improved |
| enlarge either prediction distance or order of the pool (ablation) | TAP · FLUX.1-dev DrawBench (N=7) | ImageReward: 0.89 -> 0.95 | Improved |
| include zeroth-order predictors in the pool (ablation) | TAP · FLUX.1-dev DrawBench | ImageReward | Improved |
| shift prediction window left of k instead of right (ablation) | TAP · FLUX.1-dev DrawBench | ImageReward | Improved |
| use finer distance step delta=0.1 instead of 1 (ablation) | TAP · FLUX.1-dev DrawBench | ImageReward: about +0.005 | No clear effect |
| use TAP instead of TaylorSeer | TAP vs TaylorSeer · Qwen-Image 3.57x | PSNR: 18.02 -> 20.13 | Improved |
| use TAP instead of TaylorSeer | TAP vs TaylorSeer · Qwen-Image 3.57x | ImageReward: 1.18 -> 1.23 | Improved |
| use TAP instead of TaylorSeer | TAP vs TaylorSeer · HunyuanVideo VBench | VBench score: 64.89 -> 65.46 | Improved |
| push acceleration from N=8 to N=10 | TAP · FLUX.1-dev DrawBench | ImageReward: 0.99 -> 0.91 | Hurt |
| use TAP on distilled FLUX-schnell | TAP vs TaylorSeer · FLUX.1-schnell | PSNR: 22.67 -> 26.26 | Improved |
Closest papers by meaning
- Token Perturbation Guidance for Diffusion Models (NeurIPS 2025 · similarity 0.90)
- Encoder-Decoder Diffusion Language Models for Efficient Training and Inference (NeurIPS 2025 · similarity 0.89)
- Tortoise and Hare Guidance: Accelerating Diffusion Model Inference with Multirate Integration (NeurIPS 2025 · similarity 0.89)
- Hierarchical Variational Policies for Reward-Guided Diffusion (NeurIPS 2026 · similarity 0.89)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.89)
- Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion (NeurIPS 2024 · 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.
