Research map / Image generation / Fast diffusion sampling
FADE: Fractional Anomalous Dynamics Extrapolation for Training-Free Diffusion Transformers Acceleration
Jinlong Yang, Jinke Wu, Lizilin, Yao Zhou
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FADE, a training-free feature-caching method for accelerating Diffusion Transformer sampling by forecasting features with fractional-order (anomalous dynamics) extrapolation instead of integer-order operators.
Problem
Iterative sampling in Diffusion Transformers is computationally expensive; existing feature-caching methods that reuse or forecast features with integer-order operators adapt poorly to the feature dynamics across denoising steps.
Approach
Replaces integer-order difference/extrapolation operators used in feature-caching forecasters with fractional-order extrapolation modeling anomalous (non-integer-order) feature dynamics, requiring no retraining.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use stage-aware LUT-calibrated fractional order instead of integer order gamma=1 (ablation) | FADE · DiT-XL/2 FID, N=6 | FID: 3.17 -> 2.87 | Improved |
| use fixed global fractional order 0.85 instead of gamma=1 (ablation) | FADE · DiT-XL/2 FID, N=6 | FID: 3.17 -> 3.06 | Improved |
| use stage-aware LUT instead of polynomial fit of gamma(t) (ablation) | FADE · DiT-XL/2 FID, N=6 | FID: 3.00 -> 2.87 | Improved |
| calibrate with directional cosine loss instead of MSE (ablation) | FADE · DiT-XL/2 FID, N=6 | FID: 2.97 -> 2.87 | Improved |
| use FADE instead of HiCache at same cost | FADE · DiT-XL/2 FID, N=4 | FID: 2.51 -> 2.41 | Improved |
| use FADE instead of FORA at N=6 | FADE · DiT-XL/2 FID, N=6 | FID: 9.22 -> 2.87 | Improved |
| increase calibration set size from 1 to 10 samples (ablation) | FADE · FLUX.1 ImageReward, N=6 | ImageReward: 0.9963 -> 1.0019 | Improved |
| increase calibration set size from 10 to 20 (ablation) | FADE · FLUX.1 ImageReward, N=6 | ImageReward: 1.0019 -> 1.0008 | No clear effect |
| add global refinement factor lambda=1.3 (ablation) | FADE · FLUX.1 ImageReward, N=5 | ImageReward: 0.9974 -> 1.0071 | Improved |
| increase lambda from 1.3 to 1.4 (ablation) | FADE · FLUX.1 ImageReward, N=5 | ImageReward: 1.0071 -> 0.9858 | Hurt |
| use FADE instead of HiCache at N=7 m=1 | FADE · FLUX.1 PSNR, N=7 | PSNR: 18.2778 (HiCache) vs 18.1925 (FADE) | Mixed |
| use FADE instead of TaylorSeer on HunyuanVideo | FADE · HunyuanVideo VBench, N=6 | VBench: 79.87 -> 80.12 | Improved |
Experimental setup
Builds on: feature caching for diffusion transformers
Closest papers by meaning
- BlockDance: Reuse Structurally Similar Spatio-Temporal Features to Accelerate Diffusion Transformers (CVPR 2025 · similarity 0.89)
- Tracing the Roots: Leveraging Temporal Dynamics in Diffusion Trajectories for Origin Attribution (NeurIPS 2025 · similarity 0.88)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.88)
- HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025 · similarity 0.88)
- Learning to Solve Generative ODEs Beyond the Linear Span (NeurIPS 2026 · similarity 0.88)
- Why DiT Models Underperform as Representation Learners without Long Skip Connections (NeurIPS 2026 · similarity 0.87)
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.
