Research map / Image generation / Fast diffusion sampling
Relational Feature Caching for Accelerating Diffusion Transformers
Byunggwan Son, Jeimin Jeon, Jeongwoo Choi, Bumsub Ham
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 relational feature estimation (RFE) to TaylorSeer (ablation) | TaylorSeer · ImageNet DiT-XL/2 NFC 11 | sFID: 5.85 -> 5.22 | Improved |
| add relational cache scheduling (RCS) to TaylorSeer (ablation) | TaylorSeer · ImageNet DiT-XL/2 NFC 11 | sFID: 5.85 -> 5.21 | Improved |
| combine RFE and RCS (ablation) | RFE only · ImageNet DiT-XL/2 NFC 11 | sFID: 5.22 -> 4.88 | Improved |
| replace linear extrapolation with RFE for output change estimate (ablation) | caching forecast · ImageNet DiT-XL/2 NFC 14 | sFID2FC: 3.40 -> 2.02 | Improved |
| use RFE instead of fixed scaling w(t)=1.0 (ablation) | caching forecast · ImageNet DiT-XL/2 NFC 11 | FID2FC: 0.73 -> 0.55 | Improved |
| schedule by input prediction error (RCS) instead of input distance (ablation) | cache scheduling · ImageNet DiT-XL/2 NFC 14 | sFID2FC: 2.53 -> 1.88 | Improved |
| sum input prediction errors over all modules instead of first module only (ablation) | RCS · ImageNet DiT-XL/2 NFC 11 | FID2FC: 0.62 -> 0.61 | No clear effect |
| skip full computation more often in shallow layers (ablation) | RFC · ImageNet DiT-XL/2 NFC 8 (m=2) | FID: 3.40 -> 11.70 | Hurt |
| skip full computation more often in deep layers (ablation) | RFC · ImageNet DiT-XL/2 NFC 8 (m=2) | FID: 3.40 -> 10.20 | Hurt |
| use RFC instead of TaylorSeer under large interval | TaylorSeer · ImageNet DiT-XL/2 NFC 7 | FID: 4.90 -> 3.40 | Improved |
| use RFC instead of TaylorSeer on distilled FLUX.1 schnell, 3 NFC | TaylorSeer · FLUX.1 schnell DrawBench | PSNR: 22.2134 -> 27.1185 | Improved |
| use RFC instead of TaylorSeer on LSUN-bedroom U-Net, NFC 10 | TaylorSeer · DDIM LSUN-bedroom | FID2FC: 37.81 -> 11.64 | Improved |
Closest papers by meaning
- ProCache: Constraint-Aware Feature Caching with Selective Computation for Diffusion Transformer Acceleration (AAAI 2026 · similarity 0.92)
- Accelerating Diffusion Transformers with Token-wise Feature Caching (ICLR 2025 · similarity 0.91)
- Forecast Then Calibrate: Feature Caching as ODE for Efficient Diffusion Transformers (AAAI 2026 · similarity 0.91)
- Let Features Decide Their Own Solvers: Hybrid Feature Caching for Diffusion Transformers (ICLR 2026 · similarity 0.91)
- Learning-to-Cache: Accelerating Diffusion Transformer via Layer Caching (NeurIPS 2024 · similarity 0.90)
- HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025 · similarity 0.90)
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.
