Research map / Image generation / Fast diffusion sampling
FeRA: Frequency-Energy Constrained Routing for Effective Diffusion Adaptation Fine-Tuning
Bo Yin, Xiaobin Hu, Xingyu Zhou, Yu HE, Peng-Tao Jiang, Yue Liao, Junwei Zhu, Jiangning Zhang, Ying Tai, Shuicheng YAN
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 |
|---|---|---|---|
| replace LoRA with FeRA frequency-energy routed experts | LoRA on SD 2.0 (20M) · Cyberpunk style | FID: 165.17 -> 154.19 | Improved |
| replace LoRA with FeRA | LoRA on FLUX.1 (20M) · Cyberpunk style | FID: 136.43 -> 128.43 | Improved |
| use FeRA instead of full fine-tuning at 20M params | SD 2.0 · Barbie style | FID: 191.00 -> 187.83 | Improved |
| use FeRA instead of SaRA in DreamBooth | SD 2.0 · DreamBooth Dog | CLIP-I: 0.790 -> 0.900 | Improved |
| use FeRA instead of SaRA in DreamBooth | SD 2.0 · DreamBooth Dog | CLIP-T: 25.97 -> 25.95 | No clear effect |
| add frequency-energy indicator (FEI) routing instead of timestep routing (ablation) | hard router · Cyberpunk, SD 2.0 | FID: 138.50 -> 132.70 | Improved |
| add soft router instead of hard routing (ablation) | timestep routing · Cyberpunk, SD 2.0 | FID: 138.50 -> 134.90 | Improved |
| combine FEI and soft router (ablation) | FeRA · Cyberpunk, SD 2.0 | FID: 138.50 -> 126.40 | Improved |
| use one expert per band (3 experts) instead of 2, 4 or 5 (ablation) | FeRA · Cyberpunk, SD 2.0 | FID: 142.8 (2 experts) -> 126.4 (3 experts) | Improved |
| use 3 frequency bands instead of 2, 4 or 5 (ablation) | FeRA FEI · Cyberpunk, SD 2.0 | FID: 148.5 (2 bands) -> 126.4 (3 bands) | Improved |
| add frequency-energy consistency loss with small weight (ablation) | FeRA · Cyberpunk, SD 2.0 | convergence | Improved |
| increase FECL weight to larger values (ablation) | FeRA · Cyberpunk, SD 2.0 | training stability | Hurt |
Closest papers by meaning
- FreqTS: Frequency-Aware Token Selection for Accelerating Diffusion Models (AAAI 2025 · similarity 0.88)
- D-DOIT: Training-free Adaptation of Discrete Diffusion via Doob's h-Transform (NeurIPS 2026 · similarity 0.87)
- Diffusion Tuning: Transferring Diffusion Models via Chain of Forgetting (NeurIPS 2024 · similarity 0.87)
- Bayesian Power Steering: An Effective Approach for Domain Adaptation of Diffusion Models (ICML 2024 · similarity 0.87)
- SADA: Stability-guided Adaptive Diffusion Acceleration (ICML 2025 · similarity 0.87)
- Tortoise and Hare Guidance: Accelerating Diffusion Model Inference with Multirate Integration (NeurIPS 2025 · 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.
