Research map / Image generation / Text-to-image diffusion
DeRaDiff: Denoising Time Realignment of Diffusion Models
Ratnavibusena Don Shahain Manujith, Teoh Tze Tzun, Kenji Kawaguchi, Yang 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 |
|---|---|---|---|
| realign a reward-hacked beta=250 model with DeRaDiff toward beta=2000 | SDXL beta=250 · HPSv2+PartiPrompts 1000 prompts | FID: 72.98 -> 50.71 | Improved |
| realign a beta=500 model with DeRaDiff toward beta=2000 | SDXL beta=500 · HPSv2+PartiPrompts 1000 prompts | FID: 61.61 -> 47.13 | Improved |
| realign a beta=750 model with DeRaDiff toward beta=2000 | SDXL beta=750 · HPSv2+PartiPrompts 1000 prompts | FID: 50.04 -> 48.43 | Improved |
| use beta=2000 anchor to approximate beta=500 target (ablation) | DeRaDiff · PickScore | absolute difference (%): 0.0611 -> 1.3451 | Hurt |
| use beta=2000 anchor to approximate larger-beta targets (lambda<=1) | DeRaDiff · PickScore | absolute difference (%): 0.1399 | No clear effect |
| use beta=2000 anchor to approximate larger-beta targets (lambda<=1) | DeRaDiff · CLIP | absolute difference (%): 0.3041 | No clear effect |
| use beta=2000 anchor to approximate larger-beta targets (lambda<=1) | DeRaDiff · HPS | absolute difference (%): 0.2688 | No clear effect |
| extrapolate lambda beyond 1 (ablation) | DeRaDiff · SDXL | image quality | Hurt |
| use lambda in [0,1] interpolation instead of extrapolation (ablation) | DeRaDiff · SDXL | approximation accuracy | Improved |
| align once and use DeRaDiff for N=3 beta values instead of aligning three models | DeRaDiff · SDXL alignment compute | GPU-hour savings: 66.7% | Improved |
| align once and use DeRaDiff for N=10 beta values | DeRaDiff · SDXL alignment compute | GPU-hour savings: 90% | Improved |
Closest papers by meaning
- Training-Free Diffusion Model Alignment with Sampling Demons (ICLR 2025 · similarity 0.88)
- Inference-Time Alignment of Diffusion Models with Direct Noise Optimization (ICML 2025 · similarity 0.88)
- Stitched Value Model for Diffusion Alignment (NeurIPS 2026 · similarity 0.86)
- APT: Adaptive Personalized Training for Diffusion Models with Limited Data (CVPR 2025 · similarity 0.85)
- Decoding-time Realignment of Language Models (ICML 2024 · similarity 0.85)
- A Variational Perspective on Solving Inverse Problems with Diffusion Models (ICLR 2024 · similarity 0.85)
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.
