AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Reward Sharpness-Aware Fine-Tuning for Diffusion Models

Kwanyoung Kim, Byeongsu Sim

CVPR 2026 · Reinforcement learning › RL from feedback · approach: Theory of RLHF · 2 citations (OpenAlex)

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 changedWhereMeasuredEffect
add RSA-FT (input and weight perturbation) to AlignPropAlignProp (SD1.5) · HPDHPSv2.1 / ImageReward: HPSv2.1 24.93 -> 32.02; ImageReward 0.032 -> 0.528Improved
add RSA-FT to Draft-LVDraft-LV (SD1.5) · DrawBenchHPSv2.1 / PickScore: HPSv2.1 25.59 -> 26.67; PickScore 20.96 -> 21.12Improved
add RSA-FT to ReFLReFL (SD1.5) · HPDHPSv2.1: 34.95 -> 35.81Improved
add RSA-FT to DRTuneDRTune (SD1.5) · HPDPickScore: 21.91 -> 21.92No clear effect
add RSA-FT to ReFL on SDXLReFL (SDXL) · HPDHPSv2.1: 31.21 -> 32.66Improved
add RSA-FT to DRTune on SDXLDRTune (SDXL) · HPDHPSv2.1: 32.36 -> 33.74Improved
add RSA-FT to ReFL on SD3ReFL (SD3) · DrawBenchHPSv2.1: 29.45 -> 29.81Improved
add RSA-FT on SD3 with DRTuneDRTune (SD3) · HPDPickScore: 22.33 -> 22.41No clear effect
combine image-space and parameter-space flattening versus each alone (ablation)RSA-FT · SD1.5 RDRLreward metricsImproved
higher reward sharpness during fine-tuningDRaFT-K SD1.5 with HPSv2 · training dynamicsPickScore / ImageReward correlation: Pearson r = -0.802 (PickScore), -0.669 (ImageReward)Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy