AI Paper Atlas

Research map / Image generation / Text-to-image diffusion

Deep Reward Supervisions for Tuning Text-to-Image Diffusion Models

Xiaoshi Wu, Yiming Hao, Manyuan Zhang, Keqiang Sun, Zhaoyang Huang, Guanglu Song, Yu Liu, Hongsheng Li

ECCV 2024 · Image generation › Text-to-image diffusion · approach: Inference-time T2I improvement · 27 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
supervise early denoising steps with DRTune instead of only last stepsStable Diffusion 1.5 · Symmetry rewardsymmetry error: 0.7250 (DRaFT-10) -> 0.0551Improved
add stop gradient on denoising network input (ablation)DRaFT-T · reward training convergenceconvergenceImproved
add stop gradient on denoising network input (ablation)AlignProp · reward training convergenceiterations to reach similar reward: roughly 50% of iterationsImproved
increase training steps K beyond 15 without stop gradient (ablation)DRTune variants · gradient norm of LoRA parametersgradient normHurt
choose K = 0.1T training steps (ablation)DRTune · T from 20 to 50final rewardImproved
choose early stop m = 0.4T (ablation)DRTune · various Tfinal rewardImproved
use equally spaced training timesteps instead of random subset (ablation)DRTune · reward trainingconvergence speedImproved
fine-tune SDXL 1.0 with DRTune on HPS v2.1SDXL 1.0 · user study (400 prompts)win rate vs SDXL: 69% win rateImproved

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