AI Paper Atlas

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

Benchmarking and Improving Fine-Grained Text-to-Image Alignment via Paired Reinforcement Learning

Kaihang Pan, Wendong Bu, Yuruo Wu, Kai Shen, Yang Wu, Yun Zhu, Zehan Wang, Yunfei Li, Hang Zhao, Juncheng Li, Siliang Tang

ICML 2026 · Image generation › Text-to-image diffusion · approach: T2I evaluation and controllable adapters

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
train with FocusDiff (paired data + Pair-GRPO) on X-OmniX-Omni · DeltaBench averagesa: 61.8 -> 76.1Improved
train with FocusDiff on X-OmniX-Omni · DeltaBench averagesa-sg stability gap: 8.6 -> 4.3 pointsImproved
expand group to include images from both paired prompts (ablation)Pair-GRPO · DeltaBench averagesa / sg: 73.5 -> 76.1 sa; 67.9 -> 71.8 sgImproved
provide ground-truth images with annealed probability during RL (ablation)Pair-GRPO · DeltaBench averagesa / sg: 75.1 -> 76.1 sa; 70.0 -> 71.8 sgImproved
use Pair-GRPO instead of vanilla GRPO (ablation)GRPO · DeltaBench averagesa / sg: 71.9 -> 76.1 sa; 66.0 -> 71.8 sgImproved
train on FocusDiff-Data instead of commonly-used prompts (ablation)vanilla GRPO · DeltaBench averagesa / sg: 63.0 -> 71.9 sa; 55.4 -> 66.0 sgImproved
add FocusDiff training to X-OmniX-Omni · GenEvaloverall: 0.83 -> 0.90Improved
add FocusDiff training to X-OmniX-Omni · DPG-Benchaverage: 87.65 -> 88.35Improved
use FocusDiff-trained model under extended sampling (2 to 8 images per prompt)X-Omni-FocusDiff · DeltaBenchgeometric mean: 71.8 -> 69.8No clear effect
use Pair-GRPO instead of vanilla GRPO for training stability (ablation)GRPO · RL training rewardreward trendImproved
FocusDiff-trained model versus Qwen-Image on GenEvalQwen-Image · GenEvaloverall: 0.91 vs 0.90Hurt

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