AI Paper Atlas

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

PSR: Scaling Multi-Subject Personalized Image Generation with Pairwise Subject-Consistency Rewards

Shulei Wang, Longhui Wei, Xin He, Jianbo Ouyang, Hui Lu, Zhou Zhao, Qi Tian

CVPR 2026 · Image generation › Text-to-image diffusion · approach: Multi-subject personalization · 8 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 PSR reinforcement learning stage after SFT (ablation)Ours-SFT · PSRBench overallsubject consistency: 0.559 -> 0.673Improved
add PSR reinforcement learning stage after SFT (ablation)Ours-SFT · PSRBench overallaesthetic preference (HPSv3): 0.794 -> 1.124Improved
add PSR reinforcement learning stage after SFT (ablation)Ours-SFT · PSRBench overallsemantic alignment: 0.712 -> 0.783Improved
SFT at 512 resolution on multi-subject dataFLUX.1 Kontext · PSRBench overallaesthetic preference (HPSv3): 0.870 -> 0.794Hurt
use frame-wise positional offset instead of h-w offset (ablation)FLUX.1 Kontext multi-image SFT · PSRBench ablation (Position)semantic alignment: 0.469 -> 0.508Improved
use frame-wise positional offset instead of width-only offset (ablation)FLUX.1 Kontext multi-image SFT · PSRBench ablation (Position)semantic alignment: 0.459 -> 0.508Improved
use frame-wise positional offset instead of height-only offset (ablation)FLUX.1 Kontext multi-image SFT · PSRBench ablation (Position)semantic alignment: 0.437 -> 0.508Improved
use frame-wise positional offset instead of h-w offset (ablation)FLUX.1 Kontext multi-image SFT · PSRBench ablation (3 subjects)semantic alignment: 0.831 -> 0.870Improved
use frame-wise positional offset instead of h-w offset (ablation)FLUX.1 Kontext multi-image SFT · PSRBench ablation (4 subjects)semantic alignment: 0.808 -> 0.821Improved
use frame-wise positional offset instead of h-w offset (ablation)FLUX.1 Kontext multi-image SFT · PSRBench ablation (2 subjects)semantic alignment: 0.929 -> 0.922No clear effect
combine PSR with semantic and aesthetic rewards (ablation)PSR-only RL · PSRBench qualitativereward hackingImproved
PSR-trained model vs strongest DreamBench baselineOurs (PSR) · DreamBench multi-IPDINO: 0.506 -> 0.529Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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