Visual Jev rewards improve multi-subject image generation verification
Visual Jev Rewards: Reference-Bound Verification for Multi-Subject Image Generation
Computer Vision and Pattern RecognitionArtificial IntelligenceMachine Learning
Summary
Generating images with multiple specific people doing certain actions or having relationships is tricky. The authors created a way to check if the generated image truly matches the requested subjects and their interactions. They trained a model to give yes/no answers about whether the right people are doing the right things together, which helps improve image generation. Their method raised performance scores on a test set but did not show clear gains over simpler scoring methods in human evaluations. This work offers an initial approach to better verify complex image content during generation.
What this means in practice
- •For image generation engineers: Use Visual Jev rewards to train models that better verify multiple subjects interacting correctly in generated images.
- •For advertising content teams: Improve automated generation tools to create images with specified people and interactions matching campaign requests.
Authors
Baoteng Li, Wenzhuo Wu, Kongming Liang, Zhanyu Ma
Abstract
Multi-subject image generation requires rewards that verify whether requested attributes, actions, and relations hold for the specified reference subjects. Subject presence alone does not establish that the correct subjects participate in a requested interaction. We present reference-bound Visual Jev rewards that turn these visual decisions into generator training signals. Each subject-related question receives a positive label only when the requested condition and the relevant reference identities hold jointly. We construct fixed questions offline, train a Qwen3.5-4B verifier with binary supervision, and directly read Yes probabilities from its language-model head. Their mean supplies a GRPO reward while retaining individual judgments for inspection. Using 200 MICo-150K training tasks and 30 updates, the framework raises a GPT-5.4 composite score from 41.78 to 52.50 on a manually selected 897-task MICo-Bench subset; direct 27B rewards yield 51.84. Each reward is tested in one GRPO run, and offline human evaluation does not establish a statistically significant advantage over direct scoring. The study provides an initial implementation and evaluation of Visual Jev as a reference-bound reward for multi-subject image generation.