Unified Personalized Reward Model for Vision Generation

Yibin Wang^1,2, Yuhang Zang⁴, Feng Han^1,2, Yujie Zhou^3,4, Jiazi Bu^3,4,
Cheng Jin^1,2, Jiaqi Wang²

¹Fudan University, ²Shanghai Innovation Institute,
³Shanghai Jiaotong University, ⁴Shanghai AI Lab

Paper UnifiedReward-Flex Pref-GRPO

🤗

🤗

Image Generation Personalized Reasoning

data-overview

Video Generation Personalized Reasoning

data-overview

Reward Model Comparison

pipeline

Text-to-Image Generation GRPO

pipeline

pipeline

Text-to-Video Generation GRPO

pipeline

pipeline

Training Progress Visualization

pipeline

pipeline

BibTeX

Video Comparison

2 guys talking near a big tree, animation style

Wan2.1-T2V-14B

GRPO w/UnifiedReward-Flex

Alien couple performing a massive concert in a violet cyberpunk world, vibrant, psychdellic 4k, 1080p

Wan2.1-T2V-14B

GRPO w/UnifiedReward-Flex

An Iron man is playing the electronic guitar, high electronic guitar

Wan2.1-T2V-14B

GRPO w/UnifiedReward-Flex

Origami dancers in white paper, 3D render, on white background, studio shot, dancing modern dance

Wan2.1-T2V-14B

GRPO w/UnifiedReward-Flex

Robot dancing in Times Square

Wan2.1-T2V-14B

GRPO w/UnifiedReward-Flex

all AI models fighting in mortal kombat

Wan2.1-T2V-14B

GRPO w/UnifiedReward-Flex

human girl talk to cute dragon, pixar, disney

Wan2.1-T2V-14B

GRPO w/UnifiedReward-Flex

unicorn running in the beautiful garden with rainbow

Wan2.1-T2V-14B

GRPO w/UnifiedReward-Flex