Research map / Large language models / Reasoning and chain of thought
T2I-R1: Reinforcing Image Generation with Collaborative Semantic-level and Token-level CoT
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Applies GRPO with a bi-level chain-of-thought (semantic-level text planning plus token-level patch generation) and an ensemble of vision-expert rewards to Janus-Pro-7B to improve text-to-image generation.
Problem
Text-to-image generation lacks explicit reasoning; unified multimodal models keep understanding and generation decoupled, and image generation has no rule-based reward for RL.
Approach
Identify two CoT levels for autoregressive image generation (semantic-level textual planning before generation and token-level patch-by-patch generation) and jointly optimize both with a single GRPO objective (BiCoT-GRPO), using an averaged ensemble of human-preference, object-detector, and VQA rewards.
Key results
- T2I-R1 reaches 0.8130 color binding on T2I-CompBench vs 0.6359 for Janus-Pro-7B, with an average ~13% gain reported.
- On WISE, T2I-R1 achieves an overall score of 0.54 vs 0.35 for Janus-Pro-7B and 0.50 for FLUX.1-dev.
- Token-level-only GRPO reduces image diversity (Vendi), while adding semantic-level CoT raises it (8.203 vs 6.976 baseline).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add semantic-level CoT to token-level CoT training (ablation) | Janus-Pro-7B · T2I-CompBench | Color: 0.7752 -> 0.8130 | Helped |
| add semantic-level CoT to token-level CoT training (ablation) | Janus-Pro-7B · WISE | Culture: 0.3500 -> 0.5600 | Helped |
| add semantic-level CoT to token-level CoT training (ablation) | Janus-Pro-7B · Diversity | Vendi Score: 6.255 -> 8.203 | Helped |
| add token-level CoT optimization to semantic-level CoT (ablation) | Janus-Pro-7B · T2I-CompBench | Color: 0.8082 -> 0.8130 | Helped |
| add token-level CoT optimization to semantic-level CoT (ablation) | Janus-Pro-7B · WISE | Culture: 0.4900 -> 0.5600 | Helped |
| add token-level CoT optimization to semantic-level CoT (ablation) | Janus-Pro-7B · Diversity | Vendi Score: 8.177 -> 8.203 | No clear effect |
| add ORM reward to the HPM + object detector + VQA ensemble (ablation) | T2I-R1 · T2I-CompBench | Complex: 0.3993 -> 0.3921 | Hurt |
| use HPM + object detector + VQA rewards instead of HPM + object detector rewards (ablation) | Janus-Pro-7B · T2I-CompBench | Complex: 0.4005 -> 0.3993 | No clear effect |
| train with token-level CoT only, without semantic-level CoT (ablation) | Janus-Pro-7B · Diversity | Vendi Score: 6.976 -> 6.255 | Hurt |
Surprising findings
- Zero-shot semantic-level CoT on the baseline Janus-Pro gives only marginal improvement, implying BiCoT-GRPO training is needed for the two CoT levels to cooperate.
- Token-level-only GRPO substantially reduces image diversity.
Limitations
- Extension to video generation is not addressed; no unified understanding-generation video model exists for BiCoT-GRPO.
- Long video inference time makes GRPO impractical for video.
- Reward ensemble choice relies on a small human study (8 students, 30 prompts per subtask).
Experimental setup
Datasets: T2I-CompBench training prompts 6,786 text prompts (no images)
Benchmarks: T2I-CompBench WISE GenAI-Bench TIIF-Bench
Metrics: T2I-CompBench attribute/relation/complex scores WISE overall score GenAI-Bench accuracy TIIF-Bench score Vendi Score diversity human rank-based visual quality
Baselines: Janus-Pro-7B FLUX.1-dev Show-o Show-o + PARM EMU3 SD-XL PixArt-alpha CoMat
Builds on: GRPO DeepSeek-R1 Janus-Pro Image Generation with CoT
Closest papers by meaning
- GoT-R1: Unleashing Reasoning Capability of Autoregressive Visual Generation with Reinforcement Learning (ICLR 2026 · similarity 0.92)
- ThinkGen: Generalized Thinking for Visual Generation (CVPR 2026 · similarity 0.91)
- Delving into RL for Image Generation with CoT: A Study on DPO vs. GRPO (NeurIPS 2025 · similarity 0.91)
- Janus-Pro-R1: Advancing Collaborative Visual Comprehension and Generation via Reinforcement Learning (NeurIPS 2025 · similarity 0.91)
- Let's Verify and Reinforce Image Generation Step by Step (CVPR 2025 · similarity 0.90)
- MindOmni: Unleashing Reasoning Generation in Vision Language Models with RGPO (NeurIPS 2025 · similarity 0.90)
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.