Research map / Vision-language and multimodal / Multimodal LLMs

ThinkGen: Generalized Thinking for Visual Generation

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

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.

Proposes ThinkGen, a decoupled MLLM (Qwen3-VL-8B-Think) plus DiT (OmniGen2-DiT-4B) framework for think-driven visual generation, trained with supervised pretraining and a separable GRPO (SepGRPO) that alternates RL on the MLLM and DiT.

Problem

Chain-of-thought reasoning for visual generation is typically tailored to a single scenario and fails to generalize across diverse generation and editing tasks, requiring manual activation.

Approach

An MLLM writes CoT rewrites of user intent, the post-</think> hidden states (refined by learnable prepadding states) condition a DiT, and the two modules are trained with alternating GRPO (MLLM first with DiT frozen, then DiT with MLLM frozen) using multi-scenario rule-based rewards.

Key results

• With CoT reasoning, ThinkGen reaches WISE 0.76 (vs 0.55 without CoT) and RISEBench 13.0 (vs 3.6 without), a new state of the art on WISEBench.
• ThinkGen* achieves GenEval 0.89, CVTG word accuracy 0.84, and ImgEdit 4.21 overall, comparable to GPT-4o's 4.20 on ImgEdit.
• Training the MLLM with MLLM-GRPO on 10K reasoning data lifts WISE from 0.55 to 0.74, whereas SFT on the same data does not improve WISE (0.58).

Surprising findings

• Directly applying SFT to DiT with reasoning data gives little reasoning gain (WISE 0.55 to 0.58), while RL on the MLLM gives large gains, suggesting the DiT lacks transferable world knowledge.
• Stage 4 MLLM-GRPO without CoT at inference slightly reduces GenEval (0.88 to 0.86) and WISE, and the CoT-enabled gains are what drive improvement.

Limitations

• Reported text-rendering and reasoning-editing gains are uneven; logical reasoning on RISEBench stays at 1.1 with CoT.
• Pseudo-CoT empty think blocks in supervised pretraining provide no real reasoning traces.

Experimental setup

Datasets: ShareGPT-4o-Image SFHQ FLUX-Reason-6M DenseFusion UltraEdit OmniConsistency Echo4o GPT-Image-Edit X2Edit NHR Nano-banana-150k Echo-4o-Image

Benchmarks: WISEBench RISEBench GenEval DPG-Bench CVTG ImgEdit

Metrics: WISE score RISEBench accuracy GenEval overall DPG overall CVTG word accuracy CVTG NED ImgEdit overall score

Baselines: BAGEL OmniGen2 STAR MetaQuery-XL BLIP3-o-8B Janus-Pro FLUX.1-dev GPT-4o

Builds on: GRPO Flow-GRPO BAGEL OmniGen2 Qwen3-VL

Closest papers by meaning

• Thinking-while-Generating: Interleaving Textual Reasoning throughout Visual Generation (CVPR 2026 · similarity 0.92)
• T2I-R1: Reinforcing Image Generation with Collaborative Semantic-level and Token-level CoT (NeurIPS 2025 · similarity 0.91)
• Janus-Pro-R1: Advancing Collaborative Visual Comprehension and Generation via Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
• MindOmni: Unleashing Reasoning Generation in Vision Language Models with RGPO (NeurIPS 2025 · similarity 0.90)
• GoT-R1: Unleashing Reasoning Capability of Autoregressive Visual Generation with Reinforcement Learning (ICLR 2026 · similarity 0.89)
• DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models (ICML 2026 · similarity 0.89)