ThinkGen: Generalized Thinking for Visual Generation

TL;DR

ThinkGen presents a think-driven approach to visual generation that explicitly uses Chain-of-Thought reasoning from a Multimodal Large Language Model to guide Diffusion Transformer synthesis. A Visual Generation Instruction Refinement module extracts concise, task-specific instructions, enabling effective conditioning for the DiT. The separable GRPO (SepGRPO) training paradigm alternates reinforcement learning between the MLLM and DiT across multiple generation scenarios, improving both instruction quality and image fidelity. Across diverse benchmarks, ThinkGen achieves state-of-the-art results, particularly in reasoning generation, text rendering, and image editing, demonstrating strong generalization and practical impact for versatile, reasoning-guided generation systems.

Abstract

Recent progress in Multimodal Large Language Models (MLLMs) demonstrates that Chain-of-Thought (CoT) reasoning enables systematic solutions to complex understanding tasks. However, its extension to generation tasks remains nascent and limited by scenario-specific mechanisms that hinder generalization and adaptation. In this work, we present ThinkGen, the first think-driven visual generation framework that explicitly leverages MLLM's CoT reasoning in various generation scenarios. ThinkGen employs a decoupled architecture comprising a pretrained MLLM and a Diffusion Transformer (DiT), wherein the MLLM generates tailored instructions based on user intent, and DiT produces high-quality images guided by these instructions. We further propose a separable GRPO-based training paradigm (SepGRPO), alternating reinforcement learning between the MLLM and DiT modules. This flexible design enables joint training across diverse datasets, facilitating effective CoT reasoning for a wide range of generative scenarios. Extensive experiments demonstrate that ThinkGen achieves robust, state-of-the-art performance across multiple generation benchmarks. Code is available: https://github.com/jiaosiyuu/ThinkGen

Figures (9)

• Figure 1: Comparison between BAGEL deng2025emerging and our ThinkGen. ThinkGen achieves superior performance when adopting CoT reasoning (w/ think) across a wide range of generation scenarios.
• Figure 2: ThinkGen enables think-driven generation across a wide range of scenarios, including text-to-image generation, text rendering, image editing, reasoning generation, reasoning editing, and reflection.
• Figure 3: Overview of ThinkGen. Within ThinkGen, the MLLM and DiT architectures are decoupled for autoregressive CoT generation and diffusion-based image generation. The MLLM receives text/images as input and outputs generation instructions tailored to the preferences of the DiT. Through a process called visual generation instruction refinement (VGI-Refine), the hidden states corresponding to these instructions are extracted and concatenated with Prepadding States, forming the conditional information for DiT's image generation. For clarity, we omit the text encoder and vision encoder components within MLLM and DiT.
• Figure 4: The training recipe of ThinkGen consists of three supervised pre-training stages: Connector training (stage 1), Pre-training (stage 2), and High-quality fine-tuning (stage 3), as well as two SepGRPO stages: MLLM-GRPO (stage 4) and DiT-GRPO (stage 5).
• Figure 5: Visualization of the MLLM-GRPO process, the reward score steadily increases.