💡🎞️ Awesome_Image_Generation_with_Thinking

Image Generation with Thinking.

Welcome to the Awesome-Image-Generation-with-Thinking repository! This repository represents a comprehensive collection of research focused on empowering models to think during image generation. We explore current works and summarize them into three approaches: explicit reflection, reinforcement learning, and unified multimodal models.

---

🔔 News

• [2025-06] We created this repository to maintain a paper list on Awesome-Image-Generation-With-Thinking. Contributions are welcome!

---

📜 Table of Contents

• ✍️ Survey
• 🗒️ Explicit Reflection
• 🧠 Reinforcement Learning
• 🚀 Unified LMMs
• 📚 Benchmarks

---

📖 Survey

• Delving into RL for image generation with CoT: A study on DPO vs. GRPO (May, 2025)
  

🧠 Reinforcement Learning

Reinforcement learning has been proven to be a crucial step in enhancing reasoning capabilities. Here, we summarize methods that utilize reinforcement learning, such as GRPO, into image generation process.

• Can we generate images with CoT? Let’s verify and reinforce image generation step by step (Jan., 2025)
  

• SimpleAR: Pushing the frontier of autoregressive visual generation through pretraining, SFT, and RL (Apr., 2025)
  

• T2I-R1: Reinforcing image generation with collaborative semantic-level and token-level CoT (May, 2025)
  

• Flow-GRPO: Training flow matching models via online RL (May, 2025)
  

• DanceGRPO: Unleashing GRPO on visual generation (May, 2025)
  

• GoT-R1: Unleashing reasoning capability of MLLM for visual generation with reinforcement learning (May, 2025)
  

• Co-Reinforcement learning for unified multimodal understanding and generation (May, 2025)
  

• ReasonGen-R1: CoT for autoregressive image generation model through SFT and RL (May, 2025)
  

🗒️ Explicit Reflection

Reflection is an essantial step in thinking processes. Explicit reflection, which leverages modalities such as text, object coordinates, and image with editing instructions, is a typical approach.

• Visual programming: Compositional visual reasoning without training (CVPR, 2023)
  

• ViperGPT: Visual inference via python execution for reasoning (ICCV, 2023)
  

• From reflection to perfection: Scaling inference-time optimization for text-to-image diffusion models via reflection tuning (Apr., 2025)
  

• ImageGen-CoT: Enhancing text-to-image in-context learning with chain-of-thought reasoning (Jan., 2025)
  

• GoT: Unleashing reasoning capability of multimodal large language model for visual generation and editing (Mar., 2025)
  

• Visual planning: Let's think only with images (Mar., 2025)
  

---

🚀 Unified LMMs

Unified LMMs inherently excel at text-to-image controllability, hence we collect a list of relevant works.

• Multi-modal generation via cross-modal in-context learning (May, 2024)
  

• Emu: Generative pretraining in multimodality (ICLR, 2024)
  

• DreamLLM: Synergistic multimodal comprehension and creation (ICLR, 2024)
  

• Making LLaMA see and draw with seed tokenizer (ICLR, 2024)
  

• MiniGPT-5: Interleaved vision-and-language generation via generative vokens (Mar., 2024)
  

• Generative multimodal models are in-context learners (CVPR, 2024)
  

• Unified-IO 2: Scaling autoregressive multimodal models with vision, language, audio, and action (CVPR, 2024)
  

• SEED-X: Multimodal models with unified multi-granularity comprehension and generation (May, 2025)
  

• Chameleon: Mixed-Modal early-fusion foundation models (May, 2025)
  

• Transfusion: Predict the next token and diffuse images with one multi-modal model (Aug., 2024)
  

• Show-o: One single transformer to unify multimodal understanding and generation (ICLR, 2025)
  

• VILA-U: A unified foundation model integrating visual understanding and generation (Mar., 2025)
  

• Emu3: Next-token prediction is all you need (Sep., 2024)
  

• Janus: Decoupling visual encoding for unified multimodal understanding and generation (Oct., 2024)
  

• JanusFlow: Harmonizing autoregression and rectified flow for unified multimodal understanding and generation (Mar., 2025)
  

• TokenFlow: Unified image tokenizer for multimodal understanding and generation (CVPR, 2025)
  

• MetaMorph: Multimodal understanding and generation via instruction tuning (Dec., 2024)
  

• LMFusion: Adapting pretrained language models for multimodal generation (Feb., 2025)
  

• Janus-Pro: Unified multimodal understanding and generation with data and model scaling (Jan., 2025)
  

• MINT: Multi-modal chain of thought in unified generative models for enhanced image generation (Mar., 2025)
  

• Transfer between modalities with metaqueries (Apr., 2025)
  

• BLIP3-o: A family of fully open unified multimodal models-architecture, training and dataset (May, 2025)
  

• Emerging properties in unified multimodal pretraining (May, 2025)
  

• Thinking with generated images (May, 2025)
  

• Show-o2: Improved native unified multimodal models (Jun., 2025)
  

• ShareGPT-4o-Image: Aligning multimodal models with GPT-4o-level image generation (Jun., 2025)
  

• OmniGen2: Exploration to advanced multimodal generation (Jun., 2025)
  

• Ovis-U1: Unified Understanding, Generation, and Editing (Jun., 2025)
  

---

📚 Benchmarks

Essential resources for understanding the broader landscape and evaluating progress in visual reasoning.

• ELLA: Equip diffusion models with LLM for enhanced semantic alignment (Mar., 2024)
  

• T2I-CompBench: A comprehensive benchmark for open-world compositional text-to-image generation (NeurIPS, 2023)
  

• GenEval: An object-focused framework for evaluating text-to-image alignment (NeurIPS, 2023)
  

• Commonsense-T2I challenge: Can text-to-image generation models understand commonsense? (COLM, 2024)
  

• WISE: A world knowledge-informed semantic evaluation for text-to-image generation (May, 2025)
  

• TIIF-Bench: How does your T2I model follow your instructions? (Jun., 2025)
  

• OneIG-Bench: Omni-dimensional nuanced evaluation for image generation (Jun., 2025)
  

Star History