Research map / Vision-language and multimodal / Multimodal LLMs

MM-R1: Unleashing the Power of Unified Multimodal Large Language Models for Personalized Image Generation

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Instruction-tuned multimodal assistants · 2 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 MM-R1, which adds a cross-modal chain-of-thought (X-CoT) reasoning step and GRPO with reward signals to a unified MLLM (Lumina-mGPT) for zero-shot personalized image generation.

Problem

Personalized image generation with unified multimodal LLMs typically requires subject-specific fine-tuning or per-subject tokens, which limits scalability; the paper asks whether intrinsic reasoning can enable zero-shot personalization.

Approach

Structure personalization as a two-stage cross-modal reasoning process (subject grounding via text and a focus image, then conditioned generation), train it with an automatically built X-CoT dataset via SFT cold-start, and refine with GRPO using format, PickScore text-alignment and DreamSim subject-similarity rewards.

Key results

• On DreamBench, MM-R1 reaches DINO 0.786 and CLIP-T 0.313, the best among compared methods including Flux-Kontext (0.682 DINO) and GPT-4o (0.722 DINO).
• On Kontext-Bench, MM-R1 achieves DINO 0.562, highest among listed methods, with CLIP-I 0.750 and CLIP-T 0.296.
• Ablation: removing X-CoT cold-start (GRPO only) drops DreamBench DINO from 0.786 to 0.732; removing GRPO (X-CoT only) gives DINO 0.727.

Surprising findings

• Simply prompting the base Lumina-mGPT to reason before generating (no training) already improves over direct generation, e.g., DreamBench DINO 0.631 to 0.650.

Limitations

• Evaluation limited to two benchmarks and single-image reference setting; no human evaluation in main text.
• Reported gains on CLIP-T are mixed: MM-R1 scores 0.296 on Kontext-Bench, below Flux-Kontext (0.308) and GPT-4o (0.303).

Experimental setup

Datasets: Subjects200K X-CoT dataset (constructed) DreamBench Kontext-Bench

Benchmarks: DreamBench Kontext-Bench

Metrics: DINO CLIP-I CLIP-T

Baselines: DreamBooth Blip-Diffusion IP-Adapter Flux-Kontext Yo'Chameleon UniCTokens GPT-4o

Builds on: GRPO Chain-of-Thought Lumina-mGPT Subjects200K

Closest papers by meaning

• Yo'Chameleon: Personalized Vision and Language Generation (CVPR 2025 · similarity 0.91)
• Unified Personalized Understanding, Generating and Editing (CVPR 2026 · similarity 0.90)
• Contextualized Visual Personalization in Vision-Language Models (ICML 2026 · similarity 0.90)
• RePIC: Reinforced Post-Training for Personalizing Multi-Modal Language Models (NeurIPS 2025 · similarity 0.90)
• RAP: Retrieval-Augmented Personalization for Multimodal Large Language Models (CVPR 2025 · similarity 0.89)
• MM1.5: Methods, Analysis & Insights from Multimodal LLM Fine-tuning (ICLR 2025 · similarity 0.89)