Research map / Vision-language and multimodal / Multimodal LLMs

Think, Plan, Paint: Layout-Aware Reasoning for Controllable Image Generation in Unified Models

NeurIPS 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 ATLAS, a unified MLLM that performs Think-Plan-Paint image generation using layout (label-box tokens) as a shared representation, with RL-based layout alignment (Flow-GRPO), instantiated on BAGEL-7B and Hunyuan Image 3.0 (80B).

Problem

Unified MLLMs struggle to follow complex spatial instructions and logical constraints in controllable image generation, and existing layout-based unified models plan layouts directly from prompts without explicit reasoning.

Approach

Treat layout as a cross-modal bridge: a layout-aware chain-of-thought (Think) precedes layout generation (Plan) using shared positional coordinate tokens, which then conditions diffusion-based image synthesis (Paint), with the Paint stage further aligned by RL using layout-position, counting, negative-region, semantic and quality rewards.

Key results

• ATLAS-80B reaches 91.21 on ATLAS-Reasoning spatial and 99.17 on T2I-CompBench spatial, comparable to Nano Banana Pro on T2I-CompBench.
• ATLAS-7B obtains 72.31 on ATLAS-Reasoning versus 25.43 for PlanGen; the abstract reports an average 65.31% improvement over layout-based unified MLLMs.
• On GenEval, ATLAS-80B achieves 0.92 overall and ATLAS-7B 0.90, versus 0.88 for BAGEL.

Surprising findings

• The Think phase helps the 7B model much more than the 80B model in layout planning (average gain 17.6 vs 5.62 under Qwen3-VL evaluation).
• ATLAS-7B scores lower than BAGEL on RefCOCO (90.5 vs 91.7 for Qwen2.5-VL-7B baseline) and POPE, showing understanding trade-offs.

Limitations

• Bounding boxes are the only layout interface; other layout representations are not explored.
• Relies on LLM-based (Qwen3-VL-235B) annotation and judging, with only author-audited samples (500 layouts, 200 reasoning chains).

Experimental setup

Datasets: T2I-CompBench ImgEdit-Bench

Benchmarks: T2I-CompBench ATLAS-Reasoning DPG-Bench GenEval ImgEdit-Bench RefCOCO MMMU POPE

Metrics: T2I-CompBench category scores layout planning score rule-based clause accuracy HR@0.5 overall score

Baselines: PlanGen LayoutGPT GenArtist BAGEL Hunyuan Image 3.0 SDXL SD v3.5 FLUX.2-Dev GPT-Image-1 Nano Banana Pro Show-O2 Qwen2.5-VL-7B

Builds on: BAGEL Hunyuan Image 3.0 Flow-GRPO PlanGen

Closest papers by meaning

• Thinking with Programming Vision: Towards a Unified View for Thinking with Images (CVPR 2026 · similarity 0.89)
• LayoutVLM: Differentiable Optimization of 3D Layout via Vision-Language Models (CVPR 2025 · similarity 0.88)
• GIR-Bench: Versatile Benchmark for Generating Images with Reasoning (ICLR 2026 · similarity 0.88)
• Rethinking The Training And Evaluation of Rich-Context Layout-to-Image Generation (NeurIPS 2024 · similarity 0.88)
• 3D Space as a Scratchpad for Editable Text-to-Image Generation (CVPR 2026 · similarity 0.88)
• LayerCraft: Enhancing Text-to-Image Generation with CoT Reasoning and Layered Object Integration (NeurIPS 2025 · similarity 0.88)