Research map / Vision-language and multimodal / Multimodal LLMs

VisionDirector: Vision-Language Guided Closed-Loop Refinement for Generative Image Synthesis

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 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.

Introduces LGBench, a 2,000-task benchmark of long multi-goal T2I and I2I prompts, and VisionDirector, a training-free VLM-planner closed loop for image generation/editing with GRPO fine-tuning of the planner.

Problem

Current text-to-image and image editing models fail to satisfy long, multi-goal instructions typical of professional design briefs, and existing benchmarks use short prompts with few goals.

Approach

A VLM planner decomposes long instructions into structured goals, chooses one-shot vs staged execution, runs micro-grid sampling with verification and rollback after each edit, and is further post-trained with GRPO to shorten edit trajectories.

Key results

• On LGBench, Qwen-Image reaches 71.8% finish and Flux-Kontext 55.9%; Qwen-Image + VisionDirector reaches 74.8% finish, Flux-Krea + VID gains +8.1 points finish.
• On GenEval, VisionDirector reports overall 0.94 versus 0.87 for Qwen-Image.
• GRPO-tuned planner reduces steps from 4.2 to 3.1 (about 26%) and edits per task from 3.3 to 2.5; ImgEdit overall 4.35 versus 4.27 for Qwen-Image-Edit.

Surprising findings

• Flux-Dev reaches only 0.8% success on text-related goals in LGBench, showing near-total failure on typography.
• Reported GenEval gains are large relative to base Qwen-Image (0.87 to 0.94) despite the framework being training-free for most of the pipeline.

Limitations

• Benchmark prompts are authored by Claude and verified by a Qwen VLM, so verifier and generator bias may affect results.
• Several baseline numbers are taken from original leaderboards and mixed evaluation settings; GRPO details (group size, reward model, training data size) are sparsely specified.
• Verification relies on VLM judges with confidence filtering, and results on closed-source models are not reported for LGBench.

Experimental setup

Datasets: LGBench

Benchmarks: LGBench GenEval ImgEdit

Metrics: goal finish rate success rate (>=80% goals) per-goal-type pass rate GenEval overall score ImgEdit overall score (1-5) steps goal coverage edits per task

Baselines: Qwen-Image Flux-Krea Flux-Dev Qwen-Image-Edit Flux-Kontext GPT Image 1 [High] Seedream 3.0 FLUX.1 Kontext [Pro] OmniGen2 BAGEL

Builds on: GRPO Qwen-Image Qwen3-VL DeepSeek-R1

Closest papers by meaning

• Leveraging Unpaired Data for Vision-Language Generative Models via Cycle Consistency (ICLR 2024 · similarity 0.87)
• Generative Photography: Scene-Consistent Camera Control for Realistic Text-to-Image Synthesis (CVPR 2025 · similarity 0.86)
• Elucidating the design space of language models for image generation (ICML 2025 · similarity 0.86)
• Imagine Before You Draw: Visual Prompt Engineering for Image Generation (NeurIPS 2026 · similarity 0.86)
• FireEdit: Fine-grained Instruction-based Image Editing via Region-aware Vision Language Model (CVPR 2025 · similarity 0.86)
• ImgEdit: A Unified Image Editing Dataset and Benchmark (NeurIPS 2025 · similarity 0.86)