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

VisionDirector enables automatic step-by-step refinement for both Image Editing and Image Generation tasks, achieving superior multi-goal alignment compared to closed-source models.

Abstract

The field of visual content creation has progressed rapidly with the rise of diffusion models. However, professional image creation often relies on long, multi-goal instructions specifying global composition, local object placement, typography, and stylistic constraints. While modern diffusion models achieve strong visual fidelity, they frequently fail to satisfy such tightly coupled objectives.

To systematically expose this gap, we introduce LongGoalBench (LGBench), a dual-modality benchmark comprising 2,000 tasks (1,000 T2I + 1,000 I2I) with over 29,000+ annotated goals and automated goal-level verification. LGBench stresses multi-attribute alignment rather than simple prompt fidelity, with each task requiring satisfaction of 10–23 quantitative goals.

We further propose VisionDirector, a training-free, vision-language guided closed-loop framework that decomposes long instructions into structured goals, dynamically plans generation or editing actions, and verifies goal satisfaction after each step—achieving 30% improvement on GenEval and 60% improvement on ComplexBench over state-of-the-art methods.

Describe

Inspect

Revise

Edit

Converge

Toward

Optimal

Rendering

LGBench: LongGoal Benchmark

A director-style stress test for multi-goal image generation and editing

2,000

Total Tasks
(1K T2I + 1K I2I)

29,000+

Annotated Goals

18.0

Avg Goals per T2I Prompt

418

T2I Subcategories

Goal Type Distribution:

T2I: Additive objects (31.8%), Textual overlays (16.8%), Visual effects (16.6%), Color constraints (11.8%), Lighting (11.5%)
I2I: Effects (34.4%), Color grading (21.5%), Typography (17.5%), Lighting refinement (15.8%)

Comparison with Existing Benchmarks

Benchmark	Modalities	Prompt Complexity	Goals per Task	Scale
LGBench (Ours)	T2I + I2I	Long-chain	10–23	2,000
DrawBench / VBench	T2I	Single sentence	1	200–300
TIFA / GenEval	T2I	Short + QA	1–2	500–1000
MagicBrush / EditEval	I2I	Short directive	Few	<500

Framework Overview

VisionDirector introduces a director-style vision–language agent that decomposes long, multi-goal instructions into structured goals and supervises generation and editing through a closed-loop process with verification and rollback. The framework operates without retraining diffusion backbones, leveraging the strong perceptual capabilities of VLMs.

Adaptive Planning Behavior

VisionDirector adaptively switches between one-shot generation and staged refinement as instruction complexity increases, demonstrating rational planning behavior. The VLM dynamically decides when to iterate based on goal verification feedback.

Qualitative Results

VisionDirector consistently improves multi-goal adherence on complex generation and editing tasks

Image-to-Image Editing

Multi-stage editing with comparison against Qwen-Image, Seeddream, GPT-Image, and KONTEXT-MAX

Text-to-Image Generation

High-fidelity generation with multi-goal alignment

Complex Scene Generation

Handling extremely detailed instructions: Multidimensional Ballroom, Quantum Archaeology Laboratory, Forest Cathedral

Step-by-Step Refinement

Artistic Style Transfer

BibTeX

@misc{chu2025visiondirector,
  title        = {VisionDirector: Vision-Language Guided Closed-Loop Refinement for Generative Image Synthesis},
  author       = {Chu, Meng and Yang, Senqiao and Che, Haoxuan and Zhang, Suiyun and Zhang, Xichen and Yu, Shaozuo and Gui, Haokun and Rao, Zhefan and Tu, Dandan and Liu, Rui and Jia, Jiaya},
  year         = {2025},
  eprint       = {2512.19243},
  archivePrefix= {arXiv},
  primaryClass = {cs.CV}
}