Research map / Multi-agent systems and games / Human-AI and goal-conditioned
CREA: A Collaborative Multi-Agent Framework for Creative Image Editing and Generation
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 CREA, a GPT-4o-driven multi-agent framework (Creative Director, Prompt Architect, Generative Executor, Art Critic, Refinement Strategist) that iteratively plans, generates/edits, critiques with a six-principle Creativity Index, and refines images with FLUX.1-dev and ControlNet.
Problem
Creative image editing and generation require autonomous, iterative ideation and refinement; existing prompt-to-image and editing methods produce generic or superficially stylized outputs and demand heavy user effort.
Approach
Model creativity as a multi-agent collaborative loop grounded in six creativity principles, where an MLLM judge scores outputs with a Creativity Index and a refinement agent updates fused contrastive prompts until a threshold is met, with optional user guidance.
Key results
- On creative editing, CREA reaches CLIP 0.417 and LPIPS* 0.414 versus 0.396 and 0.252 for the best-performing baseline LEDITS++ and others.
- On creative generation, CREA achieves VENDI 10.44 and LPIPS* 0.709, with FID 248.67 vs 270.69 for Flux, while SDXL has higher CLIP (0.404 vs 0.360).
- User study (50 participants) finds CREA rated highest on creativity for generation (Q2 4.16 vs 3.56 for SDXL).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CREA instead of the strongest editing baseline, LEDITS++ | Creative image editing · Creative image editing | CLIP: 0.396 $\pm$ 0.028 -> 0.417 $\pm$ 0.030 | Helped |
| Use CREA instead of SDEdit | Creative image editing · Creative image editing | LPIPS: 0.308 $\pm$ 0.068 -> 0.414 $\pm$ 0.157 | Helped |
| Use CREA instead of the strongest generation baseline, SDXL | Creative image generation · Creative image generation | CLIP: 0.404 $\pm$ 0.033 -> 0.360 $\pm$ 0.052 | Hurt |
| Use CREA instead of ConceptLab | Creative image generation · Creative image generation | LPIPS: 0.663 $\pm$ 0.076 -> 0.709 $\pm$ 0.057 | Helped |
| Use CREA instead of ConceptLab | Creative image generation · Creative image generation | Vendi: 10.38 $\pm$ 2.27 -> 10.44 $\pm$ 2.15 | No clear effect |
| Add creativity principles to the base version (ablation) | CREA · Creative image editing | LPIPS-Diversity: 0.302 $\pm$ 0.129 -> 0.312 $\pm$ 0.142 | Helped |
| Add creativity principles to the base version (ablation) | CREA · Creative image editing | Vendi: 3.19 $\pm$ 1.49 -> 3.59 $\pm$ 2.46 | Helped |
| Add contrastive prompting after the principles variant (ablation) | CREA · Creative image editing | LPIPS-Diversity: 0.312 $\pm$ 0.142 -> 0.391 $\pm$ 0.150 | Helped |
| Add contrastive prompting after the principles variant (ablation) | CREA · Creative image editing | Vendi: 3.59 $\pm$ 2.46 -> 3.50 $\pm$ 1.64 | No clear effect |
| Add self-enhancement after contrastive prompting (ablation) | CREA · Creative image editing | LPIPS-Diversity: 0.391 $\pm$ 0.150 -> 0.414 $\pm$ 0.157 | Helped |
| Add self-enhancement after contrastive prompting (ablation) | CREA · Creative image editing | Vendi: 3.50 $\pm$ 1.64 -> 3.70 $\pm$ 1.97 | Helped |
Surprising findings
- SDXL achieves higher CLIP and user-rated usability than CREA for generation, and SDXL's LPIPS is the lowest, indicating its images are similar to each other.
- Under the same FLUX backbone, CREA outperforms RF-Inversion, which the authors interpret as the gains coming from the agentic prompting rather than the base model.
Limitations
- Inherits biases of the FLUX backbone, such as darkened backgrounds and hallucinated human figures.
- Creativity is operationalized via an MLLM judge and predefined principles; the paper explicitly does not provide a new validated assessment protocol.
- Each run relies on GPT-4o API calls and takes 3-5 minutes, adding latency and cost.
Experimental setup
Benchmarks: 24 objects x 25 prompts evaluation set (600 images per task)
Metrics: CLIP score LPIPS Vendi score DINO score FID KID user study ratings LLM-as-a-judge creativity criteria
Baselines: LEDITS++ InstructPix2Pix SDEdit TurboEdit SDXL Flux ConceptLab RF-Inversion
Builds on: AutoGen ControlNet ConceptLab Chain-of-Thought prompting
Closest papers by meaning
- JarvisArt: Liberating Human Artistic Creativity via an Intelligent Photo Retouching Agent (NeurIPS 2025 · similarity 0.88)
- CREval: An Automated Interpretable Evaluation for Creative Image Manipulation under Complex Instructions (CVPR 2026 · similarity 0.86)
- Unify-Agent: A Unified Multimodal Agent for World-Grounded Image Synthesis (NeurIPS 2026 · similarity 0.85)
- Enhancing Creative Generation on Stable Diffusion-based Models (CVPR 2025 · similarity 0.85)
- Muses: 3D-Controllable Image Generation via Multi-Modal Agent Collaboration (AAAI 2025 · similarity 0.84)
- VisionDirector: Vision-Language Guided Closed-Loop Refinement for Generative Image Synthesis (CVPR 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.