Skip to content

CREA: A Collaborative Multi-Agent Framework for Creative Image Editing and Generation

🕒 Published (v1): 2025-01-01 · Source: NeurIPS · Venue: NeurIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

CREA is a five-agent collaborative framework for creative image editing and generation that replaces single-pass prompt-to-image pipelines with an iterative loop of specialized agents (Creative Director, Prompt Architect, Generative Executor, Art Critic, Refinement Strategist). It scores outputs against a six-principle Creativity Index and self-enhances until a threshold is met or iteration cap reached. CREA outperforms editing and generation baselines on CLIP, LPIPS, VENDI, and DINO at NeurIPS 2025.

Problem

Existing text-to-image and editing pipelines produce outputs that lack genuine creativity, requiring heavy user intervention to prompt-engineer novelty. No prior work defined or tackled "creative image editing" as a formal task — methods optimize for structural fidelity or style consistency rather than autonomous artistic transformation with minimal user input.

Method

CREA orchestrates five role-specific agents using AutoGen, with GPT-4o as the multimodal LLM backbone and FLUX.1-dev / ControlNet for generation/editing:

  1. Pre-Generation Planning: The Creative Director (A1) reads the input image or concept and writes a creativity blueprint \(B\). The Prompt Architect (A2) generates six contrastive sub-prompts, one per creativity principle, and fuses them via Chain-of-Thought into a single high-creativity prompt: \(P_c = \text{CoT-Fusion}(p_1, \ldots, p_6)\). The Generative Executor (A3) assesses technical feasibility and sets model parameters.

  2. Creative Generation/Editing: A3 synthesizes \(I_0 = G(P_c, \theta)\) (generation) or edits \(I_e = G(P_c, I_0, \theta)\) (ControlNet editing, Canny conditioning, scale 0.4).

  3. Post-Generation Evaluation: The Art Critic (A4) scores \(I_e\) on six creativity dimensions (Originality, Expressiveness, Aesthetic Appeal, Technical Execution, Unexpected Associations, Interpretability & Depth) each on a 1–5 scale. The Creativity Index is \(\text{CI} = \sum_{i=1}^{6} S_i\). The Creative Director may challenge scores that conflict with blueprint \(B\).

  4. Self-Enhancement: If \(\text{CI} < S_\epsilon\) (threshold 24 for editing, 26 for generation), the Refinement Strategist (A5) identifies weak dimensions and generates a prompt delta \(\Delta P\); A2 updates \(P_r = P_e + \Delta P\) and A3 regenerates. Iterates up to \(K=3\) times or until \(\text{CI} \geq S_\epsilon\). Users may inject guidance into the loop at any iteration.

Key Contributions

  • Formal introduction of the creative image editing task (first work to define it).
  • Five-agent collaborative pipeline with role-specialized private memories and shared task-state tools.
  • Six-principle Creativity Index grounded in established creativity theory (Boden, Guilford, Amabile, Ramachandran, Martindale, Berlyne, Geneplore).
  • Optional user-in-the-loop intervention during self-enhancement, enabling human-AI co-creation without full manual control.
  • Demonstrated generalization to creative video generation (via CogVideoX) and personalization workflows (DreamBooth-style).

Results

All editing results are against LEDITS++, InstructPix2Pix, SDEdit, TurboEdit; generation against SDXL, Flux, ConceptLab (600 images per task, 24 objects × 25 prompts).

Creative Image Editing: - CLIP: 0.417 vs. best baseline 0.396 (LEDITS++) - LPIPS (higher = more transformative): 0.414 vs. best baseline 0.308 (SDEdit) - VENDI (diversity): 3.70 vs. best baseline 3.19 (SDEdit) - DINO (structural retention): 0.744 vs. best baseline 0.737 (SDEdit) - FID: 294.19 (comparable range; TurboEdit best at 320.53 — note higher FID reflects greater creative departure) - User Study Q2 (aesthetic quality): 3.74 vs. best baseline 3.59 (InstructPix2Pix)

Creative Image Generation: - LPIPS: 0.709 vs. ConceptLab 0.663, Flux 0.650 - VENDI: 10.44 vs. ConceptLab 10.38 (near-tie), Flux 5.84 - FID: 248.67 (lowest among generation methods; ConceptLab 272.97, Flux 270.69) - User Study Q1 (creativity): 4.32 vs. SDXL 4.37 (near-tie), Q2: 4.16 vs. SDXL 3.56 - Pipeline runtime: approximately 3–5 minutes on a 48 GB NVIDIA L40.

Limitations

  • The paper is truncated before explicit limitations text; from what is visible in Figure 8(e) label, "Limitations" are explored but not reproduced in the provided text.
  • FID scores for editing are high (294+), indicating distributional shift from real images — inherent to creative divergence but problematic if photorealism is needed.
  • GPT-4o as the judge agent introduces cost and latency; the LLM-as-judge evaluation may not perfectly align with human creativity assessments.
  • Maximum \(K=3\) refinement iterations is a hard cap that may not suffice for highly abstract or compositionally complex concepts.
  • Relies on FLUX.1-dev + ControlNet (Canny); other conditioning signals or generative backends require separate ablation.
  • Creativity Index threshold is manually set (24/26); no adaptive mechanism for concept difficulty.

Relevance to Harnesses / Meta-Harnesses

CREA is a textbook domain-specific agentic harness: it wraps a generative backend (FLUX/ControlNet) in a structured multi-agent loop with role-specialized agents, shared state, iterative self-critique, and a quantitative stopping criterion — exactly the architecture pattern meta-harness designers study. The Creativity Index serves as an in-loop reward signal analogous to the objective functions used in code-generation or QA harnesses to trigger re-runs. The optional user-in-the-loop intervention demonstrates how meta-harnesses can expose human override points without breaking the automated pipeline, a design consideration directly relevant to building robust, controllable harnesses for any generative task.