Skip to content

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

๐Ÿ•’ Published (v1): 2025-01-01 ยท Source: NeurIPS ยท Venue: NeurIPS 2025 ยท link

Ask a follow-up

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

๐Ÿ’ฌ Ask ChatGPTโœฆ Ask Claude

TL;DR

CREA is a five-agent collaborative framework (NeurIPS 2025) for creative image editing and generation that structures the creative process as a multi-role agentic loop โ€” ideation, prompt synthesis, generation, critique, and refinement โ€” guided by six grounded creativity principles. It claims to be the first work to formalize "creative editing" as a distinct task. Across 600-image evaluations, it outperforms LEDITS++, InstructPix2Pix, SDEdit, TurboEdit, ConceptLab, Flux, and SDXL on diversity and semantic alignment.

Problem

Standard prompt-to-image pipelines produce outputs that replicate training-data patterns rather than generating novel artistic compositions; users must manually iterate prompts, requiring domain expertise. No prior work had formalized "creative editing" โ€” transforming an input image to enhance originality and artistic richness with minimal user guidance โ€” as an explicit task, and no multi-agent framework had been applied to creative image synthesis.

Method

CREA coordinates five specialized agents orchestrated via AutoGen: 1. Creative Director (A1) interprets the input image/concept and produces a creativity blueprint B. 2. Prompt Architect (A2) constructs six contrastive prompts โ€” one per creativity principle โ€” and fuses them via Chain-of-Thought into a single high-creativity prompt Pc. 3. Generative Executor (A3) selects and runs the appropriate T2I model (FLUX.1-dev + ControlNet for editing; Flux alone for generation) using Pc. 4. Art Critic (A4) scores the output 1โ€“5 per principle using a multimodal LLM-as-judge (GPT-4o), computing a Creativity Index CI = ฮฃ Si. 5. Refinement Strategist (A5) converts low-scoring dimensions into targeted prompt deltas ฮ”P when CI < threshold.

The loop iterates up to K=3 times until CI โ‰ฅ Sฮต (24 for editing, 26 for generation). The six creativity principles derive from established creativity theory: Originality (Boden, Guilford), Expressiveness (Amabile, Ramachandran), Aesthetic Appeal (Martindale, Berlyne), Technical Execution (Amabile), Unexpected Associations (Geneplore, Boden), and Interpretability & Depth (Ramachandran, Geneplore).

Key Contributions

  • Introduces "creative editing" as a new task distinct from instruction-following or style-transfer editing.
  • Five-agent collaborative pipeline grounded in six literature-backed creativity principles.
  • Iterative self-enhancement loop with optional user-in-the-loop guidance injected via the Prompt Architect.
  • Demonstrated extensibility to video generation (CogVideoX backbone) and personalization (off-the-shelf adapter).

Results

Editing (vs. LEDITS++, InstructPix2Pix, SDEdit, TurboEdit): - CLIP: 0.417 vs. best baseline 0.396 (LEDITS++) - LPIPS (diversity, higher = better for creative tasks): 0.414 vs. best baseline 0.308 (SDEdit) - VENDI: 3.70 vs. best baseline 3.19 (SDEdit) - DINO: 0.744 vs. best baseline 0.737 (TurboEdit) - User Study-Q2 (creativity): 3.74 vs. best baseline 3.59 (InstructPix2Pix)

Generation (vs. SDXL, Flux, ConceptLab): - LPIPS: 0.709 vs. best baseline 0.663 (ConceptLab) - VENDI: 10.44 vs. best baseline 10.38 (ConceptLab); SDXL only 6.63 - CLIP: 0.360 vs. SDXL 0.404 (CREA underperforms SDXL on CLIP) - User Study-Q2 (creativity): 4.16 vs. best baseline 3.56 (SDXL)

Ablation (LPIPS / VENDI): base 0.302/3.19 โ†’ +Principles 0.312/3.59 โ†’ +Contrastive 0.391/3.50 โ†’ +Self-Enhancement 0.414/3.70, confirming each component adds diversity.

Limitations

  • Inherits T2I backbone biases (e.g., unintentional background darkening, hallucinated human figures).
  • Defining and quantifying visual creativity is an open problem; the LLM-as-judge evaluation is itself a proxy.
  • Generation CLIP score underperforms SDXL, which tends to produce highly colorful generic images that align well with CLIP's proxy for creativity.
  • ConceptLab collapses on abstract concepts lacking BLIP subcategories; CREA is compared favorably, but no stronger generation baseline (e.g., GPT-4o image generation) is included.
  • Pipeline runtime is 3โ€“5 minutes per image on a 48 GB L40, significantly slower than single-pass baselines.
  • Video generation is proof-of-concept only; no quantitative evaluation is reported.

Relevance to Agentic AI / LLM Agents

CREA is a concrete instantiation of the role-specialized multi-agent pattern applied to a creative domain, demonstrating that decomposing a subjective, iterative task (creative synthesis) into discrete agent roles with private memory and explicit inter-agent communication yields measurable quality gains over single-pass pipelines. The use of LLM-as-judge inside the agentic loop โ€” where the Art Critic's scores gate further refinement โ€” exemplifies self-evaluation as a feedback signal for autonomous improvement, a recurring mechanism in agentic reasoning research. The Creative Director/Refinement Strategist interaction also illustrates how deliberative agents can challenge and override evaluative agents, adding robustness beyond simple critic-refiner dyads. For researchers tracking agentic AI, CREA shows that creativity-as-optimization (maximizing a structured score) is a tractable framing that standard agentic scaffolding (AutoGen + role prompting + memory) can execute effectively.