PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Kavana Venkatesh, Connor Dunlop, Pinar Yanardag

arXiv:2504.05306v2cs.CV

TL;DR

Prompt-driven image workflows often leave users to supply originality and artistic depth; CREA proposes creative image editing through collaboration among specialized agents. The authors report that its edits are perceived as more creative and aesthetically pleasing than baseline methods, while evaluation of what is truly creative remains beyond the paper’s scope.

  • Problem

    Prompt-driven outputs can lack originality and artistic depth, requiring users to refine prompts, tune parameters, or manually edit images to add creativity.

    Source: §1 · pp. 1–2

  • Method

    CREA frames creative image editing as a task and uses specialized agents that collaborate to conceptualize, generate, critique, and refine image outputs.

    Source: §1 · p. 2, §3 · p. 3

  • Results

    CREA’s edits are reported as consistently perceived as more creative and aesthetically pleasing than baseline methods.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The authors conclude that agentic frameworks can support more autonomous and creative AI systems and open directions for human-AI creative collaboration.

    Source: §5 · p. 10

    Assessing what constitutes a truly creative output remains beyond the scope of the paper.

    Source: §5 · p. 10

Abstract

from arXiv · show

Creativity in AI imagery remains a fundamental challenge, requiring not only the generation of visually compelling content but also the capacity to add novel, expressive, and artistically rich transformations to images. Unlike conventional editing tasks that rely on direct prompt-based modifications, creative image editing requires an autonomous, iterative approach that balances originality, coherence, and artistic intent. To address this, we introduce CREA, a novel multi-agent collaborative framework that mimics the human creative process. Our framework leverages a team of specialized AI agents who dynamically collaborate to conceptualize, generate, critique, and enhance images. Through extensive qualitative and quantitative evaluations, we demonstrate that CREA significantly outperforms state-of-the-art methods in diversity, semantic alignment, and creative transformation. To the best of our knowledge, this is the first work to introduce the task of creative editing.

1 Introduction

Creative image work still relies on users to add originality and artistic depth to prompt-driven outputs. CREA proposes creative image editing and a collaborative agent framework, with optional user guidance and applications spanning editing, generation, personalization, and video.

1 Introduction

  • Prompt-driven outputs can lack originality and artistic depth, leaving users to refine prompts, tune parameters, or edit results to add creativity.This reliance on user expertise and intervention limits accessibility as an autonomous creative assistant.Source: §1 · pp. 1–2
  • Creative image editing aims to enhance images’ creative and artistic qualities with minimal user intervention, and CREA introduces it as a new task.The framework assigns agents distinct roles, such as Creative Director and Art Critic, to conceptualize, generate, and refine outputs.Source: §1 · p. 2
  • Optional user-in-the-loop interventions let people steer creative direction while supporting artistic coherence and control.The paper frames this as collaborative human-AI co-creation.Source: §1 · p. 2
  • The paper presents CREA as versatile across image editing and generation, with potential applications in personalization workflows and creative video generation.The authors describe the method’s potential in these workflows rather than reporting specific results for them here.Source: §1 · p. 2

2 Related Work

Prior work covers text-to-image synthesis, creative generation, image editing, and LLM-assisted artistic workflows, but the supplied related-work discussion identifies gaps in creativity-focused editing and collaborative agents.

2 Related Work

  • Text-to-image models produce high-fidelity images but lack a structured mechanism to enforce creativity principles during generation.Personalized generation approaches such as DreamBooth [44] and Textual Inversion [18] target style and subject consistency rather than creative images.Source: §2 · p. 2
  • Creative-generation methods use approaches such as repulsion, hierarchical decomposition, diffusion priors, and feature-map amplification, while the paper says they require costly retraining or optimization and do not generalize broadly.Examples include ProCreate [31], Inspiration Tree [50], ConceptLab [41], and C3 [21].Source: §2 · p. 2
  • Editing methods including ControlNet [59], SDEdit [36], Blended Latent Diffusion [3], and InstructPix2Pix [8] emphasize structural fidelity or user-directed transformations rather than creativity-driven minimal-guidance edits.The related-work discussion also says creative editing had not been explicitly tackled and agentic frameworks had not been systematically explored for creative image generation or editing.Source: §2 · pp. 2–3

3 Methodology

CREA frames creative editing and generation as a collaborative, iterative process that balances novelty with coherence. Specialized agents plan, generate, critique, and refine images, with optional user guidance.

3 Methodology

  • CREA replaces static prompt-to-image workflows with a collaborative loop of specialized agents that iteratively evolve images while balancing novelty and coherence.The framework is described as producing higher-quality, more interpretable edits with less human effort and lower computational overhead than manual prompt tweaking.Source: §3 · p. 3

3.1 Multi-Agent Framework Design

  • The Creative Director plans and coordinates, the Prompt Architect builds prompts, the Executor generates or edits, the Critic scores, and the Strategist proposes refinements.Agents maintain private memories and share task status; their modular roles support independent upgrades, interpretability, and extensibility.Source: §3.1 · pp. 3–4

3.2 Collaborative Agentic Synthesis

  • CREA optimizes a Creativity Index using six principles, applying the same framework to editing and, with minor adjustments, image generation.The principles are originality, expressiveness, aesthetic appeal, technical execution, unexpected associations, and interpretability and depth.Source: §3.2 · pp. 4–5

3.3 Pre-Generation Planning and Image Synthesis

  • Before synthesis, the Director forms a blueprint and the Architect fuses six principle-conditioned prompts; the Executor then generates an image or edits an existing one.The Executor checks prompt feasibility against model-specific constraints, including guidance and ControlNet conditioning scales.Source: §3.3 · pp. 5–6

3.4 Post-Generation Evaluation

  • The Critic scores each of six creativity criteria from 1 to 5, and the Director reviews the assessment against the blueprint when the Creativity Index falls below a predefined threshold.The image is finalized or sent for further enhancement after the Director and Critic reach consensus.Source: §3.4 · p. 6

3.4.1 Self-Enhancement with Optional User-Guidance

  • When scores fall below threshold, the Strategist targets weak creativity dimensions for prompt refinement; iterations stop at the threshold or the maximum K, with optional real-time user instructions.The Strategist is also tasked with maintaining artistic coherence as user guidance is integrated into the evolving prompt.Source: §3.4.1 · pp. 6–7
  • The method surpasses state-of-the-art methods across multiple metrics for both editing and generation tasks.DINO scores are unavailable for generation because that task has no reference image; starred scores are interpreted to favor greater perceptual distance.Source: Table 1 · p. 7

4 Experiments

CREA is evaluated for creative image editing and generation, with additional experiments on user guidance, personalization, video generation, and design choices. The evaluations report creative gains across tasks while identifying trade-offs and scope limits.

4.1 Qualitative Results

  • In qualitative editing comparisons, CREA makes disentangled creative changes, while baselines often alter colors or backgrounds without reimagining the object.The comparison uses LEDITS++, InstructPix2Pix, SDEdit, and TurboEdit with a “creative <object>” prompt.Source: §4.1 · p. 7
  • Across varied generation prompts, CREA produces diverse creative images, while ConceptLab can fail on abstract concepts or those without well-defined subcategories.The methods use the same random seed for each evaluated prompt.Source: §4.1 · p. 8

4.2 Quantitative Results

  • CREA consistently outperforms all editing baselines across the evaluated metrics.The metrics assess text alignment, perceptual diversity, structural and semantic retention, and diversity across generated images.Source: §4.2 · pp. 8–9
  • For generation, CREA leads in LPIPS and VENDI, while SDXL leads in CLIP; ConceptLab’s high VENDI comes with lower prompt alignment.The comparison also reports that SDXL has the lowest LPIPS and significantly lower VENDI than CREA and ConceptLab.Source: §4.2 · p. 9
  • In the user study, CREA leads on creativity for generation and editing, while matching editing competitors on consistency; SDXL leads on generation usability.The study included 50 participants rating criteria on a five-point Likert scale.Source: §4.2 · p. 9

4.3 Additional Experiments

  • Users can steer edits with extra conditions, and CREA-generated subjects or styles can be adapted to new contexts, though the adapter may miss fine details.The personalization results use an off-the-shelf adapter; main characteristics are preserved despite some fine-detail loss.Source: §4.3 · pp. 9–10
  • Ablations examine model generalization, parameter sensitivity, iterative refinement, prompt variants, and components; the full framework scores highest across Table 2’s metrics.The tested backbones include SDXL and DALL-E; CFG values for Flux range from 3.5 to 40, and ControlNet conditioning scales from 0.1 to 0.5.Source: Table 2 · p. 10, §4.3 · p. 10
  • For video, CREA turns an initial prompt into a structured plan that guides prompt composition and CogVideoX synthesis, producing more imaginative scenes than the baseline.The plan specifies subject, action, setting, style, and optional additional details.Source: §4.3 · p. 10

5 Discussion

The discussion notes generative-backbone biases and the open problem of assessing visual creativity. It presents CREA as a flexible framework for creative collaboration, while encouraging human-in-the-loop deployment.

5 Discussion

  • Generative-backbone biases can affect outputs, and defining and quantifying visual creativity remains an open problem beyond this paper’s scope.The paper notes unintended darkened backgrounds and hallucinated human figures, and uses established creativity principles rather than proposing a new assessment protocol.Source: §5 · p. 10
  • The authors identify both potential benefits for creative access and possible labor displacement, recommending human-in-the-loop use as co-creative assistance.The proposed framing is assistance rather than replacement of human creators.Source: §5 · p. 10
  • The conclusion presents specialized-agent collaboration as a basis for more autonomous creative systems and future human-AI creative collaboration.It describes the approach as disentangled, with greater flexibility and artistic control.Source: §5 · p. 10

A Investigating Creativity Types

The analysis examines shape, color, and texture as attributes associated with creativity in generated images. It reports stronger scores on these attributes and object-dependent emphasis among generated results.

A Investigating Creativity Types

  • CREA scores significantly higher in shape, color, and texture than the compared methods, with the analysis associating these attributes with more diverse, imaginative objects.The comparison includes CREA, ConceptLab, Flux, and SDXL.Source: A Investigating Creativity Types · p. 15, Figure 9 · p. 15
  • The factors CREA emphasizes vary by object, dynamically shifting among shape, color, and texture.The paper characterizes this as context-aware enhancement of creativity.Source: A Investigating Creativity Types · p. 15, Figure 10 · p. 16
  • The LLM-as-a-judge comparison reports that CREA surpasses state-of-the-art methods across aspects for both editing and generation.The evaluation uses a multimodal LLM to simulate human-like subjective assessments.Source: Table 3 · p. 16

B CREA’s Multi-Agentic Architecture

CREA coordinates specialized agents through role-specific prompts and iterative collaboration to plan, generate or edit images, critique results, and refine them toward a creativity threshold. For editing, the executor conditions on the input image to preserve its semantic core while transforming it creatively.

B.1 CREA’s Multi-Agent Framework Design

  • CREA’s five specialized agents coordinate through tailored prompts and coherent multi-turn conversations to support the creative process.Prompts serve as the agents’ coordination interface for contextual reasoning and iterative image improvement.Source: §B.1 · p. 16

B.2 Creative Director (A1) Prompt

  • The Creative Director interprets the user’s input, sets a creativity blueprint, coordinates agents, and decides whether an output needs refinement.The blueprint specifies themes, constraints, novelty objectives, semantic anchors, and possible unexpected associations.Source: §B.2 · p. 16

B.3 Prompt Architect (A2) Prompt

  • The Prompt Architect creates six contrastive prompts and fuses them into one prompt balancing creativity principles, semantic clarity, and stylistic coherence.It also adjusts prompts using feedback from the Refinement Strategist.Source: §B.3 · p. 17

B.4 Generative Executor (A3) Prompt

  • The Generative Executor applies the fused prompt using suitable generation or editing strategies, preserving input-image structure during edits.It can configure model parameters and use models such as Flux or ControlNet.Source: §B.4 · p. 17

B.5 Art Critic (A4) Prompt

  • The Art Critic scores six creativity principles from 1–5, explains its ratings, and identifies weak dimensions for possible refinement.Its multimodal evaluation informs both the Creative Director’s approval decision and the Refinement Strategist’s next steps.Source: §B.5 · p. 17

B.6 Refinement Strategist (A5) Prompt

  • The Refinement Strategist turns critique into targeted prompt adjustments while aligning changes with the blueprint to avoid creative drift or over-editing.It passes proposed adjustments to the Prompt Architect for the next iteration.Source: §B.6 · p. 17

B.7 Algorithm

  • For generation or editing, it derives a blueprint, creates six prompts, fuses them with CoT-Fusion, synthesizes an image, sums six 1–5 scores, and refines below-threshold results for up to K iterations.A full round with three self-enhancement iterations typically takes 3–5 minutes on an NVIDIA L40 GPU using FLUX.1-dev1 and ControlNet for editing; optional user-guidance can be enabled.Source: Pre-Generation Planning · pp. 18–19, §B.8 · p. 19

C.2 Creative Image Editing

  • Generation proceeds from blueprint and prompt planning to critique and refinement; editing follows the same cycle while conditioning on the input image to retain its semantic core.The Critic scores six principles and the Director approves or requests revisions; refinement continues until the creativity threshold is met or K iterations are reached.Source: §C.1.1 · pp. 19–20, §C.1.2 · p. 22, §C.1.3 · pp. 22–23

D.1 Additional Ablations

The supplementary material points to an SDXL negative-prompt ablation and provides an example of questions used in the user studies.

D.1 Additional Ablations

  • Figure 16 examines how adding the negative prompt “A normal <obj>” affects generation with SDXL.The supplied passage identifies the ablation but does not report its outcome.Source: §D.1 · p. 24, Figure 16 · p. 25

D.2 User Study Details

  • Figure 17 shows an example of questions asked in the user studies.The supplied passage identifies the study-question example without specifying the questions.Source: §D.2 · p. 24, Figure 17 · p. 25

D.3 Additional Quantitative Results

CREA outperforms comparison systems in additional editing metrics and in a human evaluation of creative expressiveness and edit disentanglement.

Creative Editing

  • CREA significantly outperforms RF-Inversion across quantitative metrics using the same Flux-1-dev base model.The comparison attributes the reported gains to structured prompt design and agentic reasoning rather than base-model differences.Source: §D.3 · p. 25, Table 4 · p. 26

Creative Generation

  • CREA scores significantly higher than GenArtist on creative expressiveness and edit disentanglement in a study with 25 participants.GenArtist was prompted with the word “creative” for the comparison.Source: §D.3 · p. 25, Table 5 · p. 26

D.4 Additional Qualitative Results

The additional qualitative results illustrate CREA’s creative image outputs and its range from localized edits to broad scene changes.

D.4.1 Editing and Generation Examples

  • Figures 20–31 provide qualitative examples of CREA’s image generation and editing results.The examples are presented to demonstrate the method’s ability to produce diverse and highly creative images.Source: §D.4.1 · p. 26

D.4.2 Fine-Grained Edits

  • Figure 18 illustrates fine-grained and broad semantic edits in realistic scenes while maintaining spatial coherence, lighting, texture consistency, and photorealism.The examples range from adding small objects to changing furniture color or wall tone.Source: §D.4.2 · p. 26

D.4.3 GPT-4o as Both Agent and Image Generator

With GPT-4o as both reasoning agent and image generator, CREA produces creative edits and generations, including transformations that range from subtle insertions to broad scene changes while preserving visual coherence.

D.4.3 GPT-4o as Both Agent and Image Generator

  • With GPT-4o in both roles, CREA interprets prompts, decomposes them into semantic sub-goals, and synthesizes visually coherent outcomes using diffusion-based rendering.The reported results span creative image editing and creative image generation.Source: §D.4.3 · p. 26
  • Edits enrich existing images with imaginative, stylistically rich augmentations while preserving structural fidelity and realism.The examples cover subtle insertions, localized attribute transformations, and large-scale scene modifications while retaining spatial and lighting consistency.Source: §D.4.3 · p. 26, Figure 19 · p. 27, Figure 18 · p. 27
  • With minimal prompts such as “a couch” or “a car,” CREA generates surreal, high-fidelity scenes balancing novelty with realism.The figure describes conceptual abstraction combined with photorealistic detail.Source: §D.4.3 · p. 26, Figure 19 · p. 27
  • Figures 20–25 show generation examples whose captions describe the outputs as consistently highly creative and diverse.The examples use couch, table, monster, mug, chair, and car concepts.Source: Figure 20 · p. 28, Figure 21 · p. 29, Figure 22 · p. 30
  • Figures 26–31 show editing examples whose captions describe the edits as consistently highly creative and diverse.The examples use bike, table, couch, bowl, helmet, and guitar concepts.Source: Figure 26 · p. 34, Figure 27 · p. 35, Figure 28 · p. 36

Related papers

Loading 2504.05306v2…