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

We thank the reviewer for their kind and positive comments. We address the reviewer's concerns in detail below:

Q1: The evaluation relies heavily on the vague prompt “a creative object,” which lacks clear guidance and makes fulfilling user intent difficult. The paper doesn’t clarify how agents add explicit constraints, and should include more diverse, detailed prompts specifying styles, materials, or design principles to better direct creativity.

We agree that the base prompt “a creative object” is intentionally underspecified. This was a deliberate design choice to evaluate CREA’s ability to autonomously expand vague or minimal input into rich, interpretable, and multi-dimensional outputs using its internal creative blueprint, B, and multi-agent reasoning loop. CREA’s central contribution is precisely its capacity to introduce explicit constraints and stylistic signals automatically through prompt refinement, grounded in both system priors and feedback from the Critic agent. As detailed in Section 3.3 and Supplementary material, the Prompt Architect incorporates dimension-specific refinement goals (e.g., “increase structural novelty” or “emphasize surreal forms”) based on which creativity principles are underperforming. This mechanism enables CREA to infuse prompts with targeted constraints across material, structure, function, or aesthetic, often leading to generations that include hybrid materials, symbolic compositions, or biomorphic deformations. Moreover, CREA supports user-in-the-loop control (Section 3.4), where users can specify preferred styles, constraints, or materials that are propagated throughout the agentic pipeline. While “a creative object” serves as a common minimal anchor in evaluation, the framework itself is not limited to vague prompts and in fact excels at refining or grounding them. We will clarify this design choice and incorporate additional examples involving explicit stylistic and material constraints in the revision.

Q2: The paper could be strengthened if the authors provided a detailed example of how the agent interacts across a full round of the editing pipeline, including the contrastive prompts and the generated images. The broader impact and limitations section could be moved to the appendix to make space.

We agree and appreciate the suggestion. We kindly point out that a sample interaction between our agents are shown in Figures 5–8 (in Appendix), each showing a complete round of the editing pipeline, from the initial base concept, through the contrastive creativity prompts generated by the Prompt Architect, to the feedback from the Critic, the refinement goals proposed by the Strategist, and the final revised generation. These traces demonstrate how each agent contributes, how contrastive reasoning is applied, and how feedback drives iterative improvements. We also point out that the interaction logic is formalized in Algorithm 1 (Appendix).

Additionally, we note that we provided the source code in the supplementary material, and code/crea/agents folder has the full instructions used by our agents for transparency. We will explicitly reference detailed versions of these examples in the main paper.

Q3: There are missing details that make the creative process hard to follow. For example, what exactly is a contrastive prompt? It is unclear how the six principles are actually used by the agent to refine outputs. How does the Prompt Architect synthesize contrastive prompts based on distinct creativity principles?

We appreciate this important observation and clarify that contrastive prompts refer to structured prompt pairs; one representing a base or failed concept, and the other an improved variant targeting a specific creativity dimension (e.g., “a mug on a table” vs. “a biomorphic mug with swirling iridescent textures”). Six contrastive prompts are generated, one for each of the six creativity principles-such as Originality, Surprise element, as given in the predefined editable creativity template to highlight all of them accurately. These contrastive pairs enable the Prompt Architect to reason over differences and generate refined prompts that inject novelty while preserving core semantics. The six creativity principles serve as diagnostic signals: the Critic agent scores each image along these axes, and the Refinement Strategist identifies which dimensions scored below threshold. These are translated into natural language refinement goals (e.g., “increase originality through unexpected associations”), which the Prompt Architect incorporates using chain-of-thought-style fusion. This process is fully demonstrated in the multi-turn trace of Figures 5-8 of the supplementary section. We will clarify the contrastive prompting and principled refinement steps more explicitly in the revision.

Q4: The evaluation metric is a bit questionable. The method favors both a high LPIPS score (which indicates more dissimilar images) and a high DINO score (which indicates more similar images). Wouldn’t these two metrics contradict each other? Also, why are the CLIP and DINO metrics missing in Table 2?

To clarify, LPIPS and DINO metrics serve distinct, complementary purposes in our evaluation:

• LPIPS is used to quantify perceptual diversity by computing pairwise distances between generated images within the same object group. In the context of creative generation, higher LPIPS indicates greater visual diversity across outputs, a desirable property for tasks emphasizing creativity and expressive variation.
• DINO, in contrast, assesses how well key structural and semantic characteristics of the original image are retained after editing. A higher DINO score thus implies better preservation of essential semantic details, even amid substantial visual transformations.

Thus, rather than being contradictory, high scores in both LPIPS and DINO metrics together indicate successful disentangled editing, producing edits that are both meaningfully creative and semantically coherent.

Q5: The method has limited applicability. The creative editing results do not allow for explicit structural changes, and the model is focused on generating a single object rather than handling more complex scenes.

Thank you for raising this concern. CREA is designed to be a tool-agnostic and flexible framework that supports explicit structural edits and is not limited to single-object generation.

CREA enables structural changes through its Generative Executor, which picks the appropriate tool from an extensible set of tools of the user’s choice to perform the task at hand (editing, generation, style transformation etc.), while also providing the tool and task-appropriate hyperparameters to control editing. This allows CREA to perform various types of edits at different granularities, depending on the chosen editing tool.

Moreover, CREA is not inherently limited to single-object outputs. While the paper focuses on single-object examples to clearly showcase disentanglement and ensure fair, interpretable evaluation across the six creativity dimensions, the framework itself is content-agnostic. The creativity blueprint generated by the Director agent supports compositional planning, and the editing pipeline can seamlessly extend to multi-object or scene-level editing. This is further demonstrated in the video extension, where agents generate coherent multi-entity scenes based on structured prompts covering subject, action, setting, and style. Due to NeurIPS guidelines, we are not allowed to provide example images here, we will ensure to include multi-object scenes in the camera-ready version.

Dear reviewer,

We appreciate your valuable review, and we addressed all your concerns in our rebuttal. As the NeurIPS discussion phase will conclude soon, we kindly remind you that if you have any further questions or require additional clarifications on our paper, we are more than happy to address them promptly.

Best,

Authors of CREA

We thank the reviewer for their detailed and helpful feedback. Please find below our responses to each of your questions.

Q1: The core technical contribution overlaps with GenArtist, which uses a single agent architecture.

We appreciate the reviewer highlighting this point regarding the relationship between our method and GenArtist. While GenArtist indeed utilizes a single-agent architecture for general image generation and editing, we clarify that our work fundamentally differs both in design and scope.

• CREA explicitly introduces and addresses, for the first time, the task of creative image editing, which emphasizes iterative, autonomous transformations guided by structured creativity principles. In contrast, GenArtist mainly targets general-purpose multimodal editing tasks, without specifically addressing or defining creativity-oriented editing or systematically measuring outputs against formal creativity dimensions.
• While GenArtist uses a single multimodal LLM agent that directly performs editing tasks via a general-purpose prompt-based method, CREA proposes a novel multi-agent collaborative pipeline with specialized agents (Creative Director, Prompt Architect, Generative Executor, Art Critic, Refinement Strategist). This structured, role-based division of labor ensures a more systematic, interpretable, and iterative refinement process, closely mimicking human creative workflows.
• CREA incorporates explicit iterative refinement loops guided by detailed, agent-generated feedback grounded in well-established creativity theories (e.g., originality, aesthetics, unexpected associations). GenArtist, by comparison, operates without explicit structured evaluation against a formal criteria (e.g. creativity).

Due to NeurIPS rebuttal guidelines, we are not allowed to include additional visuals here. However, to further substantiate our performance differences, we conducted a human evaluation on Prolific involving 25 users, explicitly comparing our CREA framework against GenArtist (where we provide 'creativity' keyword in the editing prompt as 'a creative bike'). Our results indicate that CREA significantly outperforms GenArtist across user-assessed creativity metrics and overall visual quality.

Furthermore, to provide additional evidence that CREA’s multi-agent mechanism is not equivalent to a single-step, monolithic LLM approach, we conducted detailed ablation studies (see Table 2 in the main paper). Specifically, when moving from a strong single-step LLM baseline ("Base"), where we prompt an LLM to directly generate a creative object description, to our complete CREA pipeline, we observe substantial improvements in LPIPS-Diversity and VENDI metrics for generation, as well as simultaneous gains in LPIPS (reflecting creative diversity) and DINO (indicating structural and semantic preservation) for editing. These ablations conclusively demonstrate the unique value of CREA’s structured multi-agent design.

We will explicitly include these detailed comparative evaluations, visual examples, and extended discussions about the key differentiating factors between CREA and GenArtist in our revised manuscript.

Q2: The CREA framework’s multi-stage, multi-agent pipeline may be unnecessarily complex, as a powerful LLM like GPT-4o could likely generate comparably creative prompts with a simpler, single-step instruction. While the ablation study shows improved performance, it does not convincingly justify the added complexity, latency, and overhead.

We appreciate the reviewer’s thoughtful concern regarding the complexity of our framework. While a single-step LLM prompt, even when enriched, can indeed generate reasonably creative outputs, we respectfully argue that such a monolithic approach fundamentally lacks the capacity to model creativity as an iterative, and structured optimization process. Creativity is inherently dynamic, involving multiple, often conflicting principles - such as maximizing originality without sacrificing coherence or interpretability, which cannot be effectively captured or optimized in a single-step instruction.

CREA explicitly addresses this complexity by adopting a closed-loop, multi-agent architecture: the Art Critic scores generated images along six well-defined creativity principles; the Refinement Strategist identifies dimensions needing improvement; the Prompt Architect adaptively refines prompts based on specific feedback; and the entire iterative process is orchestrated by a blueprint-aware Creative Director. This design enables CREA to systematically disentangle and iteratively enhance distinct creative attributes through explicit feedback, contextual memory, and adaptive self-correction, capabilities that single-step prompting inherently lacks. As noted by the reviewer, our ablation studies (see Table 2 in main) provide strong empirical support for this argument. Even a robust single-step GPT-4o baseline ("Base") significantly performs worse.

Regarding the latency and overhead: CREA typically achieves convergence within 2–3 iterations, resulting in a total latency of approximately 3–5 minutes. Considering that typical user-led prompting workflows involve extensive trial-and-error, often taking substantial manual effort and time, we believe CREA represents a reasonable alternative.

Thus, CREA’s architectural complexity is not arbitrary; rather, it is essential for transforming creative image generation and editing from a single-step expansion into a structured optimization problem over interpretable, controllable, and systematically improvable creative objectives. Nevertheless, in our final manuscript, we will explicitly discuss and clarify this complexity-performance trade-off.

Q3: While the paper claims superiority over existing methods, its comparisons are potentially unfair; baselines use simple prompts, while CREA benefits from complex, agent-generated ones. Without testing baselines on equally enriched prompts, it’s unclear if the gains come from the agentic design or just better prompting.

We appreciate the reviewer highlighting this important aspect. Our intent was to fairly evaluate each method according to its inherent design: baseline models naturally depend on direct user-provided prompts and lack built-in mechanisms for structured creativity reasoning or sophisticated prompt elaboration. In contrast, CREA’s key novelty is precisely its agentic framework designed to autonomously generate enriched and multi-dimensional creativity cues.

We also emphasize that our agentic framework is designed to be flexible, enabling seamless integration with existing editing methods such as Ledits++. We have integrated Ledits++ into the CREA pipeline, and our preliminary results demonstrate that CREA enhances the creative generation capabilities of Ledits++ by providing richer, context-aware prompts grounded in explicit creativity principles. Due to NeurIPS rebuttal guidelines, we are unable to include visual examples here, but we will explicitly incorporate these additional results in our final manuscript.

Q4: Limited and Unconvincing Visualization of Prompts and Outputs/The paper showcases visual results from simple prompts but does not reveal the final, transformed prompts used by the system.

We appreciate the reviewer’s insightful comment regarding prompt transparency and interpretability. Due to space constraints in the main manuscript, we were unable to include all the final effective prompts directly in the images. However, in Appendix Fig. 5, we explicitly illustrate an example demonstrating how a simple input keyword ("couch") evolves through multiple stages of prompt elaboration and evaluation by our agents, resulting in a detailed, enriched prompt that drives the creative output.

Additionally, we kindly point out that we provided the source code in the supplementary material, and code/crea/agents folder has the full instructions used by our agents for transparency, and we will include the transformed prompts in the appendix in the final manuscript.

Moreover, we also emphasize that our supplementary materials “website” and our supplementary pdf (NeurIPS_2025__CREA.pdf) comprehensively present all 600 visual results generated in our evaluation, without any cherry-picking or selective curation. We hope that the extensive number of examples supports the robustness of our qualitative claims.

Q5: Is the most important thing to generate a "fancy" image a better-oriented prompt? Does the whole system aim to generate a suitable prompt for generation model to generate "creative" images?

We appreciate the reviewer raising this point. While generating a high-quality prompt certainly impacts the creativity of image outputs, our primary goal goes beyond merely crafting better-oriented prompts. Instead, the core novelty and strength of our framework lie in the structured, collaborative reasoning process among multiple specialized agents, each grounded in established creativity principles. This multi-agent interaction ensures not only a "fancy" or enriched prompt but also systematically addresses aspects such as originality, expressiveness, aesthetics, technical execution, unexpected associations, and interpretability. Thus, the framework's advantage is not just better prompting, but rather an explicit, interpretable, and iterative reasoning pipeline designed specifically to foster and enhance genuine creative outputs.

Moreover, we emphasize that our agentic framework can potentially generalize to other abstract and nuanced concepts that are challenging for users to articulate explicitly in text, such as the abstract concepts depicted in “GANalyze: Toward Visual Definitions of Cognitive Image Properties”, Goetschalckx et. al.

Dear reviewer,

We appreciate your valuable review, and we addressed all your concerns in our rebuttal. As the NeurIPS discussion phase will conclude soon, we kindly remind you that if you have any further questions or require additional clarifications on our paper, we are more than happy to address them promptly.

Best,

Authors of CREA

Dear reviewer,

We would like to express our sincere gratitude for your constructive feedback during the reviewing phase. If you find that all concerns have been satisfactorily addressed, we respectfully request that you update your final justification and score to conclude the discussion phase.

In our rebuttal, we provided comparisons with GenArtist as requested, as well as clarifying all the raised concerns. If you still have any remaining concerns, we will make our best effort to address them before the rebuttal period closes today.

We would also like to highlight that our appendix contains over 600 visual examples illustrating the diversity and editing capabilities of our method, accompanied by the full source code detailing the exact prompts and commands executed by our agents.

Best,

Authors of CREA

We appreciate the reviewer's insightful comments. Below, we answer your concerns:

Q1: Limited editing capabilities. While the qualitative demonstrations look nice, the editing prompts used in evaluation fail to address more general editing scenarios such as adding, moving, removing objects and etc -- operations beyond just changing appearance of the original object.

We appreciate the reviewer highlighting this important limitation. Indeed, our current demonstrations and evaluation prompts primarily focus on appearance transformations and creative stylistic variations, and do not explicitly showcase more general editing operations such as adding, or removing objects within scenes.

However, we clarify that our core agentic framework is not fundamentally restricted to appearance editing tasks. The underlying architecture is highly modular, flexible, and capable of integrating additional editing methods that explicitly support structural modifications - such as adding, or removing objects through appropriate conditioning mechanisms and spatial guidance (e.g., segmentation maps, bounding boxes, and layout controls provided by ControlNet).

To showcase this capability, we conducted additional experiments involving object addition and removal within a realistic living room scene. Specifically, we tasked our method with adding a "bear" on top of a couch and removing an existing "vase." We conducted a user study with 39 participants to evaluate CREA’s ability to perform disentangled edits in complex, multi-object scenes across various edit types, including style, color changes, object addition, and removal. The results showed overwhelming agreement that CREA preserves the overall structure of the image while making targeted edits: 97.4% rated the color change as highly disentangled, and 100% rated the object addition, removal, and creative transformations as successful without affecting other elements in the scene. These findings confirm that CREA enables fine-grained, high-fidelity edits in multi-object settings while maintaining disentanglement and visual coherence. Due to NeurIPS rebuttal guidelines, we are not allowed to upload visuals here. We will explicitly include these additional experiments, quantitative results, and corresponding visual examples in the final manuscript, further elaborating on the modularity and extensibility of CREA in broader editing scenarios.

Q2: The exact "editing" setup is not well-presented. Is the editing based on language instructions or changing phrases / expressions in the original prompts used for generation? I couldn't tell from the figures or tables as the editing prompts are not shown and the baselines are mix of instruction-based (instructpix2pix) and prompt-based editing models (turboedit, sdedit).

We acknowledge that mixing instruction-based and prompt-based baseline methods might have introduced confusion regarding the exact editing methodology. The reason for including both instruction-based and prompt-based baseline methods was to comprehensively evaluate against state-of-the-art techniques spanning diverse editing paradigms.

To clarify our exact editing setup: given an input image (e.g., a "cup") and optional user-in-the-loop guidance (e.g., "a monster cup"), our agentic framework autonomously synthesizes detailed creative editing instructions grounded in structured creativity principles. These instructions are then applied through a disentangled editing process with Flux and ControlNet, ensuring targeted and interpretable creative transformations.

Additionally, we kindly point out that we provided the source code in the supplementary material for transparency, and we will explicitly include these generated editing instructions in the Appendix by the camera-ready version to facilitate clearer comparisons.

Q3: It is hard to tell if the improvement comes from the base model difference. For example, TurboEdit adopts SDXL-Turbo while CREA adopts Flux1.-dev, a better image generator that the former -- not to mention InstructPix2Px and SDEdit.

We thank the reviewer for raising this important concern regarding the base model differences. Indeed, baseline methods such as TurboEdit, InstructPix2Pix, and SDEdit utilize various underlying diffusion models, including SDXL-Turbo and Stable Diffusion. We selected these baselines to represent state-of-the-art methods across diverse editing paradigms.

To directly investigate whether our performance improvements stem primarily from our framework or the choice of the base model (Flux-1-dev), we conducted additional ablation studies (Sec. 4.3, Fig. 6a) where we implemented our CREA framework using alternative base models such as SDXL and DALL-E. These experiments demonstrate that our agentic framework is able to generate creative images across different base models.

Additionally, we conduct an experiment using RF-Inversion (Rout et al., 2024) a Flux based image editing method to further demonstrate that our performance is not solely due to the base model.

We will include these experiments to our final manuscript.

Q4: Since CREA is an agent framework, it may take several iterations of the whole pipeline to get a sample that passes the critic's judgement threshold. What's a more fair baseline here is to select the top-1 out of k generation results using the base model and compare to CREA. While this is not necessary in the rebuttal, it seems reasonable to show the overall evaluation results from each iteration of CREA to show that the generation results keep improving thanks to the critic and refinement. Otherwise, it could be that the whole framework is doing rejection sampling and eventually stop at a good enough sample.

We thank the reviewer for raising this valuable point. We acknowledge that due to our iterative agent-based pipeline, multiple refinement rounds indeed occur before obtaining a sample that meets the critic's creativity threshold. While superficially similar to rejection sampling, we emphasize that our iterative refinement mechanism actively utilizes the critic's detailed feedback to inform targeted prompt adjustments, rather than simply discarding unsuccessful results.

We fully agree that explicitly demonstrating improvement across iterations would clearly illustrate the value of our critic-based refinement approach. In fact, we have already conducted comprehensive iterative analyses, as presented in Fig. 6(c) in the main manuscript. This figure clearly shows progressive improvement across successive iterations, confirming that our agentic refinement mechanism systematically enhances output quality rather than merely selecting from multiple random samples.

Nevertheless, we appreciate the reviewer's suggestion for additional clarity. In our revised manuscript, we will include detailed discussion on this.

Dear reviewer,

We appreciate your valuable review, and we addressed all your concerns in our rebuttal. As the NeurIPS discussion phase will conclude soon, we kindly remind you that if you have any further questions or require additional clarifications on our paper, we are more than happy to address them promptly.

Best,

Authors of CREA

We thank the reviewer for their constructive feedback during the review process. While we have not yet received a response during the discussion phase, we hope that our detailed rebuttal (addressing all stated concerns and additional experiments) has satisfactorily resolved the issues raised.

Should any concerns remain, we would be happy to provide further clarification before the discussion period concludes today.

We would also like to highlight that our appendix contains over 600 visual examples illustrating the diversity and editing capabilities of our method, accompanied by the full source code detailing the exact prompts and commands executed by our agents.

Q1: The internal logic of some agents (e.g., the critique agent) is described at a high level. More details on how each agent’s output is used, especially how conflicting feedback is resolved in the refinement step, would improve clarity.

We thank the reviewer and agree that clarifying agent logic enhances the paper. CREA’s multi-agent loop is implemented via AutoGen’s ConversableAgent API, with each agent assigned bounded memory and tools (see Supplementary Section, Page 35). The Critic Agent (A4) scores images across six creativity principles, producing a vector S = {S₁,…,S₆}, and operates with window_size = 1 to ensure unbiased, sample-specific evaluation. The Creative Director interprets these scores in context of the original blueprint and resolves conflicts when principle-level feedback misaligns with creative intent. The Refinement Strategist then identifies low-performing dimensions and generates instructions, which the Prompt Architect integrates into the next-generation prompt. This process repeats until the creativity index exceeds threshold or max steps are reached.

Full interaction traces are shown in Figures 5–8 (in Appendix) and the logic is formalized in Algorithm 1 (Appendix). Additionally, we kindly point out that we provided the source code in the supplementary material, and code/crea/agents folder has the full instructions used by our agents for transparency. We will clarify this loop more explicitly in the final manuscript to improve clarity.

Q2: Can the authors elaborate on how the critique agent produces structured feedback and how that feedback is utilized during refinement? Are the critiques learned or rule-based?

The Critique agent produces structured feedback by prompting a multimodal LLM-as-a-Judge using a predefined creativity template distilled from creativity principles from well-established literature on creative computing, scoring each generated image across six creativity principles. The critiques are not learned or rule-based but generated via zero-shot evaluations from the LLM using standardized criteria. These scores are then used by the Refinement Strategist to identify underperforming dimensions and generate targeted prompt refinements, as detailed in Appendix Figures 5-8.

We also emphasize that our source code has the full instructions of our agents, including critique under Supplementary, code/crea/agents/critic.py.

Q3: CREA involves multiple inference rounds and agent coordination, which may limit real-time usability. A discussion of system latency and computational cost is missing./Please provide measurements (e.g., average number of iterations, inference time per step) to better understand CREA's feasibility for interactive use.

We appreciate the reviewer raising this practical consideration. As detailed in Sec 4, Experimental Setup section, currently, a complete pipeline execution, including iterative refinement, takes approximately 3–5 minutes on a 48GB NVIDIA L40 GPU. We also conducted additional timing analyses and provide the following concrete measurements:

Avg iterations per run: 1.50 iterations

Avg total time per run: 176.54 seconds (2.94 minutes)

Avg time per iteration: 117.70 seconds (1.96 minutes per iteration)

However, we emphasize that this latency remains substantially lower than the typical time investment required by manual user-driven approaches, which usually involve extensive trial-and-error prompting to achieve comparable creative outputs. Thus, our framework aims to reduce the human effort and overall creative turnaround time. We will expand the revised manuscript to explicitly discuss computational cost and system latency.

Q4: While the authors use CLIP-based and perceptual metrics, creativity is inherently subjective. The human evaluation setup is relatively small (50 samples); more robust or crowd-sourced studies would strengthen claims.

We acknowledge that our primary experiments included a relatively modest user study with 50 participants (Sec. 4.2 in the main paper). To further address the subjectivity and robustness concerns, we also provided additional comprehensive evaluations using a multimodal LLM-as-a-Judge framework (Appendix Sec. A.2, Table 1), simulating human-like subjective assessments across several creativity dimensions including Originality, Expressiveness, Aesthetic Appeal, Technical Execution, Unexpected Associations, and Interpretability. Moreover, our appendix (NeurIPS_2025__CREA.pdf) provides an extensive set of qualitative results across all evaluated concepts, covering the entire set of 600 generated and edited images without cherry-picking (Appendix Figures 9–32). These comprehensive visuals reinforce our claims, alleviating concerns of selective curation.

Q5: CREA is evaluated primarily on artistic or fantasy domains. It is unclear how well it generalizes to domains like product design or real-world edits where creativity may be more constrained.

We appreciate the reviewer’s point. CREA is designed as a general-purpose creative reasoning framework, and while our examples often show artistic and fantastical domains where creativity is most visually measurable, CREA supports constraint-driven generation through its predefined creativity template, blueprint generation mechanism and user-in-the-loop controls, which can be modified to suit the user’s needs to seamlessly leverage CREA’s capabilities for any domain. We will clarify this extensibility in the revision.

Q6: It would be helpful to include examples of failure modes; when CREA produces incoherent or uncreative edits and how the system might be improved to mitigate these.

We appreciate the reviewer emphasizing the importance of transparently discussing potential failure modes. While we are unable to directly include visual examples in this rebuttal due to NeurIPS guidelines prohibiting supplementary images or URLs, we kindly point out to the examples of certain limitations (Fig. 6(e) in the main paper). These cases illustrate unintended semantic shifts or background hallucinations, such as undesired human figures spontaneously appearing or backgrounds unintentionally darkening based on the semantics of edited objects. In the camera-ready version, we plan to further expand this discussion by explicitly providing additional visual examples of failure modes.

Q7: How does CREA compare to a monolithic iterative approach (e.g., looping Plug-and-Play generations with feedback) in terms of both performance and interpretability?

We appreciate the reviewer’s thoughtful concern regarding the complexity of our framework. While a single-step LLM prompt, even when enriched, can indeed generate reasonably creative outputs, we respectfully argue that such a monolithic approach fundamentally lacks the capacity to model creativity as an iterative, and structured optimization process. Indeed, our ablation studies (see Table 2 in main) provide strong empirical support for this argument. Even a robust single-step GPT-4o baseline ("Base") significantly performs worse.

Q8: Given the potential use of such systems in generating deepfakes or manipulated visuals, have the authors considered misuse risks? Are any mitigation strategies discussed?

We appreciate the reviewer raising this important concern. Like any powerful generative system, CREA carries potential misuse risks, such as generating manipulated or misleading visuals. As noted in Section 5, we advocate for CREA to be used strictly as a human-in-the-loop co-creative assistant, not as an autonomous generator. In the revised manuscript, we will outline concrete mitigation strategies, including safety filters, content moderation, and controlled access protocols for sensitive deployments.

Q9: The authors mention the subjective nature of creativity and the challenges of evaluation, which is appropriate. However, further discussion on scalability, domain transfer, and potential ethical misuse (e.g., for disinformation or deceptive content) is recommended.

We appreciate the reviewer acknowledging our discussion about creativity's inherent subjectivity and evaluation challenges. In our revised manuscript, we will explicitly expand our discussion to cover:

• Scalability: Clearly outlining computational costs, latency trade-offs, and strategies (e.g., agent parallelization, model distillation) that could enhance system scalability.
• Domain Transfer: Discussing CREA’s potential applicability to diverse domains beyond those demonstrated, emphasizing how the modular multi-agent design allows for adaptability and extension to abstract tasks that are challenging to articulate explicitly.
• Ethical Misuse: Providing detailed consideration of risks associated with misuse, such as deepfakes or deceptive visual content, along with explicit strategies for mitigating these risks, including content moderation, watermarking, and responsible deployment guidelines.

Q10: Include a section on computational cost and a mitigation discussion on possible societal misuse.

In practice, CREA converges within 2–3 agentic refinement rounds, averaging 3–5 minutes per image, comparable to real-world manual prompting workflows. Each agent role uses compact GPT-4o prompts, resulting in 6–9 LLM calls per sample. At current OpenAI pricing, this equates to a per-image cost of roughly 0.02–0.05 USD, comparable to iterative user prompt refinement. Future work will explore distilling agent behavior into smaller or open-weight models to further reduce cost without compromising reasoning quality. We will summarize these computational and ethical considerations in the revised Broader Impact section.

Dear reviewer,

We appreciate your valuable review, and we addressed all your concerns in our rebuttal. As the NeurIPS discussion phase will conclude soon, we kindly remind you that if you have any further questions or require additional clarifications on our paper, we are more than happy to address them promptly.

Best,

Authors of CREA

Dear reviewer,

We thank you for your constructive feedback during the review process. While we have not yet received a response during the discussion phase, we hope that our detailed rebuttal (addressing all stated concerns) has satisfactorily resolved the issues raised.

Should any concerns remain, we would be happy to provide further clarification before the discussion period concludes today.

We would also like to highlight that our appendix contains over 600 visual examples illustrating the diversity and editing capabilities of our method, accompanied by the full source code detailing the exact prompts and commands executed by our agents.

Best,

Authors of CREA

We thank the reviewer for their thoughtful and constructive feedback. We address each of the reviewer's comments point-by-point below.

Q1: Fig. 1 is not showed in a very intuitive way; the video can be displayed in frames; the generated prompts are also not shown. “Baseline” is not specified either.

Thank you for the constructive feedback. In fact, in line with your suggestions, our Fig. 5(c) presents five representative frames per video and specifies CogVideoX as the baseline. However, these critical details were indeed not included in the teaser Fig. 1. For clarification, the videos in Fig 1 were generated with the following text prompts: 'A creative train chugging along its track' and 'A creative dog running towards the camera'. We will revise Fig. 1 in the camera-ready version to explicitly indicate the baseline as CogVideoX, include representative video frames, and display the corresponding prompts to improve clarity and self-containment.

Q2: The proposed method lacks fine-grained control. The editing ability can be greatly enhanced if various more visual inputs can be supported (e.g., masks, arrows, etc.)

We thank the reviewer for highlighting this important consideration regarding fine-grained spatial control. We would like to clarify that our current implementation utilizes ControlNet as the generative backend, which inherently supports various types of visual conditioning, including depth maps, segmentation masks, edge maps, and more. Thus, as suggested by the reviewer, our framework is readily capable of incorporating additional visual inputs to enhance localization and precision. Moreover, our multi-agent framework autonomously determines and optimizes conditioning parameters (e.g., conditioning scales, guidance strength), enabling the system to dynamically adjust the granularity of edits from fine to coarse as necessary.

Due to NeurIPS rebuttal guidelines, we are unable to include additional visual examples or external links here. However, in our revised manuscript, we will clearly highlight this existing capability, include examples of fine-grained editing (such as performing a ‘creative edit’ on a small object (e.g. a vase) in a busy living room scene) and explicitly discuss the straightforward extensibility of our framework to incorporate additional visual inputs for even finer control.

Q3: In Fig. 3, the user’s initial intention or purpose is not clearly illustrated, and it’s hard to evaluate the visual quality of the generated images.

We apologize for any confusion regarding Fig. 3. To clarify, Fig. 3(a) illustrates our framework's "creative editing" capabilities. Specifically, the first column contains input images provided to our method, and columns two through four depict three "creatively edited" variations per input, produced without any manual intervention by the user. Additionally, the second set within Fig. 3(a) includes input images containing complex backgrounds, highlighting our framework's effectiveness in performing creative edits even when background elements are present.

Moreover, Fig. 3(b) demonstrates the "creative generation" ability of our method. Here, the system receives only a simple textual keyword (e.g., "a car" or "a dress") without any visual input, and autonomously generates corresponding images. Regarding the visual quality, we kindly direct your attention to the supplementary materials. Specifically, our supplementary website (under “website” folder) provides high-resolution images you can view in more detail. Furthermore, the supplementary document (NeurIPS_2025_CREA.pdf) includes an additional 600 examples generated by our framework. We will incorporate these clarifications into the final manuscript.

Q4: In the middle group of Fig. 3, the background is often modified when it is not supposed to be touched.

We acknowledge that minor unintended background modifications can occur even with state-of-the-art image editing methods. Such artifacts typically arise due to inherent limitations of inversion mechanisms: these approaches aim to reconstruct input images in latent space, but often produce slightly imperfect inversions that do not exactly match the original images. A more appropriate comparison would therefore involve evaluating our edited outputs against inverted images. Unfortunately, due to the updated NeurIPS rebuttal guidelines, we are unable to provide external URLs or supplementary PDFs containing these comparative images at this stage. However, we will include these comparisons in the final manuscript.

Q5: During the evaluation phase, only using creativity criteria is limited to obtain a comprehensive evaluation on the generated image. Metrics that measure aspects such as background preservation, identity preservation are missing.

We appreciate the reviewer’s insightful suggestion. Our initial evaluation already includes comprehensive perceptual metrics such as LPIPS, CLIP, and DINO (see Table 1, main paper), all demonstrating that our method outperforms competing approaches in both editing and generation tasks. Specifically, the DINO metric is particularly effective for assessing background and identity preservation. Notably, our method achieves the highest DINO score, indicating that the edited images generated by our approach preserve essential details.

However, as recommended, we have now incorporated additional quantitative metrics to further assess critical preservation aspects: FID and KID as detailed below.

And for generation:

The newly computed metrics reinforce our prior findings, further confirming the superior performance of our proposed method.

Q6: In the user study results, the proposed method does not show a very obvious advantage compared with the other baselines.

We would like to clarify that our user studies comprised two distinct evaluation tasks with separate criteria:

For the generation Task:

• Usability (Q1): This measures how accurately the generated images represent the specified object (e.g., a cup). All methods performed similarly on this metric, as each can generate images resembling the intended object.

• Creativity (Q2): This assesses the creativity of the generated images. Our method significantly outperformed the closest competitor (4.16 vs. 3.56), with the difference being statistically significant (p-value = 0.0056).

For the editing Task:

• Preservation (Q1): Our method showed a better performance in preserving original image details during editing, though the margin over baselines is modest.

• Creativity (Q2), our method achieved the highest score, though the numerical difference is smaller. We also emphasize that other methods usually generated “colorful” edits to reflect ‘creativity’, whereas our method was able to generate a diverse range of creative outputs.

Q7: In the comparisons with the baselines, the authors simply use “creative [object]” as the prompt, which is too ambiguous to guide the generation

We appreciate the reviewer’s point regarding the prompt ambiguity. However, we clarify that the baseline methods were explicitly provided with the prompt "creative [object]" (e.g., "a creative couch") precisely to reflect our task definition - creative editing with minimal guidance. Our framework takes only [object] names (e.g. a couch) and autonomously generates creative transformations guided by underlying creativity principles. To further substantiate fairness, we conducted extensive ablations (Sec. 4.3, Fig. 6d) demonstrating our method’s robustness and effectiveness across various prompting strategies.

Q8: Does CREA often tend to generate imaginative objects that do not exist in the real world?

We appreciate the reviewer’s question about the realism and applicability of CREA-generated outputs. Indeed, our framework intentionally emphasizes creative and imaginative transformations, however CREA is not limited exclusively to imaginative outputs. The underlying agentic pipeline is designed flexibly, allowing users to explicitly guide the degree of creativity or realism desired.

As demonstrated in Figure 3(a), CREA can reimagine familiar objects (like a dress or a table) in novel yet plausible ways. Also, CREA includes a user-in-the-loop mechanism (Section 3.4.1, Figure 5a) that allows users to guide outputs via stylistic cues or constraints (e.g., "+skull" to steer the generation to involve skull patterns), ensuring fine-grained control when desired. This control mechanism ensures that in high-fidelity domains, like product design or marketing, realism can be preserved while still benefiting from stylistic innovation. Quantitatively, CREA balances creativity and realism well, achieving top DINO (0.744), LPIPS (0.414), and VENDI (3.70) scores, indicating structural alignment, diversity, and coherence. While we are not allowed to share more realistic edits due to NeurIPS guidelines, we will add more realistic examples to the camera ready.

Q9: I am curious about the performance of a workflow that uses GPT-4o as both the agents and image generator.

As suggested, we experimented with GPT-4o for both agents and image generator, and obtained similar results. While we are not able to include the images here due to NeurIPS guidelines, these experiments will be included in the camera ready version.

Dear reviewer,

We appreciate your valuable review, and we addressed all your concerns in our rebuttal. As the NeurIPS discussion phase will conclude soon, we kindly remind you that if you have any further questions or require additional clarifications on our paper, we are more than happy to address them promptly.

Best,

Authors of CREA

We are glad to hear that the reviewer found our rebuttal detailed, as we devoted substantial time to addressing all concerns as thoroughly as possible. We also recognize that some points in the most recent review may stem from misunderstandings about our framework, which we hope to clarify below.

On “Most prompts shown are a creative < object >”

We respectfully note that this is not correct. The creative < object > format was only used for baselines for comparison. Since these baselines lack mechanisms to generate creative objects in a principled way, we followed prior state-of-the-art works (e.g., [1]) in adopting this simplified prompting strategy for them.

In contrast, our method’s prompts are highly detailed and autonomously generated by our agents according to creativity principles. This is the core contribution of our framework; rather than relying on simple phrases such as creative < object >, our system generates its own detailed, creativity-guided transformations. Evidence of this can be seen in Fig. 5 of our Appendix, which shows examples of the exact prompts our framework produces. Quoting the exact prompt in Fig 5 (Appendix) generated by our framework below:

Picture a luxurious couch crafted from an intricate blend of materials—velvet cushions that morph into soft clouds gently floating above, juxtaposed with richly textured marble armrests. This central piece, hovering against a pure white background, holds the eye with its elegant curves and vibrant emerald vines that cascade over its form, ending in delicate cherry blossom petals. Unexpectedly, geometric kaleidoscopes can be glimpsed within the cushions, offering a changeable pattern that weaves in and out of perception. 

This prompt was generated entirely by our multi-agent framework in accordance with creativity principles; something that would otherwise require substantial time, effort, and domain expertise from a human user. We acknowledge that this distinction could have been made clearer in the main paper and will revise the manuscript to explicitly include the exact prompts used.

[1]: Han et.al. 2025 "Enhancing Creative Generation on Stable Diffusion-based Models" (CVPR'25)

On “Unfair Comparison”

We respectfully disagree. Our work is the first to define creative editing as a research problem explicitly guided by creativity principles. Baselines do not offer capabilities beyond the creative < object > format. To ensure fairness, we used exactly the same established practices in creative image generation literature ([1] from CVPR'25) when designing baseline prompts.

Furthermore, we conducted comprehensive ablations to test alternative prompting strategies for baselines, including rare < object >, innovative < object >, unusual < object >, and negative prompts such as a normal < object > to steer generation toward more creative outputs. These results are presented in Fig. 6(d) of the main paper and Fig. 3 in the Appendix.

While our comparison protocol follows established state-of-the-art practices, we welcome the reviewer’s suggestions for alternative comparison strategies and would be happy to explore them.

[1]: Han et.al. 2025 "Enhancing Creative Generation on Stable Diffusion-based Models" (CVPR'25)

On “Limited Editing Capability”

Our framework is fully capable of editing complex, multi-object scenes (e.g., a realistic living room containing numerous distinct objects). While NeurIPS guidelines prevent us from sharing new images in the rebuttal, we conducted additional experiments on such complex scenes (for a similar request done by Reviewer mq9P earlier).

Specifically, we ran a user study with 39 participants evaluating CREA’s disentangled editing across multiple edit types (style, color change, object addition, removal, and creative transformations). Results showed overwhelming agreement that CREA preserved global structure while making targeted edits: 97.4% rated color changes as highly disentangled, 100% rated object addition, removal, and creative transformations as successful without undesired changes to other elements. We will include these quantitative results, visual examples, and GenAI-Bench comparisons in the final manuscript to demonstrate the framework’s versatility.

That said, we also wish to emphasize that the main contribution of our work is not to serve as a universal editing framework for all types of edits, but to establish the first creativity-driven editing framework in the literature. We provided over 600 visual results in the supplementary materials to illustrate its range and fidelity. We hope the reviewer will take this extensive evidence into account, as it already demonstrates the breadth of our framework’s capabilities, and we are prepared to expand these results in the final version with the more complex transformations created during the rebuttal period.