{"id":"d5a2a894-2dc4-4660-bd0c-b401a9d8f990","arxiv_id":"2504.15189","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"LACE integrates generative AI into Photoshop with layer-based editing and two collaboration modes, and a pilot study finds users prefer it over text-only workflows.","lead":"This paper presents LACE, a Photoshop-integrated tool that lets artists generate AI images as editable layers and switch between step-by-step and real-time AI collaboration. In a 21-person pilot study, most participants preferred LACE over text-only tools, reporting more control and stronger ownership of the results.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Image-to-image confound undermines attribution of LACE's benefits to its interface design; require a control condition that isolates interface from generative mode.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the comparison between W3 (LACE) and W1/W2 (text-to-image) confounds the interface design with the generative mode. This is the most direct threat to the central claim because the paper explicitly claims LACE's design (layer-based prompting, collaboration modes) drives the improvements. The concrete test would disentangle these factors by adding an image-to-image condition without layer-based editing. The verdict remains CONDITIONAL because the current evidence is preliminary and the longer version may address this, but the concern is substantive and should be resolved before accepting the attribution. The paper's own statement that comprehensive findings are in a longer version reinforces the conditional status, not a rejection, since the pilot is clearly labeled as preliminary and the system contribution has independent value.","tokens_in":3931,"tokens_out":3172,"duration_ms":28243,"concrete_test":"Add a fourth condition W4: image-to-image with latent consistency using the same base model and parameters but a minimal, non-layer interface (e.g., standard image-to-image in Photoshop without layer separation, or a simple web UI). Re-run the within-subject study (N=21 or larger) with the same tasks and measures. If W4 scores comparable to W3 (W3-W4 not significant), the advantage is from the image-to-image mode, not LACE's layer-based design; if W3 significantly exceeds W4, the design contribution is supported. Also report effect sizes and confidence intervals for all pairwise comparisons.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that LACE's layer-based, Photoshop-integrated interface improves usability, ownership, and satisfaction. The within-subject comparison (W1/W2 text-to-image vs W3 image-to-image) varies two factors simultaneously: the interface (layers vs text prompting) and the generative mode (image-conditioned vs text-only). 'All generative model parameters remained consistent' does not equate the conditioning modality. If image-to-image generation itself provides more control and consistency (e.g., via the input canvas), the observed preference for W3 could reflect the generative mode rather than the LACE design. Because the paper attributes the benefit to LACE's design (Section 2: layer-based prompting, collaboration modes), this confound directly threatens the central claim. No condition in the reported study isolates interface from mode, and the paper does not report data on mode preference within LACE to support the turn-taking/parallel stage claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents LACE, a Photoshop-integrated system for AI-assisted image generation that imports AI outputs as editable layers and supports both turn-taking and parallel human-AI collaboration modes. The authors report a within-subject pilot study with 21 participants who completed one of three art tasks using three workflows: W1 (text-to-image), W2 (text-to-image with latent consistency), and W3 (LACE, image-to-image with latent consistency). Quantitative results show significant Friedman test differences in satisfaction, ownership, usability, and artistic perception, and the paper claims that LACE significantly improves these outcomes compared with text-based workflows. The paper also reports qualitative preferences for turn-taking during early ideation and parallel interaction during refinement, and it explicitly states that this shorter workshop paper presents key insights while comprehensive findings are deferred to a longer version on arXiv.","tokens_in":4085,"tokens_out":2508,"duration_ms":24545,"significance":"If the central claim were fully supported, LACE would be a valuable contribution to professional creative tools by addressing three documented pain points: limited expressiveness of text prompting, difficulty in maintaining coherence during iterative refinement, and incompatibility with established artist workflows. The system's grounding in the COFI framework, its modular AI pipeline, and its integration into Photoshop are concrete design innovations. The study also provides a plausible qualitative account of how artists switch between turn-taking and parallel collaboration modes. Notably, the paper is transparent about its preliminary nature and points to a longer version for full analysis. However, the significance is currently bounded by a confounded experimental comparison and a small, underpowered sample, so the headline quantitative claims must be read as exploratory rather than conclusive.","major_comments":[{"comment":"The comparison between W3 (LACE) and W1/W2 varies two factors at once: the interface (layer-based, Photoshop-integrated editing versus text prompting) and the generative mode (image-to-image with latent consistency versus text-to-image). The statement in Section 3 that 'All generative model parameters remained consistent across conditions' does not equate the conditioning modality; image-to-image generation may itself offer more control, consistency, or perceived agency than text-to-image regardless of the LACE interface. Therefore, the paper's central claim that LACE's design improves usability, ownership, and satisfaction is not uniquely supported by the reported data. The authors should either add a control condition that isolates interface from generative mode (for example, LACE with a text-to-image pipeline, or a non-LACE image-to-image workflow) or substantially soften the causal language to describe the observed preference for the integrated LACE system.","section":"Section 3, Workflow comparison"},{"comment":"The reported Friedman test p-values are unadjusted for multiple comparisons, no effect sizes are given, and no details are provided about the post-hoc tests or their directions (e.g., which workflows differ pairwise and by how much). With N=21 and four outcome variables, the claim that 'LACE outperforms both text-only (W1) and text-based iterative (W2) approaches' is stronger than the reported statistics support. Please report the test statistic, degrees of freedom, adjusted p-values or confidence intervals, and effect sizes; alternatively, explicitly label the findings as exploratory. As written, the quantitative evidence is not sufficient to support the definitive 'significantly improves' statement in the abstract and introduction.","section":"Section 3.1, Quantitative analysis"},{"comment":"The paper itself states that 'the comprehensive findings and detailed analysis are presented in a longer version available separately on arXiv.' This is an explicit acknowledgment that the current manuscript lacks the full analysis needed to substantiate the headline quantitative claim. In light of this, the abstract's assertion that 'LACE significantly improves usability, user ownership, and overall satisfaction compared to baseline AI workflows' overstates the evidence presented in this paper. The authors should either include the comprehensive analysis in this manuscript or rephrase the claims to indicate that these are preliminary pilot results requiring confirmation.","section":"Abstract and Introduction"}],"minor_comments":[{"comment":"Figures 8, 9, and 10 are referenced in the appendix but lack in-text explanations and detailed captions; please add captions that describe the measures, scales, and what visual comparison the reader should draw.","section":"Figures 8-10"},{"comment":"The p-values for the Friedman test are reported without the associated chi-square statistic or degrees of freedom; adding these values would allow readers to assess the magnitude of the effects.","section":"Section 3.1, Statistical reporting"},{"comment":"The sampled qualitative results are presented as figures only; including representative participant quotes or a brief thematic summary would strengthen the qualitative findings.","section":"Appendix A"},{"comment":"The Spearman correlation analysis is mentioned but no coefficients or p-values are reported; please include the actual correlation matrix or at least the key coefficients.","section":"Section 3.1.2"}],"recommendation":"major_revision","confidential_remarks":"This is a workshop paper with a promising system and a relevant research question, but the experimental design conflates interface and generative mode. The authors should be encouraged to add a proper control condition in a future full version and to tone down the claims in this shorter paper. The reference to a longer arXiv version is appropriate for a workshop, but the current manuscript's abstract and introduction should not overstate what the pilot data can establish."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nLACE is a design contribution worth a look. The system—a Photoshop plugin that keeps AI outputs on separate layers and supports turn-taking and parallel collaboration, grounded in COFI—addresses a real pain point: artists want generative control inside tools they already use. The paper describes the system clearly and the workflow figures help. No fundamental new principle, but a sensible, practical integration.\n\nThe soft spot is the pilot study. The comparison runs W1/W2 (text-to-image) against W3 (LACE, image-to-image with latent consistency). Two variables change at once: the interface and the generative conditioning. 'All generative model parameters remained consistent' doesn't equate text conditioning with image conditioning. If image-to-image alone gives more control, part of LACE's preference could be an artifact of the generative mode, not the layer-based UI. There's no condition that holds the mode fixed while varying interface. The stress-test note lands; this is the load-bearing flaw.\n\nThe statistics are also thin. N=21, Friedman with unadjusted p-values, no effect sizes, no correction, Likert self-report. The paper itself says the comprehensive analysis is in a longer version, which is honest, but then the abstract's word 'significantly' overstates what this pilot can show.\n\nCredit where due: the qualitative observations—turn-taking for early ideation, parallel for refinement—are plausible and worth exploring. The preference data, even confounded, is a signal. The paper doesn't hide its limitations and the citation pattern is normal.\n\nWho should read it: HCI and creativity-support researchers, especially people building professional creative tooling. I'd send it to peer review—the design discussion is valuable and the confound is addressable in a follow-up. But the current version's central claim should be softened, or better, backed with an isolating control condition. My own verdict is conditional; I lean positive on the design, skeptical on the evidence.","headline":"A useful design contribution with a real, nameable confound: LACE's benefits over text-only baselines are plausible but the pilot doesn't isolate the interface from the image-to-image generative mode.","tokens_in":4563,"tokens_out":2227,"would_cite":false,"duration_ms":20822,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LACE embeds generative AI in Photoshop as editable layers, and a 21-person pilot finds artists rate this workflow significantly higher on usability, ownership, and satisfaction than text-only prompting.","keywords":["human-AI co-creativity","generative AI","creative professionals","Photoshop","layer-based editing","turn-taking collaboration","parallel interaction","artistic control"],"falsifier":"Run the same within-subject comparison with the generative model mode held constant: give the text-based conditions the same image-to-image latent-consistency backend that LACE uses, or give LACE a text-to-image backend. If the usability and ownership advantage disappears, the effect cannot be attributed to LACE's layer-based design.","tokens_in":3748,"feed_emoji":"🎨","tokens_out":7849,"duration_ms":63265,"temperature":0.7,"pith_summary":"This paper claims that the main barrier to generative AI in professional art is not image quality but control: text prompts cannot express fine visual intent, and flattened outputs cannot be iterated on. LACE answers by bringing AI-generated images into Photoshop as separate layers that artists can edit, mask, blend, and composite while the AI pipeline keeps generating. A within-subject pilot with $N=21$ participants found that LACE significantly outperformed two text-based workflows on usability, ownership, and satisfaction, with 71.4% of participants choosing it as their favorite. The paper's central observation is that artistic agency over AI outputs, not the prompt alone, is what makes generative tools feel usable to professionals.","feed_headline":"LACE's editable layers beat text prompting in artist pilot","feed_subtitle":"In a 21-person pilot, it scored higher on usability, ownership, and satisfaction than text-only workflows.","key_machinery":"The central mechanism is LACE (Latent Auto-recursive Composition Engine), a Photoshop-integrated system that imports each AI output as an individual layer rather than a flattened image, so artists can isolate foreground, mid-ground, and background elements with Photoshop's native masks, blending modes, curves, and adjustments. It also implements two COFI-grounded collaboration modes: turn-taking, where prompts and outputs alternate step by step, and parallel mode, where AI adapts to snapshots of the evolving canvas in real time. A modular AI pipeline lets users swap or configure the underlying generative model. This mechanism separates the what of generation from the where and how of composition, letting iterative refinement happen through image editing instead of prompt rewriting.","core_discovery":"The central claim, on the authors' own terms, is that LACE's combination of layer-based prompting and flexible turn-taking and parallel collaboration modes materially improves the human-AI creative process. In their pilot, Friedman tests showed significant differences across workflows for satisfaction ($p=0.039$), ownership ($p=0.009$), usability ($p=0.003$), and artistic perception ($p=0.005$), with post-hoc comparisons favoring LACE over both text-only and text-based-iterative workflows. Participants preferred turn-taking in early ideation and parallel modes for later refinement, and usability ratings correlated with ownership and satisfaction only in LACE, not in text-based workflows. The authors interpret this as evidence that direct manipulation of layered AI outputs gives artists a sense of control that text prompting cannot.","pith_inferences":["A stronger test would hold the generative backend fixed across all conditions, giving the text-based workflows the same image-to-image latent-consistency pipeline LACE uses, to show whether the preference is driven by the interface or by the generative mode.","If user agency is the active ingredient, then other direct-manipulation interfaces over AI outputs, not only layer-based editing, should reproduce some of the ownership gain; this pilot does not test that.","The mixed feedback on the pixel-art task suggests layer editing can cost extra time, so a follow-up could measure whether presets or training close that gap for novices.","Comparing self-reported ownership with expert-rated output quality would separate perceived control from objective quality, a distinction the current metrics leave open."],"forward_implications":["If LACE is right, professional artists can adopt generative AI without leaving the editing environment they already use in production.","Layer-based outputs mean later revisions no longer require regenerating a whole flattened image; artists can keep the elements they like and replace the rest.","Because ownership and usability are correlated only in LACE, tools that give users direct control over generated content may reduce the detached feeling reported in pure text workflows.","Task-stage preferences suggest co-creative systems should offer both turn-taking and parallel modes rather than a single interaction paradigm.","The same layer-and-direct-manipulation pattern could extend beyond image editing, as the authors note, to 3D rendering and animation pipelines."],"supporting_citations":[{"why":"Supplies the COFI framework that defines the turn-taking and parallel collaboration modes LACE implements.","marker":"[8]"},{"why":"Documents that users do not perceive ownership of AI-generated text, the problem LACE's layer-based editing is meant to address.","marker":"[4]"},{"why":"Documents creative practitioners' reluctance to adopt AI tools that disrupt established workflows, motivating Photoshop integration.","marker":"[7]"},{"why":"Shows that generative workflows are highly sensitive to input parameters, motivating LACE's emphasis on iterative coherence.","marker":"[9]"},{"why":"Supplies fine-grained visual prompting as a generative control mechanism in LACE's pipeline.","marker":"[10]"},{"why":"Supplies disentangled image-editing control used in the AI pipeline.","marker":"[5]"}],"fun_headline_variants":["LACE layer control outranks text prompts in 21-artist study","Artist pilot: LACE's turn-taking and parallel modes win on usability","LACE's layer-based prompts give artists a greater sense of control","Pilot: LACE beats text prompts on ownership and satisfaction","LACE's hybrid modes: turn-taking for ideas, parallel for detail"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the three workflows differ only in interface and interaction mode, so any preference for LACE is caused by its design rather than by the fact that W3 uses image-to-image generation with latent consistency while W1 and W2 use text-to-image generation.","fun_headline_variants_meta":{"raw":{"variants":["LACE layer control outranks text prompts in 21-artist study","Artist pilot: LACE's turn-taking and parallel modes win on usability","LACE's layer-based prompts give artists a greater sense of control","Pilot: LACE beats text prompts on ownership and satisfaction","LACE's hybrid modes: turn-taking for ideas, parallel for detail"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000848,"raw_usage":{"total_tokens":3606,"prompt_tokens":778,"completion_tokens":2828,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":394,"completion_tokens_details":{"reasoning_tokens":2734}},"tokens_in":394,"tokens_out":2828,"duration_ms":17162,"temperature":1.0,"reasoning_tokens":2734,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:30:21.403813+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same within-subject comparison with the generative model mode held constant: give the text-based conditions the same image-to-image latent-consistency backend that LACE uses, or give LACE a text-to-image backend. If the usability and ownership advantage disappears, the effect cannot be attributed to LACE's layer-based design.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the COFI framework that defines the turn-taking and parallel collaboration modes LACE implements."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that generative workflows are highly sensitive to input parameters, motivating LACE's emphasis on iterative coherence."},{"cited_title":"A man reaching for a painting","cited_arxiv_id":null,"evidence_quote":"Supplies fine-grained visual prompting as a generative control mechanism in LACE's pipeline."}],"review_version":1}