Pith. sign in

REVIEW 4 major objections 5 minor 18 references

LACE: Exploring Turn-Taking and Parallel Interaction Modes in Human-AI Co-Creation for Iterative Image Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read The paper claims that blending turn-taking and parallel human-AI interaction in one image-generation tool gives users higher perceived ownership, usability, and art perception than turn-taking alone.

desk verdict A real system with a promising idea, but the pilot evaluation confounds interaction mode with editing capability, so the headline claims about mode are not yet supported. read the letter →

arxiv 2504.14827 v1 pith:E5GW75P7 submitted 2025-04-21 cs.HC

classification cs.HC
keywords LACEhuman-AIco-creationturn-takinginteractionparallelgenerativeimagegenerationcreativitysupporttoolsiterativerefinementuserstudy
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

LACE is a Photoshop-integrated system that lets artists generate images with AI in two interleaved ways: turn-taking, where user and AI alternate contributions, and parallel, where the AI keeps producing candidate images in the background while the artist edits on independent layers. The paper's central claim is that this hybrid style gives users a stronger sense of ownership, higher perceived usability, and a stronger sense that the result is art than purely turn-taking text-to-image workflows. In a within-subjects pilot with 21 participants, the hybrid workflow scored significantly higher on those measures ($p \le 0.002$ for ownership, $p \le 0.005$ for usability, $p \le 0.009$ for art perception). The authors interpret this as evidence that flexible participation modes let artists switch between AI-led ideation and hands-on refinement without losing control.

What carries the argument

The central mechanism is the dual-feedback loop: an AI loop continuously generates candidate images from the current prompt while an artist loop lets the user edit the canvas in real time; generated candidates are cached and can be imported into Photoshop as new layers, and imported results feed back into the AI pipeline. An influence weight between 0 and 1 controls how much the next generation reflects the artist's current canvas, making local editing an explicit input to the model. Because the interface never forces an explicit mode switch, the same pipeline behaves as turn-taking when the user ignores incoming candidates and as parallel when the user edits while new suggestions arrive.

What would settle it

Run the same within-subjects comparison with a fourth condition: a turn-taking workflow that still offers Photoshop layers, direct canvas editing, and parameter adjustment but generates only one output per user turn. If that condition matches the hybrid workflow's scores, the gains come from editing capabilities rather than parallel interaction; if the hybrid still wins, the parallel mode is doing the work.

Watch

Extended reading notes

Core claim

The paper claims that participation style is a real design lever in human-AI co-creation: supporting both sequential alternation and simultaneous work in a single pipeline improves how much control, usability, and artistic value people perceive in the output. Concretely, LACE combines turn-taking and parallel modes in one Photoshop-based workflow, and in the study the hybrid workflow (W3) significantly outperformed two turn-taking baselines on perceived ownership, usability, and art perception, with overall Friedman differences across workflows on satisfaction, ownership, usability, and art perception. The authors frame the result as preliminary, not a full comparison of the two modes, but as evidence that flexible interaction modes can enhance creative control and authorship.

Load-bearing premise

The load-bearing premise is that the higher scores for LACE come from the flexibility of turn-taking versus parallel interaction, but LACE also gives users Photoshop layers, direct canvas editing, and parameter adjustment that the two turn-taking baselines lack, so the effect of the participation mode itself is not isolated.

Editorial extensions

If this is right

  • Artists with a clear vision can edit directly while AI offers alternatives, while artists exploring can fall back on turn-taking to let the AI generate starting points.
  • A single tool can serve both early ideation and late refinement, so interface designers no longer have to choose one participation style.
  • Delivering AI outputs as editable layers, rather than final images, is one way to preserve user authorship in generative workflows.
  • Preferred mode depends on task: open-ended design challenges favored turn-taking, while representational and abstract tasks favored the hybrid parallel workflow.
  • Self-reported art perception, usability, and ownership move together with the hybrid workflow, suggesting these dimensions are aligned in co-creation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • My reading: the study treats W3 as parallel/hybrid, but the baselines also lack layer-based editing, so the cleanest test of the participation-mode claim would add a turn-taking-with-layers condition.
  • My reading: if the effect generalizes, the design pattern of background suggestions plus editable layers could transfer to other generative media such as music, 3D scenes, or video, where authorship is also threatened by one-shot outputs.
  • My reading: users' mode preference may be predictable from creative stage, so a future system could automatically suggest switching from turn-taking to parallel when canvas edits start accumulating.
  • My reading: the preference data suggest a complementary mapping, with turn-taking for divergent early exploration and parallel for convergent refinement, though the authors note this is context-dependent and not a strict division.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces LACE, a Photoshop-integrated co-creative system that supports both turn-taking and parallel interaction modes for iterative image generation. The authors report a within-subjects pilot study with 21 participants who used three workflows: W1 (basic turn-taking), W2 (iterative turn-taking), and W3 (LACE's parallel/hybrid mode). They measured ownership, satisfaction, usability, alignment with expectations, and art perception, and report significant Friedman and Wilcoxon results favoring W3 on ownership, usability, and art perception. The paper claims that flexible interaction modes enhance creative control, authorship, and workflow compatibility, and it discusses task-dependent preferences for turn-taking versus parallel interaction.

Significance. If the central claim were valid, LACE would be a useful contribution to co-creative systems, and the empirical comparison of participation modes would inform the design of human-AI creative tools. The paper has strengths: a concrete system integrated with professional software, a within-subjects design with random task assignment, non-parametric statistical tests with reported effect sizes, qualitative participant quotes, and explicit acknowledgment of several limitations. However, the load-bearing empirical comparison is confounded: W3 differs from W1 and W2 not only in interaction style but also in the available editing capabilities. The reported evidence therefore supports a weaker claim about preferences for a more capable editing system rather than the stated claim about participation modes. The paper is a reasonable workshop-scale system and pilot study, but the central contribution about turn-taking versus parallel interaction needs reframing or additional controlled evidence.

major comments (4)
  1. [§4, Table 1, §5] The headline comparison of W3 versus W1/W2 varies two factors at once: participation style (parallel/hybrid versus turn-taking) and editing interface (Photoshop layers, direct canvas editing, and parameter adjustment are available only in W3). The reported ownership (p=0.002/p=0.001), usability (p=0.005/p<0.001), and art perception (p=0.001/p=0.009) differences therefore cannot be attributed specifically to the parallel/hybrid interaction mode. This is not a peripheral concern: the abstract and introduction frame the contribution as being about interaction modes. The manuscript itself supplies evidence of the confound in Appendix A, which attributes W2's lower quality to 'uninterpretable latent interactions' and W3's enhanced agency to 'parameter adjustments or direct canvas edits.' The current design cannot rule out that the results are due to direct editing tools rather than to flexible participation styles.
  2. [§5] The Friedman test for satisfaction is reported as significant (p=0.039), but no post-hoc pairwise test is reported for satisfaction. The text says participants reported 'significant improvements in key metrics' and lists satisfaction among them, but the pairwise support for satisfaction is missing. Please either report the pairwise comparison or remove satisfaction from the list of metrics with demonstrated pairwise improvement.
  3. [§5] Multiple comparisons are not adjusted. For each dependent variable, three workflows yield three pairwise Wilcoxon tests, and six such tests are reported in the key comparisons (and more if expectation and other metrics are included). With p-values such as 0.009, some results may not survive a family-wise error correction. Please report adjusted p-values or explicitly justify the uncorrected exploratory comparisons.
  4. [Abstract, §1] The abstract and introduction state that LACE demonstrates significant improvements 'compared to standard AI workflows,' but the experiment compares W3 only with W1 and W2, which are LACE-internal turn-taking workflows, not standard AI tools such as MidJourney or Stable Diffusion. In addition, the introduction says the study 'does not provide a full comparative analysis of parallel versus turn-taking modalities,' which is in tension with the wording of the abstract. Please align the claims with what was actually compared.
minor comments (5)
  1. [§4] There are several presentation issues in the procedure description: 'For polit test, we recruited' contains a typo, and the workflow table is referenced as 'see table ??' without a resolved table number. Please correct these.
  2. [§4, §5] The questionnaire items are described only as 7-point Likert scales for ownership, satisfaction, usability, expectations, and art perception. Please provide the exact items or an appendix reference, since single-item measures have reliability limitations that should be acknowledged.
  3. [Figure 5] The figure caption says red asterisks indicate significant differences, but it does not state which pairwise comparisons are shown or whether the chart displays means, medians, or distributions. Please clarify the axis, the statistic plotted, and the specific comparisons represented by the asterisks.
  4. [§5] The text says 'Kendall’s W and Cohen’s reported as effect sizes,' but the effect size for Cohen's is not named (Cohen's d versus Cohen's r). The reported r values appear to be rank-biserial or similar; please specify the effect size formula used.
  5. [§6] The limitations section is candid and appropriate, but it could be more explicit that the W3 confound is a threat to the central participation-mode claim, not merely a limitation on generalizability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper reports an empirical within-subjects user study with no fitted parameters, mathematical derivation, or load-bearing self-citation chain.

full rationale

LACE makes no formal derivation that could reduce to its own inputs. The central claim is an empirical comparison of three workflows using Likert-scale questionnaires and non-parametric tests (Friedman and Wilcoxon signed-rank). No parameter is fitted from the outcome data and then renamed as a prediction; no quantity is defined in terms of the outcome it is supposed to explain; and the cited frameworks (COFI and the CST review) are external prior work, not self-citations by the present authors. The paper even presents its results as preliminary observations rather than as a forced consequence of a model. The main threat identified in the reader's take is confounding: Workflow 3 differs from W1 and W2 not only in participation style but also in available editing capabilities such as Photoshop layers, direct canvas editing, and parameter adjustment, and the paper's own Appendix attributes part of W2's weakness to uninterpretable latent interactions rather than to turn-taking itself. That is a validity concern about whether the manipulation isolates the intended construct, not a circularity concern: the empirical measurements are not definitionally equivalent to the claimed construct. The limitations section openly acknowledges that parallel and hybrid modes were grouped into one workflow and that future work should separate the conditions, further confirming that the authors do not attempt to derive the effect from its own definition. Because there is no self-referential derivation, no fitted input presented as a prediction, and no author-imported uniqueness theorem, the appropriate circularity score is 0.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

There are no fitted numbers; the study is empirical. The listed axioms capture the interpretive and procedural assumptions the results depend on, especially the mode-attribution assumption that is not isolated by the experimental design.

assumptions (4)
  • domain assumption Self-reported 7-point Likert ratings of ownership, satisfaction, usability, expectation, and art perception are treated as valid measures of those constructs.
    All quantitative conclusions in Section 5 come from these ratings; no validation study or questionnaire items are provided.
  • domain assumption The 21-participant within-subjects sample is adequate for the Friedman and Wilcoxon analyses.
    Sample is small, convenience-recruited, mostly students in CS or digital art; generalizability is acknowledged as limited in Section 6.
  • ad hoc to paper Observed advantages of W3 can be attributed to flexible interaction modes rather than to other features of LACE.
    Table 1 gives W3 additional capabilities such as layer import, direct editing, and parameter adjustment that are not present in W1/W2; without a controlled comparison, this attribution is an assumption.
  • domain assumption The procedure described in Section 4 is internally consistent and was executed as written.
    The text contradicts itself on whether participants were assigned one task or completed all three prompts, and on time limits, so the reader must assume a consistent protocol.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LACE: Exploring Turn-Taking and Parallel Interaction Modes in Human-AI Co-Creation for Iterative Image Generation." pith.science (2026). https://pith.science/paper/E5GW75P7

@misc{pith2026250414827,
  author       = {Pith},
  title        = {Pith review of: LACE: Exploring Turn-Taking and Parallel Interaction Modes in Human-AI Co-Creation for Iterative Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E5GW75P7}},
  note         = {Machine review of arXiv:2504.14827}
}
read the original abstract

This paper introduces LACE, a co-creative system enabling professional artists to leverage generative AI through controlled prompting and iterative refinement within Photoshop. Addressing challenges in precision, iterative coherence, and workflow compatibility, LACE allows flexible control via layer-based editing and dual-mode collaboration (turn-taking and parallel). A pilot study (N=21) demonstrates significant improvements in user satisfaction, ownership, usability, and artistic perception compared to standard AI workflows. We offer comprehensive findings, system details, nuanced user feedback, and implications for integrating generative AI in professional art practices.

Figures

Figures reproduced from arXiv: 2504.14827 by the authors.

Figure 1
Figure 1. Main interface of LACE, integrated into Adobe Photoshop [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The LACE system supports two interaction modes: [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. System architecture of LACE 3 LACE SYSTEM LACE builds on literature in digital art [4, 12], AI-assisted image creation [2, 3], and control in generative models [6, 18]. It enables precise control over image generation, supports iterative refinement for creative continuity, and integrates seamlessly into professional workflows by extending Adobe Photoshop. Drawing on the Co-Creative Framework for Interaction (COFI) [… view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: The charts compare Workflow 1 and Workflow 2 [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Preference of Workflow by Tasks involvement. Similarly, usability scores for LACE are significantly higher compared to W1 ( 𝑧 = -2.54 , 𝑝 = 0.005 ,𝑟 = 0.56 ) and W2 ( 𝑧 = -3.33 , 𝑝 < 0.001 , 𝑟 = 0.73 ). For art perception, LACE is rated higher than both W1 ( 𝑧 = -3.24 …
Figure 7
Figure 7. Figure 7: Results from Task 1. Prompt: A man reaching for a painting in an art gallery, accompanied by a dog sniffing another artwork on the floor.” [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Results from Task 2. Prompt: “An abstract composition that embodies the dynamics and motion associated with joy.” [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Results from Task 3. Prompt: “A pixel art game scene with a bustling cityscape featuring assorted architectural styles.” [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 10 canonical work pages

  1. [1]

    Abigail Batley and Richard Glithro. 2024. Exploring the Synergy of AI Generative Fill in Photoshop and the Creative Design Process Utilising Informal Learning. In DS 131: Proceedings of the International Conference on Engineering and Product Design Education (E&PDE 2024), Hilary Grierson, Erik Bohemia, and Lyndon Buck (Eds.). The Design Society, 1–6. http...

  2. [2]

    Katica Bozsó, András Béres, and Bálint Gyires-Tóth. 2024. Semantic segmentation mask-guided diffusion models: A pathway to enriched datasets in autonomous systems. 79–84. https://doi.org/10.3311/WINS2024-014

  3. [3]

    Yu Cao, Xiangqiao Meng, PY Mok, Xueting Liu, Tong-Yee Lee, and Ping Li. 2023. AnimeDiffusion: Anime Face Line Drawing Colorization via Diffusion Models. arXiv preprint arXiv:2303.11137 (2023)

  4. [4]

    Ryan Daniel. 2022. The creative process explored: Artists’ views and reflections. Creative Industries Journal 15, 1 (2022), 3–16

  5. [5]

    Fan, Monica Dinculescu, and David Ha

    Judith E. Fan, Monica Dinculescu, and David Ha. 2019. CollabDraw: An Environ- ment for Collaborative Sketching with an Artificial Agent. In Proceedings of the 2019 Conference on Creativity and Cognition . 556–561

  6. [6]

    Aosong Feng, Weikang Qiu, Jinbin Bai, Kaicheng Zhou, Zhen Dong, Xiao Zhang, Rex Ying, and Leandros Tassiulas. 2024. An Item is Worth a Prompt: Versatile Image Editing with Disentangled Control. arXiv preprint arXiv:2403.04880 (2024)

  7. [7]

    Gerhard Fischer. 2004. Social creativity: turning barriers into opportunities for collaborative design. In Proceedings of the Eighth Conference on Participatory Design: Artful Integration: Interweaving Media, Materials and Practices - Volume 1 (Toronto, Ontario, Canada) (PDC 04). Association for Computing Machinery, New York, NY, USA, 152–161. https://doi...

  8. [8]

    Peyman Gholami and Robert Xiao. 2024. Streamlining Image Editing with Layered Diffusion Brushes. arXiv:2405.00313 [cs.CV] https://arxiv.org/abs/2405.00313

Show all 18 references
  1. [9]

    Takashi Ikegami and Hiroyuki Iizuka. 2007. Turn-Taking Interaction as a Coop- erative and Co-Creative Process. In Infant Behavior and Development . 278–288

  2. [10]

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. 2022. Repaint: Inpainting using denoising diffusion proba- bilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11461–11471

  3. [11]

    MidJourney. 2022. MidJourney - Official Website. https://www.midjourney.com/. Accessed: 2024-09-15

  4. [12]

    Takeshi Okada and Sawako Yokochi. 2024. Process Modification and Uncontrol- lability in an Expert Contemporary Artist’s Creative Processes. The Journal of Creative Behavior (2024)

  5. [13]

    Jeba Rezwana and Mary Lou Maher. 2019. A User-Centered Framework for Human-AI Co-Creativity. Journal of Creative Computing 12, 3 (2019), 45–60

  6. [14]

    Jeba Rezwana and Mary Lou Maher. 2023. Designing Creative AI Partners with COFI: A Framework for Modeling Interaction in Human-AI Co-Creative Systems. In Proceedings of the CHI Conference on Human Factors in Computing Systems . ACM, 1–28

  7. [15]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis With Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 10684–10695. GenAICHI 2025, April ...

  8. [16]

    Ben Shneiderman. 2002. Creativity support tools. Commun. ACM 45, 10 (Oct. 2002), 116–120. https://doi.org/10.1145/570907.570945

  9. [17]

    Skov, and Jesper Kjeldskov

    Niels van Berkel, Mikael B. Skov, and Jesper Kjeldskov. 2021. Human-AI inter- action: intermittent, continuous, and proactive. Interactions 28, 6 (Nov. 2021), 67–71. https://doi.org/10.1145/3486941

  10. [18]

    An abstract composition that embodies the dynamics and motion associated with joy

    Lingfeng Yang, Yueze Wang, Xiang Li, Xinlong Wang, and Jian Yang. 2024. Fine- grained visual prompting. Advances in Neural Information Processing Systems 36 (2024). LACE GenAICHI 2025, April 27, 2025, Yokohama, Japan and Online A Appendix: Sampled Qualitative Results In this a...

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.