{"id":"407c6bb4-8d14-4c09-8c9c-a913d37afc1f","arxiv_id":"2501.18588","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A sketch-driven text-to-image interface with analogical inspiration and sketch scaffolding increases designers' self-reported inspiration, exploration, and co-creation compared to a ControlNet baseline.","lead":"Inkspire is a new sketch-based tool that helps product designers use AI image generators without getting stuck on one idea. In a user study, designers reported more inspiration, exploration, and a stronger sense of collaboration with Inkspire than with a standard ControlNet interface.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central comparison is confounded: Inkspire adds per-stroke auto-regeneration and sketch scaffolding that the ControlNet baseline lacks, so the reported exploration/inspiration gains may be driven by these interface affordances rather than by the analogical sketching loop itself.","rationale":"The reader's weakest assumption identifies exactly the concern that matters most: the ControlNet baseline is not matched on the interface affordances that Inkspire introduces, so the study supports a whole-system comparison but not a mechanistic attribution to analogical sketching. The interaction logs and BERTScore prompt-diversity measures are valuable converging evidence that the two conditions produced different workflows, and the paper is honest about its limitations, including the lack of ablation in Section 7.1. Self-report measures are also analyzed without multiple-comparison correction, and several reported p<0.01 effects would not survive a Bonferroni correction; however, the baseline confound is the more load-bearing issue because it threatens the causal interpretation of the central claim rather than just the statistical strength of individual tests. A per-stroke ControlNet condition would be a concrete, feasible check: it holds constant the most salient difference, automatic regeneration, while removing the analogy panel and scaffold. If the benefit survives that check, the analogical and scaffolding contributions are confirmed; if not, the paper should be reframed as showing the value of the whole interaction paradigm rather than of analogical sketching specifically. This is a revision-level concern, not a rejection-level one, so the conditional verdict stands.","tokens_in":21454,"tokens_out":4784,"duration_ms":53078,"concrete_test":"Add a third condition, 'Per-Stroke ControlNet': use the same ControlNet canvas and prompt box as the baseline, apply the dynamic guidance scale from Eq. 1, and regenerate after every stroke, but omit the analogy menu and the scaffolding underlay. Run the same within-subjects protocol (n=12) and compare exploration and inspiration against Inkspire. If the significant advantages disappear or effect sizes drop below ds≈0.5, the reported benefits are attributable primarily to the auto-regeneration affordance rather than to Inkspire's analogical and scaffolding design. If the advantages persist with this strengthened baseline, the fairness concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that Inkspire helps designers explore more and avoid fixation through analogical sketching and the sketch-to-design-to-sketch loop. The evidence is a within-subjects comparison against a ControlNet baseline that differs on at least four dimensions at once: the analogy menu, the sketch-scaffold underlay, automatic per-stroke regeneration with the dynamic guidance scale of Eq. 1, and the baseline's requirement that users finish a sketch before generating. Section 7.1 acknowledges that ablation studies are future work, but the confound is not merely a missing nuance: it concerns the most interactionally salient difference. In Inkspire, every stroke triggers a near-real-time generation, so the designer receives immediate visual feedback after minimal investment. In the baseline, participants must complete a full sketch and then explicitly generate, which the logged behavior confirms: baseline users averaged roughly 60 strokes before generating, while Inkspire users frequently interleaved one or a few strokes with generations. The self-report benefits in exploration (t(11)=3.94) and inspiration (t(11)=3.44) could therefore reflect the cost of committing to a full sketch in the baseline rather than the benefit of analogies or scaffolding. The paper's conclusion about 'analogical sketching' is thus not uniquely supported by the system-level comparison; it is equally compatible with the hypothesis that the main active ingredient is the auto-regeneration loop. This does not invalidate the system contribution, but it does weaken the specific causal attribution in the title and abstract.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents Inkspire, a sketch-driven text-to-image tool that combines an analogical inspiration panel (LLM-generated concrete analogies for abstract concepts), a sketching canvas with per-stroke auto-regeneration using a dynamic guidance scale, and a Design2Sketch pipeline that converts AI-generated designs into sketch-style underlays. The authors report a within-subjects study with twelve participants comparing Inkspire to a ControlNet-based baseline, finding significantly higher self-reported exploration, inspiration, controllability, communication, partnership, and attribution, and concluding that Inkspire promotes a more iterative, exploratory, and co-creative workflow that helps designers overcome fixation.","tokens_in":21728,"tokens_out":5296,"duration_ms":48179,"significance":"If the results hold, Inkspire is a meaningful step toward reducing design fixation in T2I workflows, offering a concrete system with a novel Design2Sketch pipeline and a validated evaluation using standard questionnaires. The paper reports large effect sizes, transparently presents non-significant results, and grounds its design goals in a professional design team exchange. However, the significance rests on the strength of the comparative evidence, which is currently confounded by multiple simultaneous differences between the conditions; therefore the precise contribution of analogical sketching is not yet established.","major_comments":[{"comment":"The central evaluation is a system-level comparison in which Inkspire differs from the baseline on at least four dimensions at once: the analogical inspiration menu, the sketch-scaffold underlay, automatic per-stroke regeneration with the dynamic guidance scale (Eq. 1), and the baseline's requirement to complete a full sketch before generating. The significant gains in exploration (t(11)=3.94) and inspiration (t(11)=3.44) could therefore be driven by the per-stroke feedback loop alone rather than by analogical sketching. Because the abstract and conclusion attribute the outcome to 'analogical sketching' and the sketch-to-design-to-sketch loop, this confound is load-bearing. Section 7.1 correctly lists ablation studies as future work, but the claims as written go beyond what the current comparison can establish. The authors should either narrow the central claim to the integrated system or add an analytic or experimental disaggregation (e.g., a condition without the analogy panel, or a baseline with per-stroke generation) to support the current wording.","section":"§5.3, §6.1, §7.1"},{"comment":"The logged sketching behavior reveals a strong asymmetry: baseline users averaged 59.8 strokes before generating, while Inkspire users averaged 17.3 strokes and interleaved one or a few strokes between generations. The baseline's full-sketch-before-generation requirement changes the cost of exploration, so the observed self-reported benefits may reflect this interaction economics rather than the value of analogies or scaffolds. To make the baseline a fair point of comparison, the authors need to justify that this requirement is representative of current practice, and ideally include a condition that gives ControlNet a comparable per-stroke interaction (as suggested in §7.1). Without such a control, the conclusion that analogical sketching is the active ingredient is not uniquely supported.","section":"§6.2.3"}],"minor_comments":[{"comment":"The dynamic guidance scale is a hand-selected formula; a brief sensitivity discussion or reference to tuning experiments would help readers understand why the constants were chosen.","section":"§4.2, Eq. (1)"},{"comment":"The BERTScore semantic-similarity comparison is reported as 'much lower' without a test statistic or p-value; a paired t-test or equivalent should be reported.","section":"§6.2.2"},{"comment":"The claim that participants drew fewer total strokes with Inkspire is not accompanied by a significance test; if the difference is not tested, it should be described as an observation only.","section":"§6.2.3"},{"comment":"There appear to be inconsistencies between the listed analogies and the 'Total' count for several participants (e.g., P1 lists seven items but reports six; P4 lists five items but reports four). Please check and correct.","section":"Table 1"},{"comment":"There is a typo in §4.3 ('acheve' should be 'achieve') and a doubled 'and' in §2.2 ('LLMs and and analogical reasoning').","section":"§4.3, §2.2"},{"comment":"The formative exchange session is described as a day-long session with seven designers, but no details on the protocol or analysis are provided; adding a brief description would strengthen the derivation of the design goals.","section":"§3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is transparent about its limitations and reports effect sizes and non-significant results carefully. However, the central claim about analogical sketching as the driver of the observed benefits is not uniquely supported by the current system-level comparison. An editorial decision to require either a revised framing or additional evidence would strengthen the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a real systems paper. The integrated loop — analogy menu, per-stroke generation with the dynamic guidance scale, and the Design2Sketch scaffolding — is new as a package, and the Design2Sketch pipeline (semantic boundaries intersected with soft edges) is a genuinely useful engineering trick. The formative session with automotive designers yields three sensible design goals. The within-subjects study is honestly executed: validated CSI and collaboration scales, significant large effects on exploration, inspiration, controllability, communication, partnership, and attribution, and transparent reporting of null results (engagement, expressiveness, design quality at p=0.06). The authors also acknowledge in Section 7.1 that the features need ablating.\n\nThe soft spot is not hidden, but it's bigger than the limitations section lets on. The baseline differs from Inkspire on at least four dimensions at once: the analogy panel, the scaffold underlay, automatic per-stroke regeneration with the guidance-scale equation, and the baseline's need to complete a sketch before generating. Logged behavior confirms this: baseline users averaged around 60 strokes per design before hitting generate, while Inkspire users interleaved one or two strokes with generations. So the large self-report differences in exploration and inspiration are just as consistent with 'low-cost, immediate visual feedback per stroke' being the active ingredient as with analogies or scaffolding. The honest claim is that the whole Inkspire interaction beats a standard ControlNet workflow; the specific claim that analogical sketching reduces fixation is underdetermined. That is a real weakness if the title and abstract are taken literally.\n\nI'm not recommending rejection. The system is citable and the study is a reasonable first demonstration. The confound is acknowledged and addressable: a follow-up condition that gives the baseline per-stroke regeneration, or one that removes the analogy panel, would isolate the mechanisms. Multiple comparisons aren't corrected and no code/data are released, but those are minor in a CHI systems paper of this type.\n\nThis paper deserves a serious referee. I'd send it to review with a request that the discussion of the confound be strengthened and the claims calibrated.","headline":"Solid systems paper with a confounded comparison: the workflow is new and the effects are large, but the specific claim about analogical sketching isn't uniquely supported by the data.","tokens_in":22300,"tokens_out":2902,"would_cite":true,"duration_ms":27073,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Inkspire claims a sketch-to-design-to-sketch loop with analogical inspiration helps designers explore more and avoid AI fixation.","keywords":["generative AI","sketching","iterative design","co-creative design","design fixation","analogical inspiration","text-to-image models","design exploration"],"falsifier":"A controlled experiment that augments the ControlNet baseline with automatic per-stroke regeneration and sketch scaffolding but not the analogical inspiration panel would locate the source of the exploration gain. If that augmented baseline shows the same exploration and inspiration ratings as Inkspire, the central claim that analogical sketching drives the effect is falsified. Alternatively, an objective divergence measure—counting the number of structurally distinct concepts in the final designs produced by each condition—would test whether self-reported exploration translates into more varied outcomes.","tokens_in":20,"feed_emoji":"🎨","tokens_out":4526,"duration_ms":95720,"temperature":0.7,"pith_summary":"The paper argues that the standard text-prompt-and-generate workflow of text-to-image tools nudges designers into design fixation, and that a sketch-driven loop can break that pattern. It introduces Inkspire, a prototype in which designers start from an abstract concept such as 'protectiveness,' receive concrete analogical inspirations (tortoise, armadillo, bunker), and then iterate with the AI stroke by stroke, with each new pen stroke producing a fresh design whose reduced-fidelity sketch scaffold is laid under the canvas. In a within-subjects study of twelve designers comparing Inkspire with a ControlNet baseline, participants rated Inkspire significantly higher on exploration, inspiration, controllability, communication, partnership, and attribution. The paper reads this as evidence that closing the sketch-to-design-to-sketch loop keeps designers co-creating with the model instead of fixating on a single thread. If the result holds, it suggests that interface design, not just model quality, is what determines whether generative AI expands or narrows the design space.","feed_headline":"Sketch loop with analogies widens design exploration","feed_subtitle":"A 12-designer study reports higher inspiration, exploration, and co-creativity than a ControlNet baseline.","key_machinery":"Three coupled mechanisms carry the argument. First, analogical inspiration: an LLM prompted with chain-of-thought reasoning converts an abstract concept into ten visually concrete objects drawn from nature, architecture, and fashion, giving designers concept-level anchors without prompt engineering. Second, per-stroke generation with a dynamic guidance scale G(n)=7-4·0.$5^{{n/3}}$, which starts near 3 when the canvas is nearly empty and asymptotes at 7 as strokes accumulate, so ControlNet can interpret incomplete sketches and re-render after every pen stroke while keeping the seed fixed for continuity. Third, the Design2Sketch pipeline converts the high-fidelity generation into a scaffold via Scaffolding = Boundary(Seg(D)) ∩ SoftEdge(D): semantic segmentation boundaries are intersected with HED soft edges to keep only key structural lines, which are then shown as a tracing-paper-style underlay. The scaffold is the hinge of the loop—it is the mechanism that turns a 'too complete' output back into something malleable.","core_discovery":"The central claim is that a complete feedback loop—sketch guides AI generation, AI output is abstracted back into a sketch scaffold, and the scaffold guides the next stroke—supports a more iterative, exploratory, and co-creative design workflow than the current practice of sketching a full image and handing it to a ControlNet model with a text prompt. The paper reports that the twelve participants rated Inkspire significantly higher than the baseline on inspiration (t(11)=3.44, p<0.01) and exploration (t(11)=3.94, p<0.01), as well as on controllability, communication, partnership, and attribution (all p<0.01). Interaction logs show Inkspire users sketching in short bursts with frequent generations, while baseline users drew long sequences and edited prompts incrementally; prompt semantic similarity was lower under Inkspire (BERTScore 0.51 vs. 0.76). The authors attribute these differences to three mechanisms: analogical inspiration that turns abstract briefs into concrete visual anchors, per-stroke regeneration with a dynamically increasing guidance scale that tolerates incomplete sketches, and sketch scaffolding that lets designers build on a generation without being fixated on its photorealistic finish.","pith_inferences":["If the analogical menu is the driver, a prompt-based tool with analogies but no scaffolding might reproduce part of the exploration gain, suggesting a possible ablation.","The measured reduction in prompt editing and the lower BERTScore similarity suggest that the interface substitutes conceptual pivots for lexical tweaks; whether this produces objectively more novel final designs is not fully settled by self-reports.","The single-thread limitation noted in the paper implies the loop may benefit from parallel analogy branches, which the authors themselves flag as future work.","Beyond product design, the sketch-to-design-to-sketch loop could be adapted to architecture or fashion sketching if the analogical source domains are re-targeted."],"forward_implications":["Designers can start ideation from a single abstract word and a single stroke rather than a fully specified prompt or sketch.","Per-stroke regeneration creates a turn-taking rhythm that makes the AI feel like a collaborator rather than a one-shot renderer.","The scaffolding underlay could be applied to other generative domains where high-fidelity outputs cause fixation.","The dynamic guidance scale provides a general recipe for making ControlNet-style models tolerate partial input."],"supporting_citations":[{"why":"Supplies the sketch-conditioned generation backbone (ControlNet) used in both Inkspire and the baseline condition.","marker":"[72]"},{"why":"Establishes the design-fixation phenomenon with text-to-image tools that Inkspire aims to counter.","marker":"[63]"},{"why":"The Creativity Support Index is the measurement instrument for the exploration and inspiration outcomes.","marker":"[22]"},{"why":"Supplies the human-AI collaboration questionnaire items for controllability, communication, partnership, and attribution.","marker":"[44]"},{"why":"Defines analogy as structure-mapping, the conceptual basis for the analogical inspiration prompt.","marker":"[29]"},{"why":"ShadowDraw motivates the guided-sketching underlay approach that Inkspire adapts into sketch scaffolding.","marker":"[46]"},{"why":"Chain-of-thought prompting is the technique used to generate analogical inspirations from the LLM.","marker":"[64]"},{"why":"Provides empirical evidence that ambiguous sketch underlays reduce fixation, grounding the scaffolding design.","marker":"[65]"}],"fun_headline_variants":["Sketch-feedback loop beats ControlNet for design exploration","Analogical sketching widens design ideas, study finds","Inkspire: AI tool helps designers explore more with sketches","From sketch to AI and back: more design inspiration","Sketch-AI feedback loop fosters design exploration"],"cache_read_input_tokens":24448,"weakest_assumption_plain":"The baseline ControlNet condition differs from Inkspire on several dimensions at once—analogical inspirations, sketch scaffolding, per-stroke regeneration, and the dynamic guidance scale—so the reported benefits may stem from any subset of these differences rather than from the complete loop, and the paper does not include an ablation to isolate them.","fun_headline_variants_meta":{"raw":{"variants":["Sketch-feedback loop beats ControlNet for design exploration","Analogical sketching widens design ideas, study finds","Inkspire: AI tool helps designers explore more with sketches","From sketch to AI and back: more design inspiration","Sketch-AI feedback loop fosters design exploration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00099,"raw_usage":{"total_tokens":4204,"prompt_tokens":961,"completion_tokens":3243,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":3164}},"tokens_in":577,"tokens_out":3243,"duration_ms":21040,"temperature":1.0,"reasoning_tokens":3164,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T22:52:41.211367+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled experiment that augments the ControlNet baseline with automatic per-stroke regeneration and sketch scaffolding but not the analogical inspiration panel would locate the source of the exploration gain. If that augmented baseline shows the same exploration and inspiration ratings as Inkspire, the central claim that analogical sketching drives the effect is falsified. Alternatively, an objective divergence measure—counting the number of structurally distinct concepts in the final designs produced by each condition—would test whether self-reported exploration translates into more varied outcomes.","supporting_citations":[{"cited_title":"Kelly, Saumya Pareek, Qiushi Zhou, and Eduardo Velloso","cited_arxiv_id":null,"evidence_quote":"Establishes the design-fixation phenomenon with text-to-image tools that Inkspire aims to counter."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the human-AI collaboration questionnaire items for controllability, communication, partnership, and attribution."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines analogy as structure-mapping, the conceptual basis for the analogical inspiration prompt."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ShadowDraw motivates the guided-sketching underlay approach that Inkspire adapts into sketch scaffolding."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides empirical evidence that ambiguous sketch underlays reduce fixation, grounding the scaffolding design."}],"review_version":1}