{"id":"d232cafb-a458-42ad-926c-338cbeed871a","arxiv_id":"2607.03731","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A staged conversational co-design pipeline lets non-experts produce VR-ready 3D assets that increase scene engagement and modulate affect, with a clear 2D-to-3D quality drop and no authorship leniency.","lead":"CoGen3D is a three-stage AI pipeline that helps non-experts make 3D objects for VR by chatting, confirming a 2D concept image, then generating a deployable 3D mesh. A 120-person study found these objects raised scene dwell time and shifted emotion ratings, though people preferred the 2D previews over the final 3D models.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Primary with/without-asset claims rest on a between-cohort contrast that confounds asset presence with participant differences, order, and novelty.","rationale":"The reader correctly isolates the load-bearing soft spot: the primary affective and engagement results are between-cohort rather than a clean within-subject or yoked asset-presence manipulation. The pipeline itself, quality-gap findings, conversation analyses, and lack of authorship leniency are solidly reported and do not require the same causal leap. Because the reader already flags this exact assumption and issues a CONDITIONAL verdict pending stronger isolation of asset vs. cohort/novelty effects, no verdict change is warranted. The concrete within-subject re-analysis (or explicit reporting of Designers’ post numbers) is the single check that would most directly settle whether the concern lands.","tokens_in":40519,"tokens_out":565,"duration_ms":14584,"concrete_test":"Re-fit the §5.3 ART ANOVA and §5.4.1 Gamma GLMM using Designers’ within-subject post-design (own assets) versus pre-design as the asset-present arm (or report the already-collected post numbers side-by-side with the Validation contrast). If the within-subject effect sizes on dwell and SAM match or exceed the between-cohort magnitudes after the same covariates, the asset-presence attribution holds; if they shrink substantially, cohort/novelty confounds are material and the central behavioral claim weakens.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim attributes higher scene engagement (Gamma GLMM, §5.4.1: Group χ²(1)=82.19, p<.001; ~1.8–2.3× dwell) and scene-dependent SAM shifts (ART ANOVA Group×Scene interactions for valence/arousal/dominance, §5.3) to insertion of co-designed assets. The primary contrast, however, is Designers’ pre-design (asset-free) ratings/telemetry versus an entirely separate Validation cohort that only ever saw populated scenes (§4.1–4.2, §4.4, §5.3–5.4). Covariates (age, gender, VR experience) are partial; they do not eliminate unmeasured cohort differences, the fact that validators never experienced the asset-free baseline, or novelty/exploration effects of first encountering any novel objects. Designers’ own post-design (own-asset) data exist but are deliberately de-emphasized for the headline models to avoid IKEA confounds, leaving the causal isolation of “asset presence” under-supported relative to the claim’s strength.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper presents CoGen3D, an agentic human–AI co-design pipeline that scaffolds non-expert VR asset authoring through conversational intent elicitation, mandatory 2D concept-image confirmation, and image-to-3D generation with direct Unity deployment. A between-subjects study (N=120) uses six validated affective VR scenes: 60 Design participants co-create one asset per scene after an asset-free baseline, and 60 Validation participants experience scenes populated with those assets. Analyses (CLMM for ordinal satisfaction, ART ANOVA for SAM, Gamma GLMMs for dwell and interaction telemetry, conversation logs) report higher scene engagement and scene-dependent SAM shifts when assets are present, a systematic preference for 2D concept images over final 3D meshes with no Group×Modality authorship leniency, and environment-shaped conversational pacing. The authors argue that staged, intent-based co-design can democratize VR authoring.","tokens_in":40816,"tokens_out":1189,"duration_ms":19489,"significance":"If the claims hold under tighter causal framing, this is a solid systems-and-evaluation contribution to HCI/XR authoring: an open, instrumented multi-stage pipeline; a large study on validated affective stimuli with appropriate mixed models, FDR post-hocs, and covariates; and actionable design implications (2D confirmation gates, scene-aware spawn heuristics, limits of psychological ownership in generative co-creation). Strengths include end-to-end deployability, dual designer/validator evaluation of the same assets, rich telemetry (≈477k spatial updates), and conversation/prompt analyses that link target environments to design reasoning. The work is timely given rapid text-to-image and image-to-3D progress and the persistent gap between unconstrained prompting and usable immersive authoring.","major_comments":[{"comment":"§4.1–4.2, §4.4, and primary models in §5.3–5.4: The headline engagement and affective claims attribute higher dwell (Group χ²(1)=82.19, p<.001; ~1.8–2.3×) and Group×Scene SAM interactions to insertion of co-designed assets, but the primary contrast is Designers’ pre-design (asset-free) measures versus an entirely separate Validation cohort that only ever saw populated scenes. Covariates (age, gender, VR experience) do not remove unmeasured cohort differences, order, or novelty/exploration of any novel objects. Designers’ own post-design (with-asset) data exist but are de-emphasized for the primary models. Soften causal language to between-cohort association, or add and foreground within-designer pre/post analyses (and/or validators with a true asset-free arm) so asset presence is not confounded with participant and protocol differences.","section":null},{"comment":"§5.3 and Discussion §6.1 (RQ3): Relatedly, the paper states that assets “shifted emotional responses” and “significantly altered user behavior,” while Limitations §6.6 correctly flags novelty and cross-sectional design. Align the Results and Discussion wording with the Limitations: either report additional controls that isolate asset presence, or restate findings as associations under the current between-cohort design rather than as effects of generated assets per se.","section":null},{"comment":"§5.2 / Table 2–3: The no-IKEA / no Group×Modality conclusion is important and currently rests on non-significant interactions. Given the ordinal CLMM and multiple scenes, please report effect sizes or equivalence-style bounds (or power considerations) for Group and Group×Modality so readers can judge whether absence of authorship leniency is well supported versus underpowered, not only that p-values exceed .05.","section":null}],"minor_comments":[{"comment":"§4.4 heading: typo “Measuremetns” → “Measurements”.","section":null},{"comment":"§5.3: “thaDesigner group’s” → “the Designer group’s”.","section":null},{"comment":"§3.1.2 / abstract vs body: model naming is inconsistent (DeepSeek V3 vs V3.2); standardize throughout.","section":null},{"comment":"Figure 5–8: ensure all panels have readable axis labels and that color encodings for Design vs Validation are consistent across figures; some density plots are hard to parse in grayscale.","section":null},{"comment":"§5.5: generation-time SDs are large for 3D (194±124 s); briefly note how timeouts/retries were handled in the user-facing protocol so latency claims are interpretable.","section":null},{"comment":"Related Work: a short explicit comparison table (or paragraph) against VRCopilot, ImaginateAR, and Dreamcrafter on stages (conversation / 2D gate / deploy / evaluation) would sharpen the claimed gap.","section":null},{"comment":"Ethics/data: state whether conversation transcripts and assets will be released with the open-source pipeline, and any de-identification steps.","section":null}],"recommendation":"major_revision","confidential_remarks":"The between-cohort confound for the primary with/without-asset claims is the main load-bearing issue; the system contribution, modality gap, and conversation analyses look publishable after revision. Fit for a serious HCI/XR venue is good if causal language is disciplined. No concerns about circularity or fabricated theory; this is a standard empirical systems paper."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing to know: this is a competent HCI systems paper that ships a real staged pipeline (chat → confirmed 2D concept → image-to-3D → Unity deploy) and backs it with N=120, validated affective scenes, and careful mixed models. The distinctive empirical pieces are the mandatory 2D gate, the clear 2D>3D satisfaction drop, and the absence of Group×Modality authorship leniency—no IKEA effect on the quality gap. Conversation pacing also tracks scene affect, which is a nice secondary result.\n\nWhat it does well: implementation detail is unusually concrete (FLUX/Hunyuan configs, latency benchmarks, async queues). Stats match the data types—CLMM for Likert, ART ANOVA for SAM, Gamma GLMMs for dwell and paths, FDR post-hocs, covariates. The no-IKEA finding is cleanly reported and useful for co-creation work. Scene-congruent prompt analysis and designer feedback are readable and honest about the 2D→3D bottleneck.\n\nSoft spots, in proportion: the stress-test is right. The primary with/without-asset claims (roughly 1.8–2.3× dwell; SAM shifts in low-valence scenes) compare Designers’ pre-design asset-free sessions to a separate Validation cohort that only ever saw populated scenes. Covariates do not fix unmeasured cohort differences, order, or novelty of any new objects. Designers’ own post-design data exist but are de-emphasized for the headline models, so causal isolation of “asset presence” is weaker than the abstract implies. There is also no freeform-prompt baseline, so the staged agentic claim is only partly isolated. Open-source is asserted without a clear artifact pointer in the text. Sample is university convenience. None of that sinks the paper; it just means the democratization narrative outruns the tightest contrast.\n\nWho it’s for: XR authoring, creativity-support, and affective VR people who want a concrete pipeline plus behavioral telemetry. Math and citations look solid for the subfield; self-cites to their stimulus set are appropriate. I’d send it to peer review. Engage if you work on AI-XR tools or affect in VR; treat the engagement/affect multipliers as associated, not cleanly causal, until a within-subject or better-controlled follow-up.","headline":"Useful staged VR co-design system with a large study and clean quality/IKEA findings, but the headline engagement and affect claims rest on a between-cohort contrast that confounds asset presence with cohort and novelty.","tokens_in":41484,"tokens_out":591,"would_cite":true,"duration_ms":10471,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A staged conversation-to-2D-to-3D pipeline lets non-experts co-create VR assets that raise scene engagement and shift affect.","keywords":["human-AI co-design","agentic system","3D content generation","virtual reality","affective computing","creativity support tools","generative AI","image-to-3D"],"falsifier":"A within-subjects or fully crossed design in which the same participants experience both asset-free and asset-populated versions of the same scenes (or a pure between-subjects arm that also measures validators on asset-free baselines) and still recovers the reported dwell-time doubling and SAM shifts would support the claim; disappearance of those effects would falsify the attribution to co-designed assets.","tokens_in":41384,"feed_emoji":"🥽","tokens_out":728,"duration_ms":6042,"temperature":0.7,"pith_summary":"CoGen3D argues that immersive 3D authoring is bottlenecked less by generative model quality than by the interaction model: unconstrained command prompting leaves non-experts unable to articulate intent or validate designs before expensive 3D generation. The system therefore inserts a deliberate three-stage agentic pipeline—LLM-guided conversational elicitation of object semantics, style, and scene fit; mandatory confirmation of a 2D concept image; then image-to-3D conversion and direct Unity VR deployment. In a between-subjects study with 120 participants across six previously validated affective VR scenes, assets co-designed this way were associated with substantially longer scene dwell times and with scene-dependent shifts in valence, arousal, and dominance, most clearly in neutral or negative environments. Both co-designers and independent validators preferred the intermediate concept images over the final 3D meshes and showed no authorship-driven leniency toward the quality drop. The paper’s larger claim is that staging human judgment before 3D rendering can democratize VR content creation and reframe it as collaborative spatial design rather than technical modeling.","feed_headline":"Staged AI co-design puts non-experts’ 3D assets into VR","feed_subtitle":"Conversation then 2D confirmation yields assets that raise engagement and shift affect across six scenes","key_machinery":"The three-stage CoGen3D pipeline: agentic conversational intent elicitation, mandatory 2D concept-image confirmation gate, and deferred image-to-3D generation with direct VR deployment. The 2D gate is the load-bearing design choice that fronts human judgment before high-latency 3D synthesis.","core_discovery":"A staged agentic pipeline that elicits intent conversationally, requires user confirmation of a 2D concept image, then generates and deploys a 3D asset into VR enables non-experts to produce scene-congruent props whose presence is associated with higher engagement and shifted affective responses, while both co-designers and independent raters prefer concept images over final meshes with no IKEA-style ownership leniency.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["CoGen3D co-designs VR assets via chat and concept image confirm","Non-experts co-create deployable 3D props that raise VR engagement","Conversational AI pipeline puts user-confirmed assets into immersive scenes","Staged intent-to-3D pipeline shifts affect without ownership bias","Users prefer concept images over final meshes in AI co-design for VR"],"cache_read_input_tokens":32896,"weakest_assumption_plain":"The main affective and engagement contrasts treat designers’ asset-free pre-design ratings as a clean baseline for validators who only ever saw scenes already populated by other people’s assets, so any cohort, order, or novelty differences could be misread as effects of the assets themselves.","fun_headline_variants_meta":{"raw":{"variants":["CoGen3D co-designs VR assets via chat and concept image confirm","Non-experts co-create deployable 3D props that raise VR engagement","Conversational AI pipeline puts user-confirmed assets into immersive scenes","Staged intent-to-3D pipeline shifts affect without ownership bias","Users prefer concept images over final meshes in AI co-design for VR"]},"model":"grok-4.5","effort":"low","cost_usd":0.004888,"raw_usage":{"total_tokens":1380,"prompt_tokens":800,"num_sources_used":0,"completion_tokens":81,"cost_in_usd_ticks":48880000,"prompt_tokens_details":{"text_tokens":800,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":499,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":800,"tokens_out":81,"duration_ms":4547,"temperature":1.0,"reasoning_tokens":499,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T00:19:26.428668+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"A within-subjects or fully crossed design in which the same participants experience both asset-free and asset-populated versions of the same scenes (or a pure between-subjects arm that also measures validators on asset-free baselines) and still recovers the reported dwell-time doubling and SAM shifts would support the claim; disappearance of those effects would falsify the attribution to co-designed assets.","supporting_citations":[],"review_version":1}