{"id":"115cc91f-3a1c-437a-98f1-36e5d85274f0","arxiv_id":"2501.13969","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"InsTex generates style-consistent textures for indoor 3D scenes using a coarse-to-fine diffusion pipeline with global image guidance, reporting faster and higher-scoring results than four baselines.","lead":"The paper presents InsTex, a two-stage pipeline that textures 3D indoor scene meshes by first generating coarse multi-view textures with a pretrained depth-to-image diffusion model and then refining them in UV space for style consistency. It claims state-of-the-art quality and 2-hour processing, versus up to 48 hours for existing methods, which matters for VR/AR, gaming, and interior design workflows.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Scene-level style consistency is asserted but not tested: the global style image / post-processing component has no ablation, and the CLIP metric shares representation with the IP-Adapter conditioning.","rationale":"The reader's weakest_assumption identifies the same load-bearing point: per-object texturing conditioned on one global style image has not been isolated by an ablation, so the central scene-consistency claim is unsupported. I agree with that reading. My stress-test adds a second, reinforcing observation: the headline CLIP metric is computed with a model from the same CLIP embedding family used by both the Stable Diffusion text encoder and the IP-Adapter image encoder, making the metric partially circular for the consistency claim. However, this circularity is not decisive by itself because the baselines also rely on Stable Diffusion and CLIP conditioning. The decisive missing evidence is an ablation of the global-image conditioning and/or the scene-level post-processing, together with a direct inter-object consistency metric. The paper is a plausible engineering contribution, and the conditional verdict remains appropriate: accept only after the missing ablation and independent consistency evaluation are provided. No code or data release further supports the need for this verification.","tokens_in":7843,"tokens_out":4081,"duration_ms":45757,"concrete_test":"Run the identical pipeline on the same 10 3D-FRONT scenes with the IP-Adapter global style image replaced by a mismatched style image (e.g., a modern dining-room image for a 'baroque bedroom' prompt) and with the style image removed entirely; render the 20 evaluation views per scene and compute pairwise CLIP image-image similarity between all object pairs in each scene, in addition to the Table I CLIP score. If the pairwise consistency score and CLIP score do not change significantly when the global style image is removed or mismatched, the scene-consistency claim is not supported by the proposed mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline 'style-consistent' claim rests on conditioning each independently textured object on a single global style image via IP-Adapter (Sections II-B and II-D), followed by a scene-level post-processing diffusion step. No ablation in Section III-F isolates either component: the only ablations shown are coarse vs. refined output and diffusion step count. Consequently, there is no demonstrated causal link between the global style image and scene-level coherence. Moreover, Table I's CLIP score measures text-image alignment in the same CLIP embedding family used by the IP-Adapter image encoder and the Stable Diffusion text encoder, so it is a weak, partially circular proxy for cross-object style consistency; the user study asks only for visual quality and prompt fidelity, not for inter-object consistency. For the SOTA claim to hold, the global-image conditioning or the post-processing must be shown to be doing the coherence work; currently it is an unsupported assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces InsTex, a two-stage pipeline for generating stylized textures for indoor scene meshes. The method decomposes a scene into individual objects, textures each object in canonical space using a depth-to-image Stable Diffusion/ControlNet prior conditioned on both a text prompt and a globally generated style image (via IP-Adapter), refines the resulting UV maps with a position-map-conditioned inpainting model, and finally recomposes the scene with an additional diffusion-based post-processing pass. The authors report qualitative results on 3D-FRONT scenes, quantitative CLIP and Inception Scores, a user study, and an ablation comparing coarse versus refined output and diffusion step counts, along with a timing comparison showing roughly 2 hours of GPU time versus 48 hours for SceneTex.","tokens_in":8015,"tokens_out":4051,"duration_ms":43126,"significance":"If substantiated, the efficiency claim alone is a meaningful practical contribution: reducing indoor-scene texturing from about two days to two hours while supporting both text and image prompts would make the method attractive for real applications. The core pipeline is clearly described and combines existing components in a sensible way. However, the paper as written does not provide enough evidence for its central 'style-consistent' and 'state-of-the-art' claims: the quantitative evaluation is small-scale and statistically unsupported, the primary metric shares a representation with the generation backbone, and the component responsible for scene-level coherence is not isolated by an ablation. The significance is therefore conditional; the manuscript needs substantial strengthening of the evaluation before the claims can be accepted.","major_comments":[{"comment":"The statement that InsTex 'significantly outperforms all baselines' is not supported by the reported evidence. The scores are aggregate means over 10 scenes and 2 prompts per scene, with no standard deviations, confidence intervals, per-scene breakdown, or significance tests. A difference of about 2 CLIP points over such a small sample cannot be assessed without this information. Please report per-scene/per-prompt results and perform appropriate statistical tests.","section":"Section III-E, Table I"},{"comment":"The central style-consistency claim rests on two components: the global-style-image conditioning used for each object and the scene-level post-processing diffusion step. The ablation studies in Section III-F only compare coarse versus refined texture output and vary the diffusion step count; they never remove or weaken the global style image or the post-processing step. Consequently, there is no demonstrated causal link between these components and the claimed scene-level coherence. Add ablations that isolate the global image guidance and the post-processing, and evaluate scene-level consistency directly rather than only per-object quality.","section":"Sections II-B to II-D and III-F"},{"comment":"The CLIP score used in Table I is computed with the same CLIP text encoder used to condition the Stable Diffusion backbone, and the IP-Adapter image encoder used for the global style image is also drawn from the CLIP embedding family. A higher CLIP score can therefore reflect alignment between the generator's conditioning space and the evaluator, rather than independent visual quality or cross-object style consistency. Supplement the evaluation with metrics that do not share this representation, or with human judgments that specifically target inter-object style coherence.","section":"Section III-C and III-E"},{"comment":"The user study is described as having 100 participants evaluating five randomly selected meshes, with no details on participant selection, no per-item variance, and no inter-rater agreement measure. The reported mean scores without statistical analysis cannot support the claim that InsTex 'significantly outperforms' the baselines in visual quality or prompt fidelity. Please report the distribution of ratings, a statistical comparison, and the stimulus-selection procedure.","section":"Section III-E, User Study"}],"minor_comments":[{"comment":"The caption says the method starts with a 'baroque bedroom' prompt, but the example prompt shown and described in the text is a 'baroque living room'; please reconcile this inconsistency.","section":"Figure 1 caption"},{"comment":"The name 'SceneText' appears to be a typo for 'SceneTex' in the description of the dataset settings.","section":"Section III-B"},{"comment":"The caption uses 'raw' where 'row' is intended ('the first raw shows...').","section":"Section III-F, Figure 4 caption"},{"comment":"The GPU is listed as 'NIVIDIA V100'; this should be 'NVIDIA V100'.","section":"Section III-A, Table II"},{"comment":"Table I cites CLIP as reference [26], while the text in Section III-C cites it as [23]; the duplicate 3D-FRONT entries [22] and [25] should be merged and all citations renumbered consistently.","section":"Reference numbering"},{"comment":"The sentence 'The main objective of this stage is to obtained the 3D location coordinates' contains a grammatical error ('to obtained' should be 'to obtain').","section":"Section II-A"},{"comment":"The phrase 'We conduct the some ablation studies' is grammatically awkward and should be rewritten.","section":"Section III-F"}],"recommendation":"major_revision","confidential_remarks":"The paper is not ready for acceptance in its current form. The proposed pipeline is reasonable and the efficiency advantage is potentially valuable, but the experimental section does not currently support the headline claims. I would ask the authors to add proper statistical reporting, an ablation isolating the global-style and post-processing components, and a less circular or supplementary evaluation before resubmission. The issues are substantive but repairable within the scope of the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi [name],\n\nThe thing to know about InsTex is that it is a solid engineering contribution with a headline speed improvement, but its state-of-the-art claim is held up by weak evidence. The pipeline combines known pieces—SD1.5, ControlNet depth-to-image, IP-Adapter, UV-space inpainting—in a new way that makes indoor scene texturing practical: roughly 2 hours per room versus 48 for SceneTex. That speedup alone is worth reporting, and the dynamic view partitioning strategy for coarse-stage inpainting is a sensible fix for projection artifacts.\n\nWhat's genuinely new is the specific system: object-level decomposition, per-object texturing with a global style image as a shared condition, and a UV-space refinement stage with a position map as an auxiliary ControlNet input. The paper also supports both text and image prompts, which is a useful feature. I haven't seen an exact prior combination of these pieces, so the engineering novelty is real, if incremental.\n\nThe soft spots are all in the evaluation. The quantitative evidence is 10 scenes x 2 prompts, with CLIP and Inception Scores and no error bars or significance tests. The user study collected 100 ratings but doesn't report how participants were selected or whether they agreed. More importantly, the headline CLIP score is computed with the same CLIP model that conditions the SD1.5 backbone via the text encoder and IP-Adapter, so part of the advantage is generator-evaluator alignment, not independent fidelity. The ablation section only shows coarse vs. refined output and diffusion step count; it never isolates the global style image or the final scene-level post-processing, so the paper's central claim of style consistency is asserted rather than demonstrated. That is a real gap, because per-object consistency does not logically imply scene-level coherence. The stress-test note is right about this.\n\nThat said, there's no internal contradiction or fatal flaw. The pipeline is coherent, the writing is clear, and the qualitative results look plausible. The missing ablations and metric details are fixable in revision. I'd send this to a serious referee. The right outcome is likely conditional acceptance after the evaluation is tightened and the style-consistency claim is either supported or scaled back.\n\nFor your purposes: worth reading if you work on 3D texturing or on evaluation pitfalls in generative 3D. I wouldn't cite it until code and data are out, but it deserves referee time.","headline":"Solid engineering with a real speed gain, but the SOTA and style-consistency claims outrun the evidence.","tokens_in":8543,"tokens_out":2338,"would_cite":false,"duration_ms":22055,"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":"InsTex textures entire indoor scenes from one prompt in about two GPU hours while keeping style consistent across objects.","keywords":["3D texture synthesis","indoor scenes","style consistency","diffusion models","depth-to-image","coarse-to-fine texture generation","scene decomposition and recomposition","prompt-conditioned generation"],"falsifier":"Run InsTex on the same 3D-FRONT scene twice with the same text prompt but two very different generated style images, then have users rate which output is more style-consistent; if the global style image is doing the claimed work, the outputs should differ clearly. A finer check is to measure CLIP-based style distance between neighboring objects across the recomposed scene and compare it with the same distance computed on the same objects rendered before recomposition, which would show whether the post-processing pass or the global style image is responsible for cohesion.","tokens_in":7662,"feed_emoji":"🎨","tokens_out":8814,"duration_ms":81203,"temperature":0.7,"pith_summary":"InsTex is a two-stage system for turning an untextured 3D indoor scene mesh into a fully stylized room from a text prompt such as 'a baroque living room' or from a reference image. The paper claims that by decomposing the room into objects, texturing each object separately with a depth-aware diffusion model under one shared style image, and then recomposing the room, it achieves both high visual quality and scene-wide style consistency. The authors report that this approach beats existing indoor texturing baselines on CLIP score, Inception Score, and user ratings, and that a typical living room takes about 2 GPU hours to texture instead of about 48 hours for the previous per-scene optimization method. If the claim holds, practical styling of indoor scenes for interior design, games, and AR/VR could move from days of compute to a much faster, prompt-driven workflow.","feed_headline":"Indoor scene texturing cut from 48 GPU hours to 2","feed_subtitle":"Objects are textured under one shared style image, keeping the room consistent and beating prior prompt fidelity.","key_machinery":"The load-bearing mechanism is the global style image: a single image generated from the scene prompt that is injected as a condition (through an image-adapter) alongside a per-object text prompt, so each independently textured object inherits a common look. The second mechanism is the position map in UV space, an image-like map whose pixels store the 3D point coordinates of the mesh, used with a separately trained position-map encoder connected by zero-convolution layers; this lets the diffusion model inpaint across UV seams as if the texture were continuous on the surface. Together, the coarse stage's dynamic view-partitioning masks and the refinement stage's position-map inpainting carry most of the texture-quality work, while the global style image is what the paper credits for style consistency across objects.","core_discovery":"The paper's central discovery claim is that scene-level style consistency for indoor textures can be obtained without optimizing the whole scene at once: texture each object in isolation using a text prompt plus one generated global style image, then stitch the objects back together and run a final diffusion pass. The coarse stage renders the normalized object from multiple viewpoints and progressively back-projects depth-conditioned diffusion outputs into UV space, guided by dynamic masks that distinguish regions to generate, update, or keep. The refinement stage inpaints the UV texture map with a position-map encoder that encodes 3D adjacency, so seams between UV fragments are treated as continuous and self-occlusion artifacts are repaired. On the 3D-FRONT benchmark the method reports the highest numbers among compared systems: CLIP score 24.18, Inception Score 3.82, user visual quality 4.70, and prompt fidelity 4.82, with a processing time of about 2 hours for a typical living room.","pith_inferences":["If the consistency gain really comes from the single global style image, then replacing that image with a layout-aware rendering (for example, one that includes object positions and rough lighting) could push style coherence further; the paper does not test this variant.","The 2-hour timing appears to count diffusion texturing and post-processing, so an end-to-end wall-clock comparison should also include scene decomposition, UV parameterization, and style-image generation before claiming full interactivity.","The position-map inpainting mechanism is not tied to indoor scenes; the same idea could be applied to single objects with complex UV seams, and an ablation that removes the position-map encoder would quantify how much of the refinement gain it contributes."],"forward_implications":["The 48-hour per-scene optimization cost for indoor texturing drops to roughly 2 hours, so multiple stylization trials of the same room become practical.","Both text-only prompts and reference-image prompts are supported, letting a user style a room from a phrase or from a photo of a desired look.","Scene consistency is achieved at the object level, meaning the same texturing pipeline can be ported to other multi-object scenes without retraining on scene-level data.","Quality and prompt fidelity improve simultaneously over the compared baselines, not one at the expense of the other."],"supporting_citations":[{"why":"Provides the preprocessed scene meshes and the per-scene optimization baseline that InsTex is compared against.","marker":"[14]"},{"why":"The pre-trained depth-to-image diffusion backbone used for coarse-view generation and inpainting.","marker":"[9]"},{"why":"The denoising diffusion framework that underlies the progressive texture painting.","marker":"[8]"},{"why":"The 3D-FRONT dataset that supplies the test scenes and prompts for all experiments.","marker":"[22]"},{"why":"Baseline whose failure modes (over-saturation, hallucinated elements) set up the qualitative comparison.","marker":"[10]"},{"why":"Baseline that textures objects but lacks global style consistency, motivating the global style image.","marker":"[12]"},{"why":"Baseline multi-view diffusion method whose blurry, dim outputs anchor the quality comparison.","marker":"[13]"},{"why":"The CLIP model used for both the CLIP-score metric and text conditioning.","marker":"[26]"},{"why":"The Inception Score metric used to measure texture quality and diversity.","marker":"[24]"},{"why":"Rendering and texture projection utilities used to back-project views into UV space.","marker":"[21]"}],"fun_headline_variants":["Two-stage diffusion textures indoor scenes in 2 hours","Style-consistent indoor texturing from text or image prompts","Coarse-to-fine diffusion slashes indoor texturing time to 2h","InsTex: 2-hour indoor scene texturing with consistent style","Diffusion-based indoor texturing beats SOTA in prompt fidelity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire consistency argument rests on one generated style image carrying enough of the room's overall look that independently textured objects will still fit together seamlessly when recomposed; no ablation isolates this assumption.","fun_headline_variants_meta":{"raw":{"variants":["Two-stage diffusion textures indoor scenes in 2 hours","Style-consistent indoor texturing from text or image prompts","Coarse-to-fine diffusion slashes indoor texturing time to 2h","InsTex: 2-hour indoor scene texturing with consistent style","Diffusion-based indoor texturing beats SOTA in prompt fidelity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00108,"raw_usage":{"total_tokens":4502,"prompt_tokens":912,"completion_tokens":3590,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":3512}},"tokens_in":528,"tokens_out":3590,"duration_ms":25572,"temperature":1.0,"reasoning_tokens":3512,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:51:57.324094+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run InsTex on the same 3D-FRONT scene twice with the same text prompt but two very different generated style images, then have users rate which output is more style-consistent; if the global style image is doing the claimed work, the outputs should differ clearly. A finer check is to measure CLIP-based style distance between neighboring objects across the recomposed scene and compare it with the same distance computed on the same objects rendered before recomposition, which would show whether the post-processing pass or the global style image is responsible for cohesion.","supporting_citations":[{"cited_title":"Scenetex: High-quality texture synthesis for indoor scenes via diffusion priors,","cited_arxiv_id":null,"evidence_quote":"Provides the preprocessed scene meshes and the per-scene optimization baseline that InsTex is compared against."},{"cited_title":"Stable diffusion,","cited_arxiv_id":null,"evidence_quote":"The pre-trained depth-to-image diffusion backbone used for coarse-view generation and inpainting."},{"cited_title":"Learning transferable visual models from natural language supervision,","cited_arxiv_id":null,"evidence_quote":"The CLIP model used for both the CLIP-score metric and text conditioning."},{"cited_title":"Improved adversarial systems for 3d object generation and reconstruction,","cited_arxiv_id":null,"evidence_quote":"The Inception Score metric used to measure texture quality and diversity."}],"review_version":1}