{"id":"acafd2a0-1ca0-40fd-bcd1-27f33a8207a7","arxiv_id":"2507.19058","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ScenePainter introduces a SceneConceptGraph that encodes multi-level scene concepts and relations, and aligns an outpainting model with them to reduce semantic drift in perpetual 3D scene generation.","lead":"ScenePainter generates long 3D view sequences from a single image while keeping the scene's objects and style consistent, using a hierarchical scene concept graph that aligns the generation model with the scene's structure. It targets the semantic drift that makes existing perpetual scene generators wander into unrelated environments.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Long-range semantic drift is never directly measured; the claim rests on a first-view prior that §3.3 updates by only one edge per step, so drift prevention is assumed rather than demonstrated.","rationale":"The strongest claim is empirical: semantic drift is overcome. The paper's only quantitative evidence for this is Table 1 and Table 3 (single-image customization fidelity) and Table 2 (user preference). Neither directly measures drift over an autoregressive view sequence, where the outpainter's outputs become its own inputs. The refinement in §3.3 updates only one relation edge and is described as user-specified, so in the default generation mode the entire sequence is guided by a prior extracted from the first frame. That is the load-bearing assumption identified by the reader, and I agree it is the least secure point. The stress-test pass did not find an internal contradiction in the method; the concern is that the central empirical conclusion is under-supported. A direct long-horizon drift evaluation would settle it. I also note the BLD-based conversion of the customized text-to-image model into an outpainter is not separately verified, and the Lprior training samples are generated via outpainting, but these are secondary to the missing drift measurement. Therefore the reader's conditional verdict remains appropriate.","tokens_in":10693,"tokens_out":5246,"duration_ms":56760,"concrete_test":"Re-run ScenePainter, WonderJourney, and SceneScape on the same 30 scenes with fixed camera trajectories of at least 50 frames, both with and without user-specified refinements. Compute a drift metric on the generated sequences: (a) pairwise CLIP and DINO similarity between the first view and every 5th generated view, and (b) CLIP text-image alignment between each view and the initial scene caption. Report per-scene distributions with bootstrap confidence intervals; if ScenePainter's first-to-last similarity is not significantly higher than WonderJourney's, the semantic-drift claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ScenePainter overcomes semantic drift by aligning the outpainter with the SceneConceptGraph. That claim depends on two linked conditions: (i) the prior learned from the first view remains a sufficient guide for every later unproject/render/outpaint step, and (ii) the test-time refinement in §3.3 keeps the graph aligned as new content appears. §3.3 explicitly restricts refinement to one relation edge between the first-level concept v0 and a user-specified second-level concept, for efficiency. This means the accumulated drift is supposed to be prevented by a fixed initialization plus at most one updated edge per step. The paper reports no direct measurement of drift over long sequences: Table 1 and the ablation in Table 3 measure single-image customization fidelity (DINO/CLIP-I on one generated image), and Table 2 is a binary user preference study with no reported number of users, no significance test, and no tie/error analysis. A method could score high on single-image fidelity and still drift by frame 50, because the outpainter is generating its own inputs after the first view. The concern is therefore not that the method is internally inconsistent, but that the load-bearing mechanism—persistent scene prior under repeated outpainting—has no quantitative support in the paper. The strongest version of the claim ('eliminates semantic drift') is broader than the evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ScenePainter, a two-stage framework for perpetual 3D scene generation from a single image. In the first stage, it constructs a hierarchical SceneConceptGraph with three levels of scene concepts and three relation types, and optimizes dedicated text embeddings and a diffusion model using masked reconstruction, prior preservation, and cross-attention losses. In the second stage, the outpainter is initialized with the optimized model and the graph is refined at test time by updating one edge at a time, enabling user-driven add, remove, or modify operations on scene concepts. The paper reports qualitative comparisons, DINO/CLIP-I fidelity scores, a user study, and ablations, and claims to overcome semantic drift in long-range 3D view sequences.","tokens_in":10981,"tokens_out":3469,"duration_ms":32564,"significance":"The paper addresses a real problem in perpetual 3D scene generation: semantic drift in autoregressive outpainting. The proposed SceneConceptGraph is an interpretable and structured scene prior, and the two-stage pipeline of construction and test-time refinement is a plausible way to condition an outpainter. The ablations indicate that the loss components and graph levels contribute to customization quality. If the long-range consistency claim were properly validated, this would be a useful contribution to 3D content generation and editing. The paper also offers a meaningful extension of customization methods from single-object or multi-object settings to scene-level concepts and relations.","major_comments":[{"comment":"The user study in Table 2 reports preference percentages (e.g., 89.3% quality over WonderJourney) without the number of participants, confidence intervals, or significance tests. Because the central claim about consistency and diversity in long-range view sequences rests on this table, the evidence is statistically unsubstantiated. Please report sample size, per-scene variance, and a significance test (e.g., bootstrap or sign test), and clarify whether ties were allowed.","section":"Section 4.3, Table 2"},{"comment":"The DINO/CLIP-I metrics in Table 1 are computed on a single generated image against the input image. Since the SceneConceptGraph and the fine-tuned model are optimized with a reconstruction loss on that same input (Eq. 5), high fidelity is partly by construction. More importantly, these scores do not measure semantic consistency across the iterated outpainting sequence of Section 3.1. The paper's headline claim of overcoming semantic drift needs a direct measurement of drift over the generated sequence, e.g., CLIP/DINO similarity between frame 1 and frame T, or per-frame semantic alignment with the SceneConceptGraph.","section":"Section 4.3, Table 1 and Section 3.2"},{"comment":"The refinement stage updates only one relation edge between the first-level concept v0 and a user-specified second-level concept per step. The paper assumes this sparse update plus the fixed initialization prevents accumulated drift, but no experiment validates that assumption. Please provide either a long-sequence study measuring drift over many outpainting steps (e.g., 50-100 frames) or an analysis of how the graph and model are updated each step and whether errors compound.","section":"Section 3.3"}],"minor_comments":[{"comment":"The values of lambda_prior and lambda_attn in Eq. (5) are not reported; please specify them in the implementation details.","section":"Section 4.5"},{"comment":"The colored dots and regions in Figure 2 are mentioned in the caption but are not clearly labeled in the figure; please add a legend for the concept nodes and relation edges.","section":"Figure 2"},{"comment":"The conclusion states that the method 'eliminates the semantic drift problem,' which is stronger than the evidence; I suggest wording such as 'substantially reduces observed semantic drift in the tested scenarios.'","section":"Section 5"},{"comment":"The qualitative comparison does not state how many scenes are shown in the figures or whether the selected examples are representative; please add a brief description of the selection protocol.","section":"Section 4.2"},{"comment":"The discussion of video generation methods is brief; consider adding citations for recent long-video models to clarify the distinction between viewpoint-centric 3D generation and object-centric video generation.","section":"Section 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely to be a good fit for a computer vision journal if the evaluation gap is addressed. The central claim about long-range drift is not directly measured, and the user study lacks statistical grounding. I would not reject, but the revision must add quantitative long-sequence consistency evaluation and proper user study statistics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is scene-level customization: prior work like Break-A-Scene and Custom Diffusion does object-level or multi-object customization, and perpetual scene generators like WonderJourney don't anchor the outpainter to a persistent scene description. ScenePainter builds a hierarchical SceneConceptGraph with three levels of concepts and typed relation edges, optimizes textual handles plus model weights per concept-relation pair, and uses that to initialize the outpainter. That is a sensible and fairly clean design, and the ablations in Table 3 show each loss term and each graph component moves the numbers, which is more than many papers in this area do. The qualitative figures also look plausible: the method clearly preserves more of the first view's identity than the baselines, and the user-controlled edits in Figure 6 demonstrate a genuinely useful capability.\n\nNow the soft spots, in proportion. The biggest one is the central claim: 'overcomes semantic drift.' The paper never measures drift over a long sequence. Table 1 is single-image customization fidelity against the input image; Table 2 is a binary user preference study with no sample size, no confidence intervals, no significance test, and no error analysis. The stress-test note is on target here. Section 3.3 explicitly refines only one relation edge per step (between v0 and a user-specified second-level concept) for efficiency. So the mechanism that is supposed to prevent drift over tens of frames is a fixed initialization plus at most one updated edge per step, with no direct evidence that this suffices. A method could score well on one-step fidelity and still drift by frame 50, because after the first view the outpainter is generating its own inputs. This is not an internal contradiction; it's a missing measurement.\n\nSecond, the automatic metrics CLIP-I and DINO are partly circular in the customization evaluation because the model is trained with a reconstruction loss against the same image used for the similarity computation. That's a known limitation of this evaluation style, but the paper doesn't acknowledge it.\n\nThird, no code or data are released. The user study numbers (89.3%, etc.) look strong, but without methodological details they're hard to weigh.\n\nOn the citation pattern: the related work is fair and points to the right prior art; no red flags there. The paper is clearly written, and the method is reasonable for what it targets.\n\nWho is this for? Researchers working on perpetual view generation, scene customization, or long-horizon outpainting. They would use the SceneConceptGraph idea and the two-stage alignment as a starting point, not as a proven recipe.\n\nMy view: worth engaging seriously, but the paper needs a direct long-sequence drift evaluation (e.g., CLIP/DINO similarity between first and later frames, or a user study with proper statistics and a per-frame drift curve) before the headline claim is supportable. I would accept it for peer review, but with the expectation of major revision.","headline":"A credible scene-level customization method with a real novelty gap vs. prior work, but the paper's central claim about eliminating long-range semantic drift is only weakly supported because drift is never directly measured.","tokens_in":11485,"tokens_out":723,"would_cite":true,"duration_ms":9362,"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":"ScenePainter claims that a hierarchical scene concept graph, learned from the first view and refined during generation, prevents semantic drift in perpetual 3D scene expansion.","keywords":["perpetual 3D scene generation","semantic drift","scene concept graph","scene-level customization","image outpainting","diffusion models","user-guided scene editing","textual inversion"],"falsifier":"Generate a long view sequence (for example, 30 or more frames) from a single image with ScenePainter, and compare a run with the SceneConceptGraph and refinement disabled against a full run using CLIP-I and DINO between the first and last views; if the ablated run matches the full method, the graph is not what prevents semantic drift. Alternatively, if editing one concept causes unrelated concepts or the global style to change in later frames, the one-edge-per-step refinement claim is violated.","tokens_in":10544,"feed_emoji":"🎨","tokens_out":5767,"duration_ms":53200,"temperature":0.7,"pith_summary":"ScenePainter addresses the semantic drift that plagues perpetual 3D scene generation, where iterated outpainting gradually turns a coherent scene into a different one. The paper argues that the fix is to give the outpainting model a scene-specific prior extracted from the first view, encoded as a hierarchical graph of concepts and their relations. It constructs this SceneConceptGraph with optimized textual embeddings and model weights, then uses it to initialize and continuously refine the outpainter. If the claim is right, long-range view sequences remain semantically consistent while users can still add, remove, or alter objects, which would make single-image 3D exploration and long video synthesis more practical.","feed_headline":"A concept graph stops 3D scene expansion from drifting","feed_subtitle":"Aligning the outpainter with learned scene relations keeps the view consistent while edits stay possible.","key_machinery":"The central object is the SceneConceptGraph, a three-layer hierarchical graph with a root node for global environment and style, second-layer nodes for category regions, and third-layer nodes for individual objects, connected by relation edges that encode spatial layout and semantic connections. It carries the argument by acting as a persistent scene memory: concept-relation pairs are converted into optimized textual embeddings and model weights, so the graph both describes the scene and steers the outpainter. During test-time refinement, only one relation edge is updated per step, keeping the process fast while letting the scene evolve through user text commands.","core_discovery":"Starting from one image, ScenePainter builds a three-level SceneConceptGraph whose nodes are the overall environment and style, category regions, and individual objects, and whose edges encode spatial and semantic relations. Each concept-relation pair is turned into dedicated textual embeddings, optimized together with a diffusion model by masked reconstruction, scene-prior preservation, and cross-attention losses, so the model learns what the scene is both as a whole and as a set of related parts. This customized model becomes the outpainter's initialization via Blended Latent Diffusion, and during view generation the graph is refined one edge at a time, updating text embeddings and model weights so newly added, changed, or removed concepts stay consistent with the original environment. The paper claims that this alignment removes the accumulated deviation that causes semantic drift and produces view sequences that are both consistent and diverse, supported by fidelity metrics and user preference over previous methods.","pith_inferences":["Beyond the paper, the SceneConceptGraph could be reused as a semantic memory for interactive world building, letting an agent accumulate what it knows about a scene and query or update it as new views arrive.","One could test the graph's generality by extending the relation types beyond spatial layout and style, for example to temporal or causal relations, which would let the same framework sustain story-like scene evolution.","The error-compounding assumption suggests a stress test: if refinement updates only one edge per frame, scenes with many simultaneous changes may eventually outrun the graph and drift; weighting or scheduling multiple edge updates could be a natural extension.","Because the method optimizes embeddings and weights at test time, its quick refinement step makes it suitable for offline keyframe generation; a distilled or faster variant would be needed for real-time navigation."],"forward_implications":["Long view sequences generated from a single image will keep a stable overall style, object set, and spatial layout instead of drifting into a different scene.","Users can edit scenes while they are generated using short text handles: add a bookshelf, remove a chair, recolor a blanket, or recombine concept relations.","Combining the graph with a 3D representation pipeline yields 3D structures that stay geometrically plausible because semantic consistency prevents the objects and surroundings from distorting.","The scene-specific prior can be transferred to any outpainting module that can be initialized from the customized text-to-image model, not just the one used in the paper.","The approach offers a practical keyframing path from one picture to long, editable 3D video sequences."],"supporting_citations":[{"why":"supplies the multi-concept extraction and phase-combined customization strategy that ScenePainter extends to scene-level relations","marker":"[1]"},{"why":"provides the mechanism that converts the customized text-to-image model into an outpainting model","marker":"[2]"},{"why":"baseline that produces consistent but monotonous scenes, framing the consistency-diversity trade-off","marker":"[10]"},{"why":"provides the embedding-optimization phase for concept and relation handles","marker":"[11]"},{"why":"is the base latent diffusion model that ScenePainter customizes and converts for outpainting","marker":"[24]"},{"why":"provides the whole-model fine-tuning phase that preserves scene identity","marker":"[25]"},{"why":"supplies the unproject-render-outpaint pipeline whose semantic drift is the paper's target problem","marker":"[36]"}],"fun_headline_variants":["Graph-guided outpainting keeps 3D scenes steady","Aligning concepts halts 3D scene drift","SceneConceptGraph anchors long-range 3D views","Relation-aware generation stops semantic drift"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the scene-specific prior learned from the first view remains a valid guide for every later outpainting step, so errors in the initial SceneConceptGraph do not compound as the camera moves.","fun_headline_variants_meta":{"raw":{"variants":["Graph-guided outpainting keeps 3D scenes steady","Aligning concepts halts 3D scene drift","SceneConceptGraph anchors long-range 3D views","Relation-aware generation stops semantic drift"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000581,"raw_usage":{"total_tokens":2712,"prompt_tokens":899,"completion_tokens":1813,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":515,"completion_tokens_details":{"reasoning_tokens":1753}},"tokens_in":515,"tokens_out":1813,"duration_ms":13437,"temperature":1.0,"reasoning_tokens":1753,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:01:00.475744+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a long view sequence (for example, 30 or more frames) from a single image with ScenePainter, and compare a run with the SceneConceptGraph and refinement disabled against a full run using CLIP-I and DINO between the first and last views; if the ablated run matches the full method, the graph is not what prevents semantic drift. Alternatively, if editing one concept causes unrelated concepts or the global style to change in later frames, the one-edge-per-step refinement claim is violated.","supporting_citations":[{"cited_title":"Break-a-scene: Extracting multi- ple concepts from a single image","cited_arxiv_id":null,"evidence_quote":"supplies the multi-concept extraction and phase-combined customization strategy that ScenePainter extends to scene-level relations"},{"cited_title":"Blended latent diffusion","cited_arxiv_id":null,"evidence_quote":"provides the mechanism that converts the customized text-to-image model into an outpainting model"},{"cited_title":"Scenescape: Text-driven consistent scene generation","cited_arxiv_id":null,"evidence_quote":"baseline that produces consistent but monotonous scenes, framing the consistency-diversity trade-off"},{"cited_title":"High-resolution image synthesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"is the base latent diffusion model that ScenePainter customizes and converts for outpainting"},{"cited_title":"Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation","cited_arxiv_id":null,"evidence_quote":"provides the whole-model fine-tuning phase that preserves scene identity"},{"cited_title":"Wonderjourney: Going from anywhere to everywhere","cited_arxiv_id":null,"evidence_quote":"supplies the unproject-render-outpaint pipeline whose semantic drift is the paper's target problem"}],"review_version":2}