{"id":"2fa587fb-9169-4284-9c76-4aa2388ed504","arxiv_id":"2603.27084","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A test-time adaptation method integrates a 3D-misaligned, AI-generated inserted view into a reconstructed 3D scene, preserving the captured region while extending it.","lead":"This paper introduces SceneExpander, a method that takes a 3D scene reconstructed from photos and expands it by adding a new AI-generated view that may not line up with the existing geometry. It keeps the original region stable while letting the new content blend in, so creators can grow a virtual scene step by step instead of regenerating it from scratch.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Expansion signal is self-referential: Eq. 8 distills toward an EMA of the student (Eq. 2), with no pixel-level loss on the inserted view, so new content is bounded by the backbone's prior; the method may only stabilize the captured region.","rationale":"The reader's weakest_assumption identifies exactly the self-referential nature of the expansion signal. My stress-test confirms this is the most load-bearing issue: the objective (Eqs. 5, 8, 10) contains no external supervision from the inserted view's pixels, so the adapted model's new content is bounded by what the backbone can already predict. This is not merely a reproducibility gap but a structural limitation that, if unaddressed, undermines the claim of expanding via integration. The paper's own ablations show positive contributions, and the human study provides some evidence, but these do not isolate the mechanism. Since the reader already rated the claim CONDITIONAL with medium correctness risk, my analysis does not change the verdict; it sharpens the specific test needed. I do not see grounds for REJECT because the method plausibly does improve preservation, and the synthetic test could vindicate the expansion mechanism.","tokens_in":13127,"tokens_out":7583,"duration_ms":83304,"concrete_test":"Construct a synthetic expansion benchmark from an ETH scene: render an inserted view with a known 3D object (e.g., a colored cube) placed at a known pose in the expanded region, then apply a random rigid misalignment to this view to simulate I_g. Run SceneExpander and the non-adapted WorldMirror on C∪{I_g}, and compute the Chamfer distance between the reconstructed point cloud in the expanded region and the ground-truth cube geometry. If the adapted model's Chamfer distance is not substantially lower than the backbone's, the self-distillation loop is not recovering the inserted object from the image; it only stabilizes the captured region, and the 'expansion' claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"SceneExpander's adaptation objective contains no term that compares the model's output to the inserted view's pixels or to any external geometric ground truth. Eq. 8 distills the student's prediction on I_g toward the EMA teacher's prediction f_{θ_G}(B)|g, where θ_G is updated by Eq. 2 as μθ_G+(1−μ)θ — an average of the student's own past parameters. The anchor loss (Eq. 5) likewise uses the frozen backbone's own predictions on C as references. Thus the only information channel from I_g to the reconstruction is the backbone's forward pass; the gradient signal only enforces consistency between the student and its own past. If, as the paper states, feed-forward models suppress new content when a misaligned view is added, the EMA teacher will share that suppression, and the student will be trained to preserve it. The ablation (Tab. 2) shows self-distillation changes metrics, but that may reflect artifact reduction or stabilization, not the introduction of content that the backbone did not already predict. The GPT-IS/H-IS scores are the sole check that the inserted content is actually realized, and they are produced by the same model family that generated the view (GPT-5.2), with human evaluation only partially mitigating this. The claim that SceneExpander 'incorporates the inserted view as an expansion constraint' therefore rests on an unverified premise: that the EMA self-distillation can transfer information from the inserted view's pixels into the geometry. Since no loss ties the output to those pixels, the premise is not guaranteed by the optimization. This is the load-bearing weak point of the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SceneExpander, a test-time adaptation method for integrating a text-guided, AI-generated inserted view into an existing multi-view 3D reconstruction. The method builds on a feed-forward reconstruction backbone (WorldMirror) and adapts it with two distillation losses: (i) anchor distillation on captured views against a frozen teacher to preserve the original scene, and (ii) inserted-view self-distillation on the generated view against an EMA teacher to 'accommodate' the new content. Additional components include subset sampling, geometry-perturbation augmentation, and stochastic restoration. Experiments on ETH scenes and online data report that SceneExpander improves preservation metrics (PSNR/SSIM/LPIPS) and expansion quality (GPT-IS, GPT-RQ, human scores) over baselines. The core claim is that this test-time adaptation robustly integrates a 3D-misaligned inserted view while keeping the captured region faithful.","tokens_in":13365,"tokens_out":4850,"duration_ms":51762,"significance":"The problem formulation—expanding a captured 3D scene by inserting a potentially inconsistent AI-generated view—is timely and practically relevant for iterative world building. The proposed anchor distillation is a sensible mechanism for preserving the known region, and the ablation study is systematic. However, the central mechanism for actually introducing the new content is under-specified and appears self-referential: the inserted-view self-distillation only distills the student toward its own EMA teacher, with no pixel-level or external geometric supervision from the inserted view. The evaluation is also weakened by using the same model family (GPT-5.2) both to generate the inserted views and to score the outputs, by manual curation of the inserted views, and by the absence of error bars or statistical tests on any headline number. If the reported improvements are real, the method is a useful step; but the current evidence does not conclusively show that the scene is actually expanded rather than merely stabilized.","major_comments":[{"comment":"The inserted-view self-distillation signal is self-referential. The soft labels in Eq. (7) are produced by the EMA teacher θ_G, which by Eq. (2) is an exponential moving average of the student's own parameters. Eq. (8) then distills the student toward those soft labels. No term compares the model's output to the pixels of I_g or to any external geometric ground truth. The only information channel from I_g to the reconstruction is the backbone's forward pass on a batch containing I_g. Since the paper itself states that feed-forward models suppress or distort content when a misaligned view is added, the EMA teacher may simply preserve that suppression. The self-distillation is therefore more accurately described as a temporal-consistency regularizer than as an 'expansion constraint.' This undermines the load-bearing claim that the inserted view contributes new scene content. The authors sh","section":"§3.3, Eqs. (2), (7), (8)"},{"comment":"The evaluation of expansion quality is circular. The same model, GPT-5.2, is used to generate the inserted views and to compute the GPT-IS and GPT-RQ scores. A VLM evaluator will likely favor outputs that resemble its own generation style, so the reported gains in GPT-IS/GPT-RQ may partly reflect self-consistency rather than true expansion quality. The human study (39 participants) is a partial mitigation, but it reports only averaged rank-derived scores without confidence intervals, inter-rater agreement, or significance tests. Tables 1 and 2 report single averages with no error bars, so the reader cannot judge whether the improvements (e.g., PSNR 21.206 vs 19.295) are statistically meaningful. Please report per-scene variance, confidence intervals, and significance tests, and consider using an independent evaluator (e.g., a different VLM or an absolute human scoring protocol).","section":"§4.1, App. C"},{"comment":"The manuscript states that the authors 'manually curate the generated candidate views and select those that best match the user instructions.' This introduces a selection bias: the evaluation is performed on views that are already likely to be visually plausible and integrable, which weakens the claim of robustness against 'free-form' and '3D-misaligned' insertions. The paper does not quantify the degree of misalignment (e.g., pose error, reprojection error, or occlusion). To support the central claim, the authors should either characterize the misalignment distribution of the curated views, or include results on uncurated generated candidates.","section":"§4.1, App. A"},{"comment":"The ablation shows that adding inserted-view self-distillation (p=0.5) changes GPT-IS from 3.56 to 3.70 and GPT-RQ from 3.24 to 3.08 compared to the previous variant, but no error bars are given and the differences are small. This does not convincingly demonstrate that self-distillation introduces new content; it could be an artifact of improved rendering stability. A more informative ablation would decouple the presence of I_g in the input batch from the self-distillation loss—for example, comparing a variant that receives I_g but has λ_G=0 (no self-distillation) against the full model—and then measuring whether the rendered expanded region actually contains the requested object. The current table conflates stabilization and expansion.","section":"Tab. 2, Ablations"}],"minor_comments":[{"comment":"Typo: 'OminiVGGT' should be 'OmniVGGT'.","section":"Tab. 1"},{"comment":"Typos: 'expanison' should be 'expansion'; 'onlince' should be 'online'.","section":"App. A"},{"comment":"The 'Restore' operation is not precisely defined. Please specify how the random subset of parameters is selected and how the rate r affects the restoration.","section":"Eq. (9)"},{"comment":"The model name 'GPT-5.2' is unusual; if this is a specific model version, please cite it. If it is anonymized, state that clearly.","section":"§4.1"},{"comment":"Typo in the caption: 'inerted' should be 'inserted'.","section":"Fig. 6"}],"recommendation":"major_revision","confidential_remarks":"The paper is clearly written and addresses a timely problem, but I am not convinced that the proposed method actually expands the scene rather than only stabilizing the captured region. The self-distillation signal is self-referential, and the evaluation has circularity and statistical weaknesses. These issues are load-bearing for the paper's central claim. I recommend asking the authors for substantial revisions, including a new loss or convincing evidence that the inserted content is actually realized, and a more rigorous evaluation protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"SceneExpander is a solid test-time adaptation recipe attached to a real workflow gap, but the headline claim overreaches: the inserted view is never actually used as a constraint. The only channel from I_g to the reconstruction is the backbone's forward pass, and the self-distillation loss (Eq. 8) just ties the student to its own EMA past. That makes the method an elaborate stabilization scheme, not a way to inject new content. If WorldMirror's prior suppresses the new region, these losses will happily stabilize the suppression.\n\nThe task framing is the best part. Iterative world building — extending a captured scene with a generated view that is explicitly misaligned — is genuinely under-explored, and the paper correctly identifies that existing world-generation and editing methods assume consistency or regenerate from scratch. The two-loss structure (anchor distillation on captured views to resist drift, inserted-view self-distillation as a soft regularizer) is a sensible decomposition, and the ablations tell a consistent story: each component helps a bit, and the full model wins. The human study (39 participants) and the Fig. 6 failure case give the paper real evidentiary weight.\n\nWhere it gets soft: the objective has no pixel-level term on I_g. No render-and-compare, no external geometry, no multi-view verification. The EMA teacher is initialized from the pre-insertion weights, so it inherits whatever suppression the backbone has, and the student is trained to match it. The GPT-IS and H-IS scores are the only check that anything from the inserted view materialized, and GPT-5.2 generated the views it later evaluates. The manual curation of inserted views adds selection bias, and Tabs. 1–2 report no error bars. These are fixable, but together they mean the central claim — that SceneExpander 'incorporates the inserted view as an expansion constraint' — is not established.\n\nNone of this is fatal to the paper's utility. If you read it as 'a way to adapt a reconstruction model to a misaligned extra image without destroying the captured region,' it works and is worth building on. The authors are honest about failure cases and hallucination risk. But the phrase 'expansion constraint' needs an actual constraint.\n\nI'd send it to peer review — the problem is important and the baseline is a reasonable starting point. A reviewer should push for either a pixel-level robust loss on the inserted view or a reframing that claims stabilization plus reliance on backbone prior. If the authors add the former, this could be a useful tool.","headline":"Real task, sensible recipe, but the inserted view is never actually used as a constraint — the central claim overreaches.","tokens_in":14066,"tokens_out":3748,"would_cite":false,"duration_ms":38464,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SceneExpander claims that a feed-forward 3D reconstruction model can be adapted at test time, using anchor distillation on captured views and self-distillation on an inserted AI-generated view, to expand a captured 3D scene with misaligned","keywords":["scene expansion","view insertion","test-time adaptation","distillation","multi-view reconstruction","3D Gaussian splatting","text-guided generation","world building"],"falsifier":"Take a captured multi-view scene with known camera poses and ground-truth geometry. Insert a real photograph of a different room (or a heavily displaced copy of a captured view) as the 'generated' view, so the misalignment is controlled and the correct expansion is known. If the adapted model renders the inserted content coherently from novel viewpoints and preserves the captured region, the central claim holds; if it produces ghosting, suppresses the inserted room, or drifts on the captured side, the claim fails.","tokens_in":12845,"feed_emoji":"🧩","tokens_out":5186,"duration_ms":49019,"temperature":0.7,"pith_summary":"The paper tackles a practical world-building workflow: starting from a real scene captured by multiple images, a user prompts an image generator to produce an additional view that extends the scene, and this view is inserted into the 3D reconstruction. The inserted view is usually not 3D-consistent with the captured geometry, so naively adding it corrupts the original reconstruction. The authors claim that test-time adaptation of a feed-forward reconstruction model, using two distillation signals—anchor distillation on captured views and self-distillation on the inserted view—can integrate the misaligned view while preserving the captured region. If correct, this would let creators iteratively expand real 3D scenes with text-controlled AI content without requiring perfectly 3D-consistent generation. The paper reports improved preservation (PSNR 21.21 vs 19.30) and expansion quality over baselines.","feed_headline":"Dual distillation lets 3D scenes absorb misaligned AI views","feed_subtitle":"A test-time teacher-student scheme keeps captured views stable while adding AI-generated content, beating its backbone on fidelity and expan","key_machinery":"The load-bearing object is the dual-teacher test-time adaptation loop built on a feed-forward reconstruction backbone named WorldMirror in the paper. An anchor teacher (frozen at pre-insertion weights) emits reference camera, depth, and normal maps for captured views; the student is penalized for deviating from them. A generation teacher, updated as an exponential moving average of the student, emits soft 'pseudo-geometry' targets for the inserted view; the student is lightly distilled toward these, so the inserted content is integrated as a weak constraint. Subset sampling of captured views, geometry-perturbation augmentation, and stochastic restoration complete the mechanism. The identity","core_discovery":"The central claim is that a parametric feed-forward 3D reconstruction model can be adapted at test time—using only internal distillation signals—to absorb a generatively synthesized view that is geometrically inconsistent with the original multi-view evidence. The adaptation freezes an 'anchor teacher' to provide fixed camera/depth/normal predictions on the captured views, pulling the student back toward the pre-insertion model; simultaneously, an EMA teacher provides soft geometric targets on the inserted view, and the student distills toward those soft labels with a small weight, so the new content is gently explained rather than hard-fitted. A stochastic-restoration step periodically nudg","pith_inferences":["The success of inserted-view self-distillation suggests a general recipe for integrating unreliable observations into a pretrained 3D model: trust the model's own iteratively refined predictions over the raw observation, and anchor against old evidence. The same principle could apply to updating radiance fields with noisy depth sensors or to online learning under distribution shift.","Because the only supervision for new content is the backbone's prior on a single image, the method's expansion quality is bounded by that prior. If the generative model produces a view with content the backbone has never seen, SceneExpander will likely imprint the backbone's hallucination—a limitation the paper acknowledges in its failure case.","A natural testable extension is long-horizon, multi-step expansion: the paper's stochastic restoration and EMA momentum would need tuning to prevent error accumulation over many sequential insertions; the paper lists this as future work."],"forward_implications":["Expanded 3D scenes can be produced from a single misaligned AI-generated view, without retraining the reconstruction model or requiring 3D-consistent generation.","The captured region is not merely tolerably preserved; on the paper's metrics it is better preserved than the backbone's own output (PSNR 19.30 → 21.21, SSIM 0.614 → 0.776), because anchor distillation actively stabilizes it.","The framework supports iterative world building: a user can chain multiple text-guided insertions, each extending the reconstruction with new content.","The ablations indicate that each component contributes: anchor distillation, subset sampling with augmentation, and stochastic restoration all improve the final trade-off."],"fun_headline_variants":["Text-guided 3Dexpansion handles misaligned insertions","Dual distillation steadies scenes as AI views join","Test-time tuning lets 3D scenes absorb new views","SceneExpander adapts to free-form view insertion"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method assumes the pre-trained reconstruction model's own view of the inserted image—iteratively refined through the EMA teacher—is a trustworthy guide to what the new content should look like in 3D, with no external ground truth or multi-view check. If that prior is uninformative or wrong, the expansion will be confidently wrong.","fun_headline_variants_meta":{"raw":{"variants":["Text-guided 3Dexpansion handles misaligned insertions","Dual distillation steadies scenes as AI views join","Test-time tuning lets 3D scenes absorb new views","SceneExpander adapts to free-form view insertion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000212,"raw_usage":{"total_tokens":1243,"prompt_tokens":723,"completion_tokens":520,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":467,"completion_tokens_details":{"reasoning_tokens":454}},"tokens_in":467,"tokens_out":520,"duration_ms":6200,"temperature":1.0,"reasoning_tokens":454,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T17:14:44.097963+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a captured multi-view scene with known camera poses and ground-truth geometry. Insert a real photograph of a different room (or a heavily displaced copy of a captured view) as the 'generated' view, so the misalignment is controlled and the correct expansion is known. If the adapted model renders the inserted content coherently from novel viewpoints and preserves the captured region, the central claim holds; if it produces ghosting, suppresses the inserted room, or drifts on the captured side, the claim fails.","supporting_citations":[],"review_version":1}