{"id":"d691f90b-0abc-4cc0-8fa2-1c272e2b5a56","arxiv_id":"2608.13255","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A training-free plugin that speeds up multi-view texture diffusion by transporting anchor views' per-step clean-signal updates to non-anchor views via geometry, preserving fidelity better than temporal caches at over 2x speedups.","lead":"GeoCache accelerates multi-view 3D texture diffusion by running the denoiser on a few anchor views per step and transporting their per-step changes to other views through geometric correspondence. It reaches up to 2.21x denoiser speedup with better fidelity than temporal caches, without retraining.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 'same transferred configuration' for SyncMVD conflicts with Table 1, so the portability claim is unverified.","rationale":"The paper's central claim about the speed-fidelity trade-off is supported by the reported numbers and ablations; the mechanism is clearly described and the delta-transport choice is ablated. However, the paper also claims that the same configuration transfers unchanged from Hunyuan to SyncMVD, which is contradicted by the configuration tuples in Table 1. This is a concrete, checkable inconsistency that undermines the portability/no-tuning claim. It does not invalidate the primary trade-off results, so the conditional verdict remains appropriate. The reader's identified assumption about anchor-only attention context is related but less directly contradicted by the paper's own motivating oracle results.","tokens_in":13114,"tokens_out":13258,"duration_ms":128664,"concrete_test":"Run GeoCache on SyncMVD with the exact Hunyuan configuration (a=2, E=5, S=10) on the same eval200 assets and seeds, and report MV-LPIPS, MV-PSNR, denoiser-loop speedup, and FLOPs. Compare to the reported E=2 and E=3 rows: if the E=5/S=10 row is within the same fidelity/speed range, the transfer claim holds and Table 1 should include it; if it is substantially worse, the method requires per-backbone scheduling and the 'transferred unchanged' statement is contradicted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and Section 4.2 assert that the Hunyuan schedule transfers unchanged to SyncMVD and leads every metric above 2x. In the paper's own notation (Section 4.1), a GeoCache configuration is the tuple (a, E, S). The Hunyuan headline is a=2, E=5, S=10, while the SyncMVD rows are a=2, E=3, S=20 and a=2, E=2, S=20. No SyncMVD row with E=5, S=10 is reported. Therefore the claim of an unchanged transfer is either false or refers to an unreported configuration, and the SyncMVD results come from different schedule hyperparameters. This does not refute the headline speed-fidelity trade-off, but it removes the evidence for the paper's no-tuning portability contribution and makes the schedule sensitivity a live question.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GeoCache, a training-free plugin that accelerates multi-view texture diffusion by running the denoiser on a rotating subset of anchor views at each step and transporting the anchors' per-step change in the predicted clean signal (Delta x0) to non-anchor views through a fixed correspondence gather built from rendered position maps. Periodic full-view refreshes bound drift, and the transported x0 is converted to the sampler's native parameterization to preserve trajectory consistency. The authors evaluate on Hunyuan3D-2.1, SyncMVD, and MVPainter against temporal caches (TeaCache, MagCache, FORA, TaylorSeer) and step reduction, reporting that GeoCache achieves the best fidelity among methods above 2x denoiser speedup on Hunyuan (2.21x, MV-LPIPS 0.0293, MV-PSNR 33.60 dB), the highest speedup on SyncMVD, and the best fidelity/cost on MVPainter. Ablations identify delta transport, refresh placement, and correspondence as load-bearing components.","tokens_in":13260,"tokens_out":7003,"duration_ms":66064,"significance":"If the reported results hold, GeoCache identifies and exploits a genuinely complementary redundancy axis, cross-view geometric evolution of x0, that step caches do not use, with a simple and parameter-free core operation (Eq. 1) whose only adjusted knobs are schedule hyperparameters. The evaluation is unusually thorough: three backbones with different samplers and view counts, matched seeds and protocols, per-asset standard deviations, component ablations, interaction cells, and honest negatives (K=1 taps, confidence threshold as a no-op). The main correction needed is the unsupported 'same transferred configuration' claim for SyncMVD, which Table 1 contradicts; once the portability evidence is either supplied or the claim is softened, the core speed-fidelity contribution is defensible and relevant to the multi-view texturing community.","major_comments":[{"comment":"The abstract and Section 4.2 claim that the Hunyuan-tuned GeoCache schedule transfers unchanged to SyncMVD and leads every metric above 2x, but Table 1 lists SyncMVD configurations a=2, E=3, S=20 and a=2, E=2, S=20, whereas the Hunyuan headline is a=2, E=5, S=10. No SyncMVD row with E=5 and S=10 is reported, so the 'transfers unchanged' claim is not supported by the data in the manuscript. Please report the exact transferred configuration on SyncMVD or revise the portability claim; as written, the schedule-sensitivity question remains open.","section":"Abstract and Section 4.2 vs Table 1"},{"comment":"Section 3.2 states that the Hunyuan operating point runs 10 UniPC steps, 'four of them full', and describes a two-step head, one mid-trajectory refresh, and a tail step, while Table 1 labels the same operating point as a=2, E=5, S=10, with E defined in Section 4.1 as the number of full-view refresh steps placed at the end of the trajectory. These statements are inconsistent about both the number and placement of full steps. Please specify the complete schedule, including which steps are full, so that the headline result is reproducible.","section":"Section 3.2 vs Section 4.1 and Table 1"}],"minor_comments":[{"comment":"The text and Figure 4 refer to 'the best temporal cache' per backbone without naming which baseline appears in each panel; since Table 1 shows the best temporal method differs across backbones (MagCache on Hunyuan, FORA on SyncMVD, and step reduction on MVPainter), please label each panel's baseline in the caption or the text.","section":"Section 4.2 and Figure 4"},{"comment":"The reconstruction formulas use sqrt(alpha_bar_t) without defining alpha_bar_t; please add the missing notation definitions or a pointer to the sampler's convention.","section":"Section 3.2"},{"comment":"The note 'bSeam below the stock reference of 0.061 signals blur' is confusing because some cells with below-stock seam values are discussed positively in the text; please clarify the intended interpretation of the superscript b.","section":"Table 2"},{"comment":"There is a typo in the first affiliation: 'Eletrical' should be 'Electrical'.","section":"Affiliation"},{"comment":"The abstract uses the phrase 'the same transferred configuration' before the configuration notation (a, E, S) is introduced in Section 4.1; consider defining the tuple earlier or rephrasing the abstract to avoid ambiguity.","section":"Abstract and Section 4.1"},{"comment":"The schedule labels in Figure 3b, such as '0,31,4 2,5', are difficult to read because of missing spacing; please add explicit separators and clarify which steps are full refreshes.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is generally well executed and the core speed-fidelity claim is credible. The portability claim in the abstract is contradicted by Table 1, so the manuscript needs a substantive correction or a revised claim; this is fixable within the scope of the revision and does not refute the method itself."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Alex,\n\nRead GeoCache. The core idea is good and new: instead of reusing outputs across denoising steps, transport the per-step delta of the predicted clean signal (x0) across geometrically corresponding views, adding that increment to each target view's own retained state. The motivating study is the right kind of evidence — it shows that copying states fails while transporting increments works. The evaluation is thorough: three backbones, matched protocols, standard deviations, ablations that isolate delta transport, and honest negatives in the appendix. On the quality side, the paper is genuinely solid.\n\nThe soft spot is the portability claim. The abstract and Section 4.2 say the Hunyuan schedule transfers unchanged to SyncMVD and leads every metric above 2x. But Table 1 shows the Hunyuan headline is a=2, E=5, S=10, while the SyncMVD rows are a=2, E=3, S=20 and a=2, E=2, S=20. No SyncMVD row with E=5, S=10 is reported. So either the claim refers to an unreported configuration, or the SyncMVD results were obtained with different schedule hyperparameters. The stress-test note is right: this removes the evidence for the no-tuning portability contribution and makes schedule sensitivity a live question. The paper's own limitations section also admits retuning on MV-Adapter, which further muddies the \"unchanged\" narrative.\n\nThe anchor-only attention context issue is real but acknowledged and empirically bounded; the authors say the schedule bounds the approximation and that multistep stability is an empirical check, not a guarantee. I'd accept that as a limitation, not a fatal flaw. The absence of released code is a practical annoyance but not a scientific one.\n\nOverall: the central mechanism is well supported, the speed-fidelity results are credible at the reported operating points, and the paper is worth a serious referee. But the schedule-transfer claim needs to be verified or walked back, and the exact configurations for every backbone should be in the main table or a clearly labeled appendix. I'd recommend conditional acceptance with a request for that correction.\n\nBring it to the reading group if you want a good example of how to do an ablation-driven acceleration paper.\n\nBest,\nSam","headline":"A genuinely new cross-view caching idea with a strong evaluation, but the claimed \"same transferred configuration\" portability is contradicted by the paper's own Table 1 and needs a fix before publication.","tokens_in":13799,"tokens_out":1174,"would_cite":true,"duration_ms":13616,"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":"The paper claims that geometrically corresponding surface points share transferable denoising evolution in multi-view texture diffusion, and that transporting per-step clean-signal changes across views yields a stronger speed-fidelity…","keywords":["training-free acceleration","multi-view diffusion","texture generation","geometric correspondence","cache","delta transport","denoising","3D generation"],"falsifier":"Take an asset with a region seen by only one anchor view, run GeoCache, and compare the transported $x_0$ of a non-anchor view at a cached step with the genuine $x_0$ from a full forward; if the error in that region exceeds the gap between GeoCache's reported 0.029 MV-LPIPS and the value-copy failure at 0.101, the correspondence-based transport is not carrying the shared evolution.","tokens_in":12892,"feed_emoji":"🎨","tokens_out":7875,"duration_ms":69506,"temperature":0.7,"pith_summary":"This paper argues that the redundancy in multi-view texture diffusion is geometric, not just temporal: the same surface point seen from different views changes in a transferable way in the predicted clean-image space, even though intermediate features stay view-specific. It introduces GeoCache, a training-free plugin that runs the denoiser on a rotating subset of anchor views at each step and transports their per-step change ($\\Delta x_0$) to every other view along geometry-based correspondences, preserving each view's own state. Periodic full-view refreshes and sampler-consistent reconstruction keep the error bounded and the denoising trajectory valid. On Hunyuan3D-2.1 it reports a $2.21\\times$ denoiser-loop speedup at the lowest MV-LPIPS among tested methods above $2\\times$, and it transfers unchanged to SyncMVD and MVPainter. A sympathetic reader would care because painting is the dominant cost of 3D asset generation, and existing step caches degrade cross-view consistency, whereas this method preserves it.","feed_headline":"Cache across views, not steps, to speed 3D texturing 2.2x","feed_subtitle":"GeoCache transports each step's change between matching surface points, keeping texture fidelity while skipping most view renders.","key_machinery":"The central object is the geometric delta-transport rule, Eq. (1): $x_0^{(v)}(t) = x_0^{(v)}(t-1) + \\mathcal{G}_{A\\to v}[\\Delta x_0^{(A)}(t)]$, where $\\mathcal{G}_{A\\to v}$ is a row-stochastic sparse gather that maps tokens in anchor views to corresponding tokens in target view $v$ using precomputed position maps (K nearest taps within 1% of the bounding-box diagonal, area-weighted). It operates on the first difference of the predicted clean signal, not on features or full states, which lets each target view keep its own content and noise while inheriting the shared surface's denoising evolution. Around this sit three supporting mechanisms: batch-sliced anchor forwards that preserve each retained view's positional identity in multi-view attention, periodic full-view refresh steps that re-ground every view, and sampler-consistent reconstruction that converts transported $x_0$ back into $\\epsilon$, $v$, or flow velocity so the solver's history stays valid. The method requires no retraining and consumes position maps the texturing pipeline already computes.","core_discovery":"GeoCache's central claim is that in geometry-conditioned multi-view texture diffusion, the predicted clean signal $x_0$ at a denoising step evolves in a way that is shared across views for the same surface point, so only a few anchor views need to run the denoiser at each step. The per-step change $\\Delta x_0$ of the anchors is transported through a fixed sparse linear gather operator built from position maps and added to each non-anchor view's own previous $x_0$, rather than overwriting it; unmatched tokens (disocclusions, background) simply keep their state until the next full refresh. Four full-view steps per trajectory bound the accumulated drift, and converting the transported $x_0$ back into the sampler's native parameterization keeps the multistep history consistent. The paper reports that on Hunyuan3D-2.1 this yields a $2.21\\times$ denoiser-loop speedup at MV-LPIPS $0.0293$ and MV-PSNR $33.60$ dB, the best fidelity of any method tested above $2\\times$, with the same schedule leading fastest and leanest on SyncMVD and best fidelity and lowest FLOPs among accelerated methods on MVPainter.","pith_inferences":["Inference: because GeoCache's saving is along the view axis, the same plugin should combine with temporal caches such as TeaCache and MagCache; the paper does not test this stacking, but the two reuse axes are orthogonal in the sense that one omits views and the other omits steps.","Inference: the delta-transport update resembles an explicit Euler step on a per-surface-point increment carried across views; this suggests a stability analysis could be formalized by bounding the spectral norm of $\\mathcal{G}_{A\\to v}$ and the deviation between anchor-only and full multi-view attention, turning the empirical refresh schedule into a derived one.","Inference: adaptive anchor selection based on visibility would likely extend the method to sparsely observed regions; the paper explicitly leaves this as future work, but the correspondence operator already carries visibility information through the tolerance test.","Inference: the same geometric axis could apply to video diffusion with a fixed multi-view rig or a moving camera, where scene points are observed from multiple frames; the method's 'views' need not be static renders of one object."],"forward_implications":["If GeoCache's claim holds, multi-view texture diffusion can be accelerated by omitting whole-view forwards rather than skipping denoising steps, so every step of the trajectory is still computed for at least the anchors and cross-view harmonization is never entirely absent.","The same transported-increment rule should transfer to any batched multi-view pipeline with geometric correspondence, since the method only changes the definition of an anchor (batch row, tile row, or token-frame slice), which the paper demonstrates by carrying the Hunyuan-tuned schedule unchanged to SyncMVD and MVPainter.","Because GeoCache composes geometric caching with mild step reduction, it can be stacked with faster samplers or distillation rather than competing with them, extending the speed-fidelity frontier beyond what either axis alone reaches.","The reported slope advantage (each additional $0.1\\times$ of speed costs GeoCache about $+3.1\\%$ MV-LPIPS, versus $+12.4$ to $+33.5\\%$ for temporal caches) implies that the gap in fidelity grows as targets move past $2\\times$, making the geometric axis the cheaper way to push speed.","On cost centers where the denoiser loop dominates (production-resolution assets), a $2.21\\times$ loop speedup translates into a larger end-to-end saving, though at the default $6\\times512^2$ resolution the measured end-to-end gain is $1.07\\times$."],"supporting_citations":[{"why":"Main substrate: the geometry-conditioned 6-view Hunyuan3D-2.1 Paint pipeline supplies the position maps, batch layout, and UniPC trajectory on which GeoCache is measured.","marker":"Hunyuan3D Team 2025b"},{"why":"SyncMVD is the second substrate and the closest antecedent, synchronizing per-view denoising through a shared UV buffer; GeoCache must beat its already view-shared design.","marker":"Liu et al. 2024"},{"why":"UniPC is the sampler whose $x_0$ prediction and multistep history GeoCache preserves through sampler-consistent reconstruction, and step reduction with UniPC is the main non-cache baseline.","marker":"Zhao et al. 2023"},{"why":"TeaCache is the timestep-based temporal cache baseline whose output reuse fails on multi-view consistency in the motivating study.","marker":"Liu et al. 2025a"},{"why":"MagCache is the magnitude-aware temporal cache baseline and the source of the seam-error failure illustrated in Figure 2a.","marker":"Ma et al. 2025"},{"why":"FORA is the fixed-interval block-reuse baseline on SyncMVD and MVPainter that defines the sub-$2\\times$ operating point.","marker":"Selvaraju et al. 2024"},{"why":"TaylorSeer is the forecasting baseline that competes with GeoCache on long Euler trajectories and on MV-Adapter.","marker":"Liu et al. 2025c"},{"why":"Reverse reprojection caching supplies the computational pattern of transporting a per-pixel quantity through correspondence with periodic refresh, which GeoCache adapts to the denoising trajectory.","marker":"Nehab et al. 2007"},{"why":"LPIPS defines the MV-LPIPS fidelity metric used for every trade-off comparison.","marker":"Zhang et al. 2018"},{"why":"MVPainter is the third substrate whose geometric control and 75-step trajectory show the largest speed-fidelity gain.","marker":"Shao et al. 2025"}],"fun_headline_variants":["Skip view renders, not steps: 2.2x faster 3D texturing","Train-free trick: reuse geometry-aligned updates for 2.2x speedup","Transport surface updates across views to cut diffusion cost","GeoCache: 2.2x faster texture diffusion via cross-view reuse","Geometry-aware caching speeds multi-view texture diffusion 2.2x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"GeoCache assumes that running the denoiser on only the anchor views at a cached step, which removes the other views from multi-view attention, leaves the anchors' predicted per-step change close enough to the full-context value that transporting it to all views does not corrupt the result; the schedule that bounds the accumulated error is chosen empirically rather than derived.","fun_headline_variants_meta":{"raw":{"variants":["Skip view renders, not steps: 2.2x faster 3D texturing","Train-free trick: reuse geometry-aligned updates for 2.2x speedup","Transport surface updates across views to cut diffusion cost","GeoCache: 2.2x faster texture diffusion via cross-view reuse","Geometry-aware caching speeds multi-view texture diffusion 2.2x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001288,"raw_usage":{"total_tokens":5360,"prompt_tokens":1143,"completion_tokens":4217,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":759,"completion_tokens_details":{"reasoning_tokens":4119}},"tokens_in":759,"tokens_out":4217,"duration_ms":27819,"temperature":1.0,"reasoning_tokens":4119,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:16:37.979121+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an asset with a region seen by only one anchor view, run GeoCache, and compare the transported $x_0$ of a non-anchor view at a cached step with the genuine $x_0$ from a full forward; if the error in that region exceeds the gap between GeoCache's reported 0.029 MV-LPIPS and the value-copy failure at 0.101, the correspondence-based transport is not carrying the shared evolution.","supporting_citations":[],"review_version":1}