{"id":"4c019d5d-4347-43c8-aedc-d70b525e3943","arxiv_id":"2504.21836","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Injecting style image features into the last four cross-attention layers of a large reconstruction model transfers artistic appearance to 3D objects without training or test-time optimization.","lead":"3D objects generated by large reconstruction models can be stylized by feeding a reference artwork's visual features into specific attention layers of the model, with no extra training. This offers a fast, optimization-free route to restyle 3D assets, which matters for interactive content creation pipelines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on an unverified layer-wise geometry/appearance separation; the layer ablation that selects 'last 4' reports only geometry loss and no style fidelity, so the key design choice is not supported.","rationale":"The reader's verdict CONDITIONAL is appropriate. The paper proposes a clever, training-free 3D stylization method with clear practical advantages: a 35-second runtime versus hours for optimization-based baselines, publicly available code, a user study with strong preference, and plausible qualitative results. Credit is due for the efficient design and for building on a valid insight from 2D attention-based style transfer. However, the central claim that LRMs exhibit a layer-wise geometry/appearance split is the load-bearing assumption. The selection of 'last 4 layers' and alpha=0.8 is based on a small, qualitative ablation: Table 1 shows only Chamfer distance for varying layer counts, without any style fidelity measure, and Fig. 5 shows a single owl example for alpha. This is not enough to establish a stable functional separation. If that separation is not robust, the method's effectiveness could depend on the specific object or style image, and the design choices would be overfit to the handful of examples shown. The paper's own limitation statement acknowledges dependence on InstantMesh's reconstructions, but does not acknowledge the deeper assumption about layer specialization, which is more central to the method's novelty. A systematic ablation with quantitative style metrics and a second architecture would directly test this assumption. Since the paper is already CONDITIONAL, my concern does not change that verdict but reinforces it.","tokens_in":13845,"tokens_out":4984,"duration_ms":51330,"concrete_test":"Run a systematic layer ablation on a diverse set of at least 15 objects x 5 style images, measuring both Gram-matrix style fidelity and Chamfer distance for every injection depth (last 1, 2, 4, 6, 8, 10, 16 layers). If the last-4 layer choice is not consistently on the Pareto frontier (or if the optimal depth varies across objects/styles), the fixed 'last 4' setting is unjustified. Additionally, repeat the same experiment on a second LRM with a different transformer depth (e.g., a 12-layer LRM or GS-LRM); if the method fails or requires a different depth, the claim of a general LRM capability is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The method's core mechanism is that the last four cross-attention blocks of the LRM encode appearance while earlier blocks encode geometry. On this hypothesis the paper selects the injection depth (last 4) and blending weight (alpha=0.8). But the evidence is thin: Table 1's layer ablation reports only Chamfer distance (a geometry measure) for a single object, with no style-fidelity metric, so the reader cannot tell whether the last-4 choice is actually necessary for style transfer or whether it is a noisy optimum. The ablations for alpha (Fig. 5) are qualitative and on one example. Moreover, the hypothesis itself is not probed directly; it is inferred from the success of the stylization results, which risks circularity: the method works for the configurations that were selected because they were observed to work. If the geometry/appearance separation is not a stable property of LRMs, then the method's effectiveness and its stated design principles (Eq. 2 with last-4 blocks) are not transferable to other architectures or style distributions. The paper's own Limitations section concedes dependence on InstantMesh's reconstruction quality, but not on this more specific architectural assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a training-free and optimization-free method for 3D appearance stylization built on InstantMesh, a large reconstruction model (LRM). The style image is encoded with the same ViT encoder used for the multi-view content images, and its key/value features are injected into the last four cross-attention blocks of the triplane decoder. The attention output with respect to the style image is blended with the content attention output through a scalar weight alpha=0.8. The paper reports quantitative results on 15 objects and 2 styles, including Gram-matrix style fidelity, Chamfer distance, F-score, and a user study, and claims superior stylization quality with a runtime of 35.5 seconds, compared with hours for optimization-based baselines.","tokens_in":14036,"tokens_out":4408,"duration_ms":47779,"significance":"If the claims hold, the method is a simple and practical contribution: it repurposes an existing LRM without any training or per-object optimization, achieving stylization in under a minute while largely preserving geometry. The idea of probing attention blocks in 3D reconstruction models for appearance-style transfer is novel and connects naturally to 2D attention-based stylization. The paper also releases code and appears reproducible, and the qualitative results are generally plausible. However, the empirical support for the key design choices and the strength of the quantitative claims is currently thinner than the narrative suggests, and several load-bearing points need to be strengthened.","major_comments":[{"comment":"The choice of injecting style features into the last four transformer blocks is not supported by the reported ablation. Table 1 reports only Chamfer distance for a single object across layer counts, and by that metric one layer (CD=0.00018) is clearly better than four layers (CD=0.00594). No style-fidelity metric is reported for the ablation, so the reader cannot determine whether the last-4 choice is needed for style transfer or is merely a noisy compromise on geometry. Please report style fidelity, content fidelity, and ideally per-object results for every layer count, and do the same for the alpha ablation in Fig. 5, which is currently only qualitative.","section":"Section 3.2, Table 1"},{"comment":"The quantitative evaluation is too thin for the strength of the claims. It uses only 15 objects and 2 styles, with no error bars, confidence intervals, or significance tests, and the user study reports raw percentages without any statistical test. In addition, the claim that 3D approaches including ours perform on-par on content fidelity is not supported by Table 2: the Chamfer distance of Ours (0.00594) is roughly an order of magnitude larger than ARF (0.00060) and StyleRF (0.00072). Please add variance measures and per-object results, temper the content-fidelity claim, or justify why the difference is acceptable for the target use case.","section":"Section 4, Table 2"},{"comment":"The method rests on the hypothesis that earlier transformer blocks in the triplane decoder determine geometry while later blocks determine appearance. This hypothesis is asserted and supported only indirectly by the success of the chosen configuration; it is not directly probed. Since the choice of last-4 layers and alpha=0.8 depends on this assumed layer-wise separation, the paper should provide a direct test, such as an analysis of attention maps or feature statistics per block, and a robustness check showing that the conclusion is not specific to a single style set or a single LRM instance. This would also clarify whether the method is transferable to other reconstruction architectures.","section":"Section 3.2"}],"minor_comments":[{"comment":"There are several grammatical issues, e.g., 'large reconstruction models ... has a similar capability' and 'the certain attention blocks ... capture'; these should be corrected.","section":"Abstract and Introduction"},{"comment":"In Eq. (2), Q_in is used for both style and content branches but is not defined; also, clarify whether the two attention outputs are normalized before blending and whether the softmax scaling uses sqrt(d_k) consistently, as in Eq. (1).","section":"Equations (1)-(2)"},{"comment":"The figure legend says 'last 1, 2, 4 ... 16 layers' but it would help to state explicitly that these are the last N of the 16 transformer blocks in the triplane decoder.","section":"Figure 4 and Table 1"},{"comment":"The table header formatting for the run-time column is irregular, and the units 'min' versus 'sec' should be unified or clearly labeled in every row.","section":"Table 2"},{"comment":"The user study reports percentages from 27 participants but no statistical significance test; at minimum, report pairwise preference counts and a test such as a binomial test or Wilcoxon signed-rank test.","section":"User Study (Table 3)"},{"comment":"The Limitations section attributes dependence to InstantMesh's reconstruction quality but does not mention the more specific dependence on the hypothesized geometry/appearance separation in the attention blocks; this should be acknowledged.","section":"Limitations"}],"recommendation":"major_revision","confidential_remarks":"The core idea is attractive and the method is likely to work as described, but the evidence for the central design choices (last-4 layers, alpha=0.8) and the quantitative superiority claims needs substantial strengthening. I would be comfortable with acceptance after the authors either provide a much more thorough evaluation or substantially soften the claims. No concerns about attribution or misconduct; the related work is adequately covered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the attached paper. Bottom line: it is a genuinely new trick worth knowing, and the evaluation is a bit thin, but the core result holds.\n\nThe real contribution is the observation, not shown before, that the cross-attention blocks in an LRM's triplane decoder separate appearance from geometry: early blocks determine shape, later blocks carry texture/color. The authors turn that into a training-free stylization method by injecting ViT features of a style image into the last four cross-attention blocks and blending the attention outputs (alpha=0.8). No fine-tuning, no per-asset optimization, 35 seconds per stylization. The qualitative results are convincing, the user study is strongly in their favor, and the runtime advantage over NeRF-based baselines is real. Extending 2D attention-swapping to 3D reconstruction models is a sensible move and, to my knowledge, new.\n\nSoft spots are in the evidence. The quantitative evaluation uses 15 objects and 2 styles, no error bars, no significance tests. The layer ablation (Table 1) reports only Chamfer distance for a single object, so it shows geometry degradation but not whether the last-4 choice actually maximizes style transfer. Alpha is set from one qualitative example. The layer-separation hypothesis is plausible and matches 2D findings, but the paper does not probe it directly; it is inferred from the fact that the chosen configuration works. That is post-hoc selection rather than circular reasoning, and it is not fatal, but it leaves the generality of the method unverified across other LRM architectures and style distributions. The Limitations section also notes dependence on InstantMesh's reconstruction quality.\n\nWho should read it: anyone working on 3D content creation or style transfer. It is a solid within-subfield step, not a reorientation. I would send it to a serious referee; with requests for stronger stats, style-aware ablations, and a second backbone, it could be a well-cited paper. As is, it deserves review, not rejection.","headline":"A simple, genuinely new trick for training-free 3D stylization, with evaluation thinner than the claims; deserves serious peer review.","tokens_in":14565,"tokens_out":3054,"would_cite":true,"duration_ms":30618,"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":"A style image's features, injected into the last four cross-attention layers of a large reconstruction model, transfer the style to a 3D object in one forward pass while preserving its geometry.","keywords":["3D stylization","large reconstruction model","style transfer","cross-attention injection","triplane decoder","training-free","multi-view consistency","single-image 3D reconstruction"],"falsifier":"Concretely, take the same 15 objects and 30 style images used in the paper, and run the injection on the first four transformer blocks instead of the last four. If the early-block variant achieves the same style fidelity with no larger Chamfer distance (the metric for geometry preservation), then the claim that appearance features live specifically in the late layers is wrong; if early-block injection distorts the shape or fails to transfer the style, the layer-wise separation is supported.","tokens_in":13630,"feed_emoji":"🎨","tokens_out":18330,"duration_ms":153625,"temperature":0.7,"pith_summary":"The paper sets out to show that large reconstruction models, the transformer-based networks that turn a few multi-view images into a 3D representation, already know how to separate geometry from appearance. Its core claim is that the early transformer blocks of the triplane decoder—the part of the model that turns image features into a 3D field—fix the object's shape while later blocks determine its look, so feeding the image-encoder features of a style image into the last four cross-attention blocks, and blending those attention outputs with the original content attention using a factor of $\\alpha=0.8$, transfers the style with no training or per-asset optimization. The payoff is practical: once an object's multi-view images exist, stylization drops from a minutes-to-hours optimization process to a single 35-second reconstruction pass, and the stylized object stays multi-view consistent because the same 3D geometry is reused. This would make style control an interactive knob on top of existing single-image 3D generators rather than a separate, expensive post-processing stage.","feed_headline":"Restyle an existing 3D asset in 35 seconds, no retraining","feed_subtitle":"No optimization, no fine-tuning: restyle a 3D object with a reference image in a single forward pass.","key_machinery":"The load-bearing mechanism is cross-attention between triplane tokens and image tokens in the reconstruction model's transformer; the triplane is a compact 3D representation made of three axis-aligned feature planes that are rendered into an object. In each block, triplane features provide the queries $Q_{\\mathrm{in}}$, and the ViT embeddings of the six multi-view images provide the keys and values; the score $QK^T/\\sqrt{d_k}$ decides which parts of the input images explain each 3D location. The method computes a second cross-attention in the last four blocks using keys and values from the style image's ViT embeddings, then blends the two attention outputs with the scalar $\\alpha$ (set to 0.8). This leaves the geometry-forming attention in the early blocks untouched and re-routes only the appearance-forming attention, which is why the stylized object stays consistent across views without optimization.","core_discovery":"The central discovery is a layer-wise functional separation inside the triplane decoder of a large reconstruction model: earlier cross-attention blocks determine the 3D geometry, and the last blocks determine appearance. Encoding a reference style image with the same vision transformer used for the multi-view content images, then substituting its keys and values in the last four cross-attention blocks, makes the model's own priors produce a stylized 3D asset whose shape matches the unstylized reconstruction. The substitution is done as a weighted blend of two attention outputs, $$\\mathrm{Attention}(Q_{\\mathrm{in}},K,V)=\\mathrm{softmax}\\left(\\frac{Q_{\\mathrm{in}}K_{\\mathrm{style}}^T}{\\sqrt{d_k}}\\right)V_{\\mathrm{style}}\\,\\$\\alpha$+\\mathrm{softmax}\\left(\\frac{Q_{\\mathrm{in}}K_{\\mathrm{content}}^T}{\\sqrt{d_k}}\\right)V_{\\mathrm{content}}\\,(1-\\$\\alpha$),$$ with $\\alpha=0.8$; the weighting preserves fine details (for example, an owl's eyes) that a full style takeover would erase. The paper reports that this training- and optimization-free recipe achieves the best style-fidelity score among its baselines and runs in 35 seconds, while keeping Chamfer distance and F-scores close to those of per-asset 3D stylization methods.","pith_inferences":["Editorial inference: If the early/late layer separation is a general property of large reconstruction models, the same injection recipe should transfer to other triplane-based or Gaussian-splat reconstructors; testing it on a Gaussian-splat model would either broaden the method or reveal that the separation is specific to InstantMesh's architecture.","Editorial inference: The success of blending attention outputs, rather than concatenating tokens or averaging embeddings, suggests that controllable attributes are carried by the key-value pairs at particular layers, not by the input embeddings; that makes attention-layer probing a cheap way to locate other steerable attributes (lighting, material, viewpoint) in pretrained 3D generators.","Editorial inference: Because the paper fixes $\\alpha=0.8$ globally, a natural extension is to predict $\\alpha$ per object or per style; texture-heavy styles may need more style weight in later layers while color-only styles need less, and this could be validated by sweeping $\\alpha$ and measuring the style-fidelity versus Chamfer-distance trade-off.","Editorial inference: Since the style is injected at the triplane level rather than into rendered images, the stylization is inherently multi-view consistent; this may make the method a more reliable front end for downstream rendering or texturing pipelines than image-space stylization followed by reconstruction."],"forward_implications":["Stylization becomes interactive: once the six multi-view images of an object have been generated, each new style costs only one reconstruction run, roughly 35 seconds on a single A100 GPU, instead of a per-asset optimization loop.","Because only the last four cross-attention blocks are re-routed, the same 3D geometry can be restyled any number of times without accumulating geometric drift; the shape is fixed by the earlier blocks in every pass.","The blending weight $\\alpha$ gives a continuous dial between content preservation and style strength, so users can keep small structural details in the appearance (like an owl's eyes) by not pushing $\\alpha$ to 1.","The method inherits the generality of the large reconstruction model's pretraining: any style image that the ViT encoder can embed becomes a valid style source, with no style-specific fine-tuning or paired data."],"supporting_citations":[{"why":"InstantMesh is the reconstruction pipeline analyzed and modified; it supplies the multi-view diffusion plus triplane-decoder architecture and the baseline for Chamfer distance evaluation.","marker":"[Xu et al. 2024a]"},{"why":"The LRM is the large reconstruction model architecture with the transformer cross-attention blocks whose layer-wise geometry/appearance separation is investigated.","marker":"[Hong et al. 2023]"},{"why":"Zero123++ generates the six consistent multi-view images that provide the content tokens for the cross-attention injection.","marker":"[Shi et al. 2023]"},{"why":"Cross-image attention is the 2D precedent for style transfer by injecting style features into attention layers, and its Gram-matrix evaluation protocol is reused for style fidelity.","marker":"[Alaluf et al. 2024]"},{"why":"Neural style transfer with Gram matrices supplies the style-fidelity metric used in the quantitative comparison.","marker":"[Gatys et al. 2016]"},{"why":"ARF is a per-asset NeRF-optimization stylization baseline that the paper compares against in runtime and quality.","marker":"[Zhang et al. 2022a]"},{"why":"StyleRF is a second per-asset NeRF stylization baseline included to show the efficiency gain of the training-free approach.","marker":"[Liu et al. 2023b]"}],"fun_headline_variants":["35s 3D restyle from a single reference image","Training-free 3D stylization via attention swap","One-pass 3D restyle from reference style image"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a pretrained large reconstruction model separates geometry and appearance across its transformer layers, with early blocks fixing shape and later blocks fixing appearance, so injecting style features only into the last four blocks preserves the 3D form; if this layer-wise specialization is not a stable property of the model, the stylization will either fail to transfer the style or will deform the geometry.","fun_headline_variants_meta":{"raw":{"variants":["35s 3D restyle from a single reference image","Training-free 3D stylization via attention swap","One-pass 3D restyle from reference style image"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000404,"raw_usage":{"total_tokens":2135,"prompt_tokens":1009,"completion_tokens":1126,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":1073}},"tokens_in":625,"tokens_out":1126,"duration_ms":9398,"temperature":1.0,"reasoning_tokens":1073,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:51:30.026992+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Concretely, take the same 15 objects and 30 style images used in the paper, and run the injection on the first four transformer blocks instead of the last four. If the early-block variant achieves the same style fidelity with no larger Chamfer distance (the metric for geometry preservation), then the claim that appearance features live specifically in the late layers is wrong; if early-block injection distorts the shape or fails to transfer the style, the layer-wise separation is supported.","supporting_citations":[],"review_version":1}