Pith. sign in

REVIEW 3 major objections 6 minor 5 references

3D Stylization via Large Reconstruction Model

T0 review · 3 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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.

desk verdict A simple, genuinely new trick for training-free 3D stylization, with evaluation thinner than the claims; deserves serious peer review. read the letter →

arxiv 2504.21836 v1 pith:HIJMRZKX submitted 2025-04-30 cs.CV

Ipek Oztas , Duygu Ceylan , Aysegul Dundar This is my paper · ORCID
classification cs.CV
keywords 3Dstylizationlargereconstructionmodelstyletransfercross-attentioninjectiontriplanedecodertraining-freemulti-viewconsistencysingle-image
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

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.

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 (3)
  1. [Section 3.2, Table 1] 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.
  2. [Section 4, Table 2] 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.
  3. [Section 3.2] 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.
minor comments (6)
  1. [Abstract and Introduction] There are several grammatical issues, e.g., 'large reconstruction models ... has a similar capability' and 'the certain attention blocks ... capture'; these should be corrected.
  2. [Equations (1)-(2)] 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).
  3. [Figure 4 and Table 1] 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.
  4. [Table 2] 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.
  5. [User Study (Table 3)] 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.
  6. [Limitations] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the method is a feedforward feature injection into a frozen pretrained LRM, with hyperparameters selected by ablations and evaluated against external baselines.

full rationale

The paper's derivation is not circular. The core operation, Eq. 2, blends cross-attention outputs from style and content image tokens in the last four blocks of a frozen InstantMesh triplane decoder. Nothing in Eq. 2 by construction guarantees the reported style fidelity or geometry preservation; those outcomes are empirical and are measured against external baselines (Table 2) and a user study (Table 3). The choices of 'last 4 layers' and alpha=0.8 are hyperparameters selected from ablations (Table 1, Fig. 5), and hyperparameter selection based on observed performance is not a fitted parameter renamed as a prediction. The geometry/appearance separation is explicitly presented as a hypothesis ('We hypothesize that while the earlier transformer blocks are crucial for determining the geometry...'), and while that hypothesis may be weakly supported, a weak or unverified assumption is a correctness risk, not circularity. The citation to Alaluf et al. 2024 for the analogous 2D behavior is external, not self-citational, and the paper's self-citations (Dundar et al., Ceylan et al., Yildirim et al.) appear only in related work and are not load-bearing for the central claim. The Limitations section acknowledges dependence on InstantMesh's reconstruction quality, which is a scope limitation, not a circular dependency. No step in the derivation chain reduces, by definition or by construction, to its own input.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The method introduces no new entities and uses one blending weight and one layer count as hand-picked parameters. Its central assumption about layer-wise appearance/geometry separation is a domain assumption validated only by small ablations. The pretrained InstantMesh pipeline is taken as an external input.

free parameters (2)
  • alpha (style blending weight) = 0.8
    Chosen empirically in Section 3.2 and Fig. 5 to balance style transfer and content fidelity; not derived from first principles.
  • Number of style-injected transformer layers = 4 (last 4 layers)
    Selected based on Chamfer Distance ablations in Table 1 and Fig. 4; the layer count balances stylization strength and geometry preservation.
assumptions (3)
  • domain assumption Later transformer blocks of the LRM encode appearance-specific features while earlier blocks encode geometry.
    Stated as a hypothesis in Section 3.2; supported only by qualitative Fig. 4 and CD values in Table 1 on a small set of objects. If false, layer selection for style injection is unjustified.
  • domain assumption Style images encoded by the same ViT encoder are compatible with the multi-view image token distribution for cross-attention.
    The method relies on the pretrained ViT to embed arbitrary style images; no alignment or fine-tuning is performed. Distribution gap between artwork and object renderings could limit style transfer.
  • domain assumption InstantMesh's pretrained reconstructor is a valid LRM with the claimed attention structure.
    The paper builds on InstantMesh (Xu et al. 2024a) and inherits its reconstruction quality; if the underlying LRM fails to reconstruct certain objects, stylization quality will suffer (acknowledged in Limitations).

how reviews work

0 comments
Cite this review

Pith. "Pith review of 3D Stylization via Large Reconstruction Model." pith.science (2026). https://pith.science/paper/HIJMRZKX

@misc{pith2026250421836,
  author       = {Pith},
  title        = {Pith review of: 3D Stylization via Large Reconstruction Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HIJMRZKX}},
  note         = {Machine review of arXiv:2504.21836}
}
read the original abstract

With the growing success of text or image guided 3D generators, users demand more control over the generation process, appearance stylization being one of them. Given a reference image, this requires adapting the appearance of a generated 3D asset to reflect the visual style of the reference while maintaining visual consistency from multiple viewpoints. To tackle this problem, we draw inspiration from the success of 2D stylization methods that leverage the attention mechanisms in large image generation models to capture and transfer visual style. In particular, we probe if large reconstruction models, commonly used in the context of 3D generation, has a similar capability. We discover that the certain attention blocks in these models capture the appearance specific features. By injecting features from a visual style image to such blocks, we develop a simple yet effective 3D appearance stylization method. Our method does not require training or test time optimization. Through both quantitative and qualitative evaluations, we demonstrate that our approach achieves superior results in terms of 3D appearance stylization, significantly improving efficiency while maintaining high-quality visual outcomes.

Figures

Figures reproduced from arXiv: 2504.21836 by the authors.

Figure 1
Figure 1. Given a 3D neural radiance field reconstructed by a large reconstruction model (top row), our approach enables the visual stylization based on a target [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. InstantMesh provides a two stage pipeline for single-image based [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of our method. We build upon InstantMesh which takes an input image of an object and generates 6 sparse view outputs, which are then [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Qualitative results for using style image ViT-encoded embeddings in the cross-attention layers across the last 1, 2, 4, 6, 8, 10, and all 16 layers of the [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Our method blends the cross attention output with respect to original [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Visual results of our method and competing methods. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: We compare our results to alternative design choices. First, we con [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: We provide more visual results as style image, 3D object, and stylization triplets. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: We provide visual results of additional competing methods. [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Example for limitation SIGGRAPH Conference Papers ’25, August 10–14, 2025, Vancouver, BC, Canada [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

5 extracted references · 5 linked inside Pith

  1. [2]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Rave: Randomized noise shuffling for fast and consistent video editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 6507–6516. Áron Samuel Kovács, Pedro Hermosilla, and Renata G Raidou. 2024. Surface-aware Mesh Texture Synthesis with Pre-trained 2D CNNs. In Computer Graphics Forum, Vol. 43. W...

    arXiv 2024

  2. [4]

    arXiv preprint arXiv:2404.02733 (2024)

    Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733 (2024). Zhouxia Wang, Xintao Wang, Liangbin Xie, Zhongang Qi, Ying Shan, Wenping Wang, and Ping Luo. 2023. Styleadapter: A single-pass lora-free model for stylized image generation. arXiv preprint arXiv:2309.01770 (2023). Xinyue Wei, Kai Zhang, S...

    arXiv 2024

  3. [2022]

    In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10684–10695. Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. 2023. Dreambooth: Fine tuning text-to-image diffusion models for subject- driven generation. In Proce...

    arXiv 2023

  4. [2024]

    In ACM SIGGRAPH 2024 Conference Papers

    Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers. 1–12. Anand Bhattad, Aysegul Dundar, Guilin Liu, Andrew Tao, and Bryan Catanzaro. 2021. View generalization for single image textured 3d models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 6081–6090. Duygu Ceylan, Val...

    arXiv 2024

  5. [2025]

    arXiv preprint arXiv:2504.02762 (2025)

    MD-ProjTex: Texturing 3D Shapes with Multi-Diffusion Projection. arXiv preprint arXiv:2504.02762 (2025). Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. 2024. ViewCrafter: Taming Video Diffusion Models for High-fidelity Novel View Synthesis. arXiv preprint arXiv:2409.02048 (20...

    arXiv 2025

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.