Stylos: Multi-View 3D Stylization with Single-Forward Gaussian Splatting

arXiv:2509.26455 · cs.CV · Submitted 2025-09-30 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Stylos: Multi-View 3D Stylization with Single-Forward Gaussian Splatting".

Jane: Here's the gist of the paper. Stylos presents a single-forward 3D Gaussian framework for 3D style transfer that operates on unposed content from a single image to a multi-view collection, conditioned on a separate reference style image.

Tom: This method achieves geometry-aware, view-consistent stylization without requiring per-scene optimization or precomputed poses.

Jane: Framework and Architecture Stylos adopts a shared Transformer backbone with two pathways: geometry predictions retain self-attention to preserve geometric fidelity, while style is injected via cross-attention to enforce visual consistency across views.

Tom: The content tokens retain self-attention for geometric reasoning, and style is conditioned on the target style through CrossBlock modules that integrate a cross-attention layer between self-attention and MLP stages.

Title and authors: Tom: The paper we're discussing is titled Stylos: Multi-View three dee Stylization with Single-Forward Gaussian Splatting <ref:2509.26455#pg1,Stylos: Multi-View 3D Stylization with Single-Forward Gaussian Splatting>. It addresses how to take content from a single image and project it into multiple views while applying a style from another image.

Jane: So, the main goal here is to create a method that can do this three dee style transfer without needing any precomputed camera poses for each scene, which sounds like it solves a major headache in three dee computer vision <ref:2509.26455#pg1>.

Lu: The authors are Liu, Huang, Mi, Lu, Saripalli and Peng Jiang from Texas A andM University. They are introducing this new framework to handle the complexity of three dee stylization in a single forward pass pipeline without scene specific training or post-optimization <ref:2509.26455#pg1>.

Meng: So for someone trying to put this into production, it’s important that it's a single forward pass, right? We don't want multi stage inference just because it adds latency.

Tom: That's exactly what they are aiming for; a single forward pass pipeline means you get the stylized result quickly without needing those extra steps.

Jane: What’s the big takeaway for us listeners? It’s that we can achieve geometry aware and view consistent stylization just by using a single image as content and another image as style reference.

Lu: And they achieve this by proposing a shared backbone design with two pathways: one pathway handles geometry predictions keeping self attention to preserve geometric fidelity, while the other pathway injects style via cross attention for visual consistency across views.

Meng: So it's like they’re essentially building two specialized parts inside the same Transformer architecture to handle structure and appearance separately.

Tom: Right, and they also introduce a voxel based three dee style loss that aligns aggregated scene features with style statistics, which is key for enforcing that view consistent stylization <ref:2509.26455#pg1,a voxel based 3D style loss that aligns aggregated scene features>.

Jane: So we’ve seen how this works—the authors propose a voxel based three dee style loss that aligns aggregated scene features with style statistics, providing stronger view-consistent stylization while preserving geometric fidelity <ref:2509.26455#pg1,a voxel based 3D style loss that aligns aggregated scene features>.

Lu: That loss function is what ties the geometry and the appearance together in a way that is specifically tailored for multi-view data.

Meng: I wonder how stable that voxel based fusion is when we're dealing with real world scenes where things are very complex.

The paper's summary: Tom: Stylos synthesizes a stylized three dimensional Gaussian scene without doing per scene optimization or precomputed poses, which is the core promise of the paper.

Jane: So what does that mean for us in practical terms? It means we can take a single piece of content and generate stylized three dee results instantly without having to worry about complex setup beforehand <ref:2509.26455#pg1>.

Lu: The authors introduce Stylos as a single-forward-pass pipeline for three dimensional style transfer that operates on unposed content, from one image to a multi view collection, conditioned on a separate reference style image.

Meng: That sounds like it eliminates the need for us to precompute poses for every scene we want to stylize. That would save us a lot of time in the pipeline setup phase.

Tom: Yes, that’s the point; you feed it one content image and one style image, and you get a multi-view collection right away.

Jane: So, this means we can achieve high quality zero shot stylization that generalizes to unseen categories, scenes, and styles without having to retrain anything specific for each new thing.

Lu: The main contributions are three things: they propose the shared backbone design with two pathways for geometry and style injection, introduce the voxel based three dee style loss that aligns aggregated scene features with style statistics, and develop Stylos as a single forward-pass pipeline <ref:2509.26455#pg1,voxel based 3D style loss that aligns aggregated scene features>.

Meng: So it’s not just one clever trick; it’s a combination of architecture, a new loss function, and the pipeline design that makes it work together.

The paper's improvements: Tom: The improvements they propose focus on how this framework can be used to handle different types of style application more effectively than traditional methods.

Jane: What are they actually doing better compared to the older methods? Is it about the quality or something else?

Lu: They move away from classical style transfer which relies on Gram matrix correlations, and instead focus on channel wise statistics like mean and variance, which are more efficient alternatives to things like AdaIN.

Meng: So they're prioritizing efficiency in terms of speed and resource usage over just chasing the absolute peak PSNR score?

Tom: They are aiming for an efficient alternative that still delivers high visual fidelity by using these channel wise statistics for the style injection.

Jane: That’s smart, especially if we think about running this on hardware where we need to be mindful of how much computation we're doing per frame.

Lu: And the voxel level three dee style loss is what really makes a difference in terms of enforcing cross view coherence and geometry aware stylization <ref:2509.26455#pg1>.

Meng: So that specific loss function is what allows it to produce sharper boundaries and a stronger sense of three dee geometry compared to image level or scene level losses <ref:2509.26455#pg1>.

Conclusion: Tom: To wrap up, Stylos is a single forward pass pipeline for three dimensional style transfer that operates on unposed content, from one image to a multi view collection, conditioned on a separate reference style image.

Jane: So the main implication is that we can get geometry aware and view consistent stylization without requiring per scene optimization or precomputed poses.

Lu: The voxel-space three dee style loss effectively aggregates multi-view features to enforce strong cross view coherence through aligning those aggregated scene features with style statistics <ref:2509.26455#pg1>.

Meng: And this means the system can scale from a single view to hundreds of views with a single style image, achieving zero shot generalization to unseen categories, scenes, and styles.

Lalam: We can also interpolate between embeddings of two distinct style images or between a reconstruction embedding and a stylized embedding for smooth transitions.

Tom: This whole paper is about achieving a favorable balance between visual quality, consistency, and efficiency for practical real time three dee stylization <ref:2509.26455#pg1>.

Jane: It seems like Stylos offers a way to do this by keeping geometry predictions on the self attentive backbone path and conditioning appearance via global cross blocks on the style path.

Lu: And it’s a framework that is scalable and generalizes effectively to novel categories, scenes, and styles.

Texas A&M University

cs.CV

Submitted: 2025-09-30

Updated: 2026-11-05

Comments: Accepted by ICLR 2026

Journal ref: Proceedings of the International Conference on Learning Representations (ICLR) 2026, pp. 70768-70801

Code: https://github.com/HanzhouLiu/Stylos

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 88/100

The gist: Stylos presents a single-forward 3D Gaussian framework for 3D style transfer that operates on unposed content from a single image to a multi-view collection…

Key concepts

Shared Transformer Backbone
This is the core neural network structure used to process both the 3D geometry and the style information. It has two main paths: one path handles geometric predictions using self-attention, while another path manages style injection through cross-attention, ensuring both aspects are processed simultaneously.
Voxel-level 3D Style Loss
This is a specific loss function that fuses features from all input views into a 3D voxel grid. By calculating style statistics directly in this voxel space, it enforces style consistency across different views and the underlying 3D structure, leading to sharper boundaries and better geometry preservation.
CrossBlock Module
This module is used to inject the target style information into the content features. It integrates a cross-attention layer between self-attention and MLP stages, which significantly boosts reconstruction accuracy compared to simpler frame attention variants when enforcing visual consistency across views.

Terminology

Summary

Stylos presents a single-forward 3D Gaussian framework for 3D style transfer that operates on unposed content from a single image to a multi-view collection, conditioned on a separate reference style image. p. 0 This method achieves geometry-aware, view-consistent stylization without requiring per-scene optimization or precomputed poses. p. 0

Framework and Architecture

Stylos adopts a shared Transformer backbone with two pathways: geometry predictions retain self-attention to preserve geometric fidelity, while style is injected via cross-attention to enforce visual consistency across views. The content tokens retain self-attention for geometric reasoning, and style is conditioned on the target style through CrossBlock modules that integrate a cross-attention layer between self-attention and MLP stages. p. 0 The geometry backbone follows the alternating-attention design of VGGT, which interleaves frame and global self-attention layers to capture both intra-frame structure and cross-view consistency. p. 3

Training Strategy

The training strategy is divided into two stages:

  1. Geometry Pretraining, where the geometric backbone is initialized with VGGT weights and trained end-to-end to learn geometry and photometric appearance. p. 4 One input view is randomly selected and color-jittered as the style reference to improve robustness. p. 4

  2. Stylization Fine-tuning, where only the Style Aggregator and the color head are updated, utilizing feature-level style and content losses in VGG space. p. 4 This stage incorporates a 3D voxel-space style loss for cross-view consistency and a CLIP-based loss for semantic alignment. p. 4

Style Loss Functions

Stylos employs three distinct style losses to enforce consistency across different levels of representation:

  1. Image-Level Style Loss, which aligns each rendered frame independently with the style reference using mean and standard deviation statistics. p. 4

  2. Scene-Level Style Loss, which concatenates per-view features into a map and calculates statistics on this aggregated map to promote global style coherence across views. p. 4

  3. Voxel-level 3D Style Loss, which fuses multi-view features into a voxel grid using differentiable unprojection and computes style statistics directly in voxel space to enforce style consistency across both views and the underlying 3D structure. p. 5

Experimental Results

Experiments demonstrate that Stylos delivers high-quality zero-shot stylization, showing strong cross-view consistency scores ranking first across all consistency metrics and all four scenes. Stylos attains either the best or second-best artistic metric values on all Tanks and Temples scenes while maintaining the fastest stylization speed. p. 7 p. 7 The 3D style loss produces sharper boundaries and a stronger sense of 3D geometry, leading to the most coherent multi-view stylization. p. 6 The method successfully transfers diverse artistic styles to scenes while preserving content geometry. p. 9

Limitations and Analysis

Stylos tends to underperform on scenes dominated by high-frequency or highly cluttered structures, such as dense foliage, thin branches, or wire-like elements. Performance decreases as the number of input views increases due to limitations of the VGGT backbone in maintaining geometric reconstruction stability. p. 1 Failures can be categorized into inaccurate scene reconstruction, generating colors absent from style images, or introducing blurry or over-smoothed style colors. p. 4 The global CrossBlock design for style injection significantly boosts reconstruction accuracy compared to frame attention variants. p. 5

Controllability and Scaling

The model supports controllable stylization through interpolation experiments, allowing smooth transitions between two distinct style images or between the content and a stylized embedding. The efficiency scaling shows that while very small batches are unstable, runtime becomes much more stable once the batch size exceeds 4–8 views. p. 6 The framework scales from a single view to hundreds of views with a single style image, achieving zero-shot generalization to unseen categories, scenes, and styles. p. 1

Conclusion

In this work, Stylos achieves geometry-aware and view-consistent stylization by keeping geometry predictions on the self-attentive backbone path and conditioning appearance via global CrossBlocks on the style path. The voxel-space 3D style loss effectively aggregates multi-view features to enforce strong cross-view coherence. p. 9 Stylos demonstrates a favorable balance between visual quality, consistency, and efficiency for practical real-time 3D stylization. p. 7 The proposed framework is scalable and generalizes effectively to novel categories, scenes, and styles. p. 0

Acknowledgement

This research used both the DeltaAI advanced computing and data resource, which is supported by the National Science Foundation (award OAC 2320345) and the State of Illinois, and the Delta advanced computing and data resource which is supported by the National Science Foundation (award OAC 2005572) and the State of Illinois.

Reproducibility

The main paper describes all model architectures in Sec. 3, provides detailed pseudo-code for the proposed 3D style loss function in Algo. p. 0 1, and will release the full implementation upon acceptance. p. 5 All datasets used are publicly available as described in Sec. p. 7 4. The paper provides extensive qualitative results in Appendix A.6 and analyzes efficiency trends with respect to the number of views per batch in Appendix A. p. 62. The authors follow the original authors’ training setups and inference pipelines for fair comparisons.

References

The paper cites numerous related works, including StyleGaussian (Liu et al., 2024), G-Style (Kovacs et al. p. 5, 2024), StylizedGS (Zhang et al. p. 5, 2025), and SGSST (Galerne et al. p. 5, 2025). Other cited works include NeRF (Mildenhall et al. p. 0, 2020), 3D Gaussian Splatting (Kerbl et al. p. 0, 2023), and VGGT (Wang et al. p. 4, 2025a). The paper also references classical style transfer methods like Gatys et al. p. 2 (2016) and various image-level and scene-level losses. p. 6 The evaluation uses metrics such as PSNR, SSIM, LPIPS, ArtScore, and ArtFID <ref:2509. p. 526455#pg11>. The paper also references datasets like CO3D (Reizenstein et al. p. 5, 2021) and DL3DV-10K (Ling et al. p. 5, 2024). The paper's contributions are summarized in the main text <ref:2509.26455#pg2>. The references section lists all cited literature, including works on pose-free 3D reconstruction and 3D style transfer losses. p. 1 The authors release model weights to support reproducibility. p. 6 The final section provides a list of references. p. 5

Appendix Details

Implementation details specify the use of VGGT as the backbone with L = 24 Alternating-Attention Transformer layers for geometry and depth prediction. The voxelization step uses a voxel size of 0. p. 4002 for differentiable voxelization. p. 4 The loss weights are set as style weight λstyle = 1. p. 40, content-reconstruction weight λcnt = 0. p. 21, CLIP consistency weight λclip = 1. p. 40, and total-variance regularization weight λtv = 10. p. 40. Efficiency analysis shows runtime grows roughly linearly with respect to batch size once the batch size exceeds 4–8 views. p. 6 The paper also details specific failure cases related to geometric reconstruction and color generation. p. 8 The authors release implementations of all three style losses to facilitate further exploration by the community. p. 4 The paper also provides extensive qualitative results in Appendix A.6. All 3DGS-based methods are evaluated on full-resolution images, while single-forward methods are evaluated at their preset input resolutions. p. 1 The paper also reports supplementary quantitative results in Table 6. p. 8 p. 8 The authors follow the original authors’ training setups and inference pipelines for fair comparisons. The paper also reports supplementary quantitative results in Table 3 and Table 4. p. 8 p. 8

Figure Analysis

Figure 1 shows the architecture overview, illustrating the core CrossBlock module that facilitates style injection by integrating a cross-attention layer between self-attention and MLP. Figure 2 compares different CrossBlock designs on the CO3D pizza scene, showing that Global CrossBlock yields the most faithful result. p. 6 Figure 8 visually confirms that the 3D style loss provides a stronger sense of 3D geometry and cleaner wooden floor textures compared to other losses. p. 7 Figure 11 illustrates failures caused by inaccurate scene reconstruction, such as missing geometric details or incorrect region colors in the multi-view results. p. 9 Figure 12 shows failures caused by generating colors that are absent from the original style images, such as a pure white region hallucination. p. 9 The paper presents additional visual results across diverse style–content pairs on Tanks and Temples datasets. p. 9 The paper also presents additional visual results on the CO3D dataset. p. 7 The paper also presents additional visual results on the DL3DV-10K dataset. p. 5 The paper also presents additional visual results on the CO3D skateboard scene. p. 7 Stylos attains either the best or second-best artistic metric values on all Tanks and Temples scenes while maintaining the fastest stylization speed. Table 3 shows Stylos demonstrates improved short-range and long-range consistency scores across the four scenes compared to baselines. p. 8 Table 4 shows Stylos achieves consistently favorable metric scores across the four scenes, indicating improved cross-view stylization consistency. p. 8 The paper's quantitative evaluation confirms that Stylos provides markedly improved cross-view stylization consistency. The authors follow the original authors’ training setups and inference pipelines for fair comparisons. The paper also reports supplementary quantitative results in Table 3 and Table 4. The paper also reports supplementary quantitative results in Table 6. The authors follow the original authors’ training setups and inference pipelines for fair comparisons.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI system can perform single-forward 3D style transfer from unposed inputs, scaling from a single view to hundreds of views with a single style image, achieving zero-shot generalization to unseen categories, scenes, and styles.

  2. The system achieves geometry aware and view consistent stylization by employing a shared Transformer backbone with two pathways: geometry predictions retain self attention for geometric reasoning while style is injected via cross attention.

  3. The system enforces view consistent stylization while maintaining geometric coherence through the introduction of a voxel-based 3D style loss that aligns aggregated scene features to style statistics, specifically the equation L3D sty (5).

  4. The system can perform controllable stylization by linearly interpolating between embeddings of two distinct style images or between a reconstruction embedding and a stylized embedding to produce a continuous spectrum of outputs with gradually increasing stylization strength.

  5. The system can distinguish between different style loss designs, showing that the 3D style loss produces sharper boundaries and a stronger sense of 3D geometry compared to image-level matching or scene-level losses.

Sources

Related papers