- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
This paper presents a method called Backdrop which aims to provide 3D stylization via a full-frame texture images. The authors make the observations that in SoTA 3D generation methods such as Trellis 2, while object reference images work, full texture images result in artifacts. They observe that this is because the background patches in object reference images observe substantial cross attention, and without this attention destablizes geometry.
Backdrop appends additional patch tokens from a blank image to overcome this challenge. It's training free and the method is shown on top of Trellis 2 with Style Sculptor and MorpAny 3D.
Critical strengths
- Interesting observation that the redundant background tokens play a crucial role for stylization. The paper does various experiments (e.g., removing background tokens from object references also result in artifacts) to show this.
- The method is training free and is adapted to both Style Sculptor and Morph Anyhting 3D.
Critical weaknesses
-
Important closely related work is missing: Oztas et al., 3D Stylization via Large Reconstruction Model, SIGGRAPH 2025. That work already demonstrates training-free 3D appearance stylization by injecting reference-image features into selected cross-attention layers of a pretrained reconstruction model. It also examines how layer selection affects geometry preservation. This work is directly relevant to the attention-blending baseline and the broader claim of enabling reference-guided stylization in frozen 3D models.
-
The pipeline is shown successfully on top of a single generator Trellis 2. The experiments on Trellis show that is behaviour is different and Backdrop may not be as necessary. And for Hunyuan3D, the proposed solution is not as effective. While these are interesting observations, it also means the solution is not general and more architecture specific.
-
The authors already mention that addition of blank image tokens, in a way, is adjusting the attention scale of the reference tokens. The zero-value-token variant and fixed attention scaling ablations already are effective. This finding is also related to the above mentioned related work, Oztas et al. Hence, the tecnhical contribution is not justified.
-
The evaluation section would benefit from choosing metrics that separately evaluate stylization quality, geometric quality.
Questions for authors
-
How does your attention-blending baseline relate to Oztas et al.’s work? Would restricting style injection to selected layers, as studied there, mitigate collapse on TRELLIS.2 without appended tokens?
-
How sensitive are the results to the number of appended tokens? Can a substantially smaller subset retain the benefits, and how does the required number relate to reference composition or style strength?
-Which architectural properties determine whether BACKDROP is needed and effective? Can you demonstrate a successful repair on another generator, rather than only an attention-allocation similarity or a different unresolved failure?
- Can you evaluate geometric-style transfer separately from texture and color transfer, ideally with human judgments?
5 comments
W3: Technical contribution.
We respectfully disagree that the technical contribution is unjustified. The proposed method is derived from a causal investigation of geometric collapse, rather than an arbitrary adjustment of attention strength. Our experiments distinguish BACKDROP from the suggested alternatives:
- Zero-value tokens (Table 5) still require BACKDROP's blank-image keys to compete for attention. This validates our mechanism rather than replacing it. Although this variant achieves comparable structural integrity, it requires modifying the internal cross-attention operations, whereas BACKDROP works directly at the conditioning interface with a single line of code.
- Fixed attention scaling (Table 6) uses factors measured from BACKDROP's own attention distributions, yet still produces 28 collapsed outputs versus only 2 with BACKDROP.
- Global style-weight adjustment (Table 10) cannot achieve the same combination of strong stylization and geometric integrity under the evaluated settings.
Crucially, BACKDROP enables query-dependent attention redistribution through the generator's own softmax, naturally adapting across queries, heads, blocks, and sampling steps without additional learned parameters or per-instance tuning. Our contribution therefore includes both identifying and causally validating a previously overlooked conditioning failure and developing a simple, effective intervention that resolves it. The one-line implementation reflects this mechanistic insight, rather than a lack of technical contribution.
W4, Q4: Separate evaluation of geometric stylization.
Our existing evaluation already distinguishes three objectives: style similarity (CSD), structural integrity (SIS), and content-shape preservation (F@2%). To further isolate geometric-style transfer from texture and color, we conducted additional evaluations using untextured surface-normal renders across three stylization methods.
Using a VGG-16 patch-statistics score normalized relative to the unstylized base (0) and reference (1), BACKDROP achieves 0.25 for attention blending and 0.28 for StyleSculptor, improving by +0.097 [0.055, 0.139] and +0.127 [0.083, 0.171], respectively, over their unaugmented counterparts.
CSD computed on the same normal renders independently supports these findings. All three methods with BACKDROP outperform the unstylized base, while attention blending and StyleSculptor also significantly outperform their corresponding unaugmented versions, by +0.028 [0.005, 0.054] and +0.022 [0.0003, 0.048], respectively. For MorphAny3D, the improvement over its unaugmented version remains statistically unresolved, consistent with the original CSD evaluation.
These additional experiments provide quantitative evidence that BACKDROP supports geometric-style transfer beyond appearance similarity and structural preservation, directly addressing the concern about insufficient separation of evaluation objectives. A human study could not be completed within the discussion period due to the required ethics approval.
Q2: Token-count sensitivity.
The choice of 1,024 tokens was not tuned; it corresponds to the complete 32 × 32 patch grid of a blank image at the encoder's native 512-pixel resolution. We additionally evaluated blank-image encodings producing 64–2,304 tokens (full table in our response to Reviewer enmS).
The results reveal a clear relationship between attention competition and structural integrity. At α = 0.85, where 418/432 outputs collapse without BACKDROP, 576 tokens already reduce collapses to 14/432, while 1,024 tokens reduce them to only 2/432. Increasing the count to 2,304 does not improve reliability (14/432 collapses). The required count also depends on style strength: at α = 0.55, only 256 tokens already reduce collapses to 6/432, versus 0/432 with 1,024 tokens. Reference composition also matters: at 576 tokens, collapses occur in 6/96 artistic, 8/276 texture, and 0/60 scene pairs.
These results directly validate the attention-competition mechanism and demonstrate that the original, untuned 1,024-token choice provides the strongest structural reliability among the tested configurations, while maintaining strong style transfer.
Dear Reviewer dqbs, as the discussion period ends soon, we'd be grateful if you could check whether our response addresses your concerns. In particular, we added Öztaş et al., the layer-selection experiment you suggested, and a geometric-style evaluation. Thank you!
Thank you for the review and for highlighting Öztaş et al. We would like to clarify several important distinctions regarding our contributions and provide additional experimental evidence addressing the concerns. We have uploaded a revised PDF with changes marked in blue.
W1, Q1: Related work and layer-selective injection.
We have included Öztaş et al. in the revision (Section 2). However, their work pursues a different technical objective from BACKDROP. They propose a stylization method that injects reference features into the last 4 of 16 cross-attention blocks to transfer appearance while preserving geometry. In contrast, BACKDROP is not another stylization method. Our contribution is identifying why existing methods fail with non-object references and resolving this failure to enable both appearance and geometric stylization.
Our attention-blending baseline follows the same weighted combination of content and reference cross-attention outputs as Eq. 2 of Öztaş et al., but applies reference conditioning throughout the shape blocks to enable geometric as well as appearance stylization.
To directly investigate the suggested alternative, we conducted new layer-selection experiments on TRELLIS.2. Without BACKDROP, restricting reference injection from all shape blocks to the late half and final quarter reduces collapses from 418/432 to 226/432 and 30/432, respectively. However, this substantially weakens stylization: the final-quarter setting achieves CSD 0.27 versus 0.33 with BACKDROP, and geometric-style similarity 0.12 versus 0.25. BACKDROP preserves 430 of 432 outputs while allowing reference injection throughout the network, achieving stronger appearance and geometric stylization without sacrificing structural integrity. These results directly demonstrate why restricting injection layers does not replace our contribution.
W2, Q3: Applicability beyond TRELLIS.2.
Our cross-backbone experiments provide evidence that the importance of background conditioning extends beyond TRELLIS.2. On Hunyuan3D 2.1, deeper blocks also allocate substantial attention to background tokens (Fig. 8), and removing these tokens during direct generation degrades geometry, including structural failures (Fig. 12). This independently confirms that background conditioning influences the generated geometry across different 3D generators.
Our analysis suggests two conditions under which BACKDROP is particularly relevant: (1) the generator's deeper blocks allocate substantial attention to background tokens, and (2) the underlying stylization method preserves content geometry when using object references. These conditions explain the observed architectural differences. TRELLIS (v1) predominantly attends to global CLS and register tokens, which remain available for non-object references, and consequently does not exhibit the same collapse. On Hunyuan3D 2.1, the tested stylization implementation does not reliably preserve content geometry even with object references, preventing a direct demonstration of stylization repair in that setting.
These findings establish that the underlying background-attention phenomenon is not unique to TRELLIS.2. BACKDROP provides a simple, training-free intervention motivated by this broader mechanism, with its effectiveness demonstrated across three stylization methods and diverse reference types on TRELLIS.2. We have clarified this experimental scope in the revision by removing the overgeneralized wording "and generators" from Figure 1 and the TL;DR, and explicitly identifying TRELLIS.2 in the abstract.