acceptodds
Withdrawn from ICLR 2027

Disentangling Texture and Color for Controllable 3D Stylization and Morphing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Separate texture and color references do not guarantee separate effects in a pretrained 3D generator: both attributes share a nonlinear decoder, and texture editing changes the state on which color acts. We formulate selective control as carrier-conditioned editing in TRELLIS structured latent space. Detail-oriented conditioning first constructs a textured carrier on the source's sparse support. Color is then fitted separately to the source and textured carriers, with preservation defined relative to each carrier. The four fitted endpoints define a two-attribute latent surface whose interaction term adapts the color direction to texture strength and exactly recovers all four corners. No pretrained weights are updated. We completed 120 balanced content–texture–palette triplets and exported 480 endpoint GLBs. A matched ten-triplet ablation measures the texture-transfer/content-preservation trade-off, and decoded control grids expose the effect of the interaction term. These results support endpoint generation and inspectable two-axis control, while matched external baselines and reference-based interior evaluations remain necessary to establish comparative superiority and independent decoded effects.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Controllable 3D Stylization, Texture–Color Disentanglement, Carrier-Conditioned Latent Editing, Interaction-Aware Style Morphing

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_r63uOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This work achieves 3D stylization with decoupled texture and color, based on the pre-trained 3D generative model TRELLIS. The authors frame this task as an editing process within the latent space of TRELLIS, where texture information is injected into the source structure first, followed by the separate injection of color information. This allows for independent control over the intensity of both texture and color.

Critical strengths
  1. The examples shown in Figure 3, along with most of the metrics presented in Table 1, demonstrate strong performance.
  2. The approach yields high-quality results without requiring updates to the pre-trained weights, making it both simple and effective.
Critical weaknesses
  1. The number and variety of results provided are limited; the method should be validated on a broader range of data.
  2. Why does the method perform significantly worse than all other approaches in terms of the SSIM metric (Table 1)?
  3. Although this is a 3D-based method, the paper only presents 2D frontal views, making it impossible to assess the actual 3D results.
Questions for authors
  1. What is the computational efficiency of this method?
Reviewer_giZHOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper argues that, when restyling a 3D object in TRELLIS, the effect of a color edit depends on the object’s current texture. Instead of reusing the color edit on the source (F00 to F01) for the textured object (F10) to obtain the colored, textured object (F11), they fit F11 directly from F10, which they call the texture carrier. After constructing the four resulting endpoints (F00, F01, F10, and F11), they use bilinear interpolation for continuous control. The contributions are (1) carrier-conditioned color fitting that preserves the details of the underlying carrier, and (2) combining the four endpoints by standard bilinear interpolation, where they interpret the cross term as how the color edit changes with texture. They generated the four endpoints for 120 triplets and used 10 triplets of the same object for quantitative results and ablations. They compared with other methods.

Critical strengths

Fitting the color edit while conditioned on the underlying object state is a well-motivated way to model the color-texture dependency. Since the decoder is nonlinear, the same color edit can give a different output when the starting texture is different. Although the method requires per-sample latent optimization, it is training-free.

Critical weaknesses

It is unclear whether the method actually improves the results over reusing the color edit and the baselines. The main argument is that fitting endpoints separately (the main contribution) is better than reusing the color edit from F00 to F01 (L219-222). The reused version is the K = 0 corner in Equation 9. Table 4 compares K = 0 with K != 0, but its metrics show how much the outputs change, not which one is closer to the references (L901-903). Against the baselines, the paper claims better results (L300-309), although the baseline results are from a different evaluation set (L081-082). In addition, the baselines seem to consider how the colors are assigned to the object’s parts (Figure 3), whereas the method optimizes F11 toward the color distribution of the palette strips (L203-215). However, the claim of better color performance mainly relies on MS-SWD and C-Hist (L302-304), which measure the overall color distribution (L272-274). Therefore, these metrics show whether the color distribution matches the palette, but not how the colors are assigned to the object’s parts. Overall, neither comparison shows whether the results are better or merely different.

Questions for authors
  1. Can you evaluate the reused version (the K = 0 corner) with the same reference-based metrics as Table 1, on the same inputs?
  2. Can you report the baselines on the same inputs as the method?
  3. Can you include a color evaluation that also measures how the colors are assigned to the object’s parts?
Reviewer_3wd9OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies separate texture and color editing in the structured latent space of a frozen TRELLIS model. It first constructs source and texture-edited carriers, then independently fits a color edit around each carrier using carrier-relative preservation constraints.

The resulting source, texture-only, color-only, and jointly edited endpoints define a two-dimensional latent control surface. An interaction term captures the difference between the two fitted color displacements, making the interpolated color direction depend on texture strength.

The authors report generation of 120 input triplets and 480 endpoint assets. Quantitative evaluation is restricted to ten triplets involving one content object, while the main quantitative interaction analysis uses a single fireplace example.

Critical strengths
  • The paper identifies a meaningful limitation of additive latent editing: a color displacement fitted around the source need not retain the same decoded effect after texture editing. Carrier-relative fitting is a reasonable approach to accommodating this dependence on the decoder’s operating point.

  • The four-endpoint construction provides an intuitive interface for continuous texture and color control. Once the endpoints are fitted, intermediate states can be produced without further optimization, although decoding remains necessary.

  • The manuscript explicitly distinguishes shared latent support from fixed decoded geometry and acknowledges that continuous latent interpolation does not guarantee perceptual independence or monotonicity. These qualifications appropriately identify the limits of what the construction establishes.

Critical weaknesses
  • The interaction-aware surface is algebraically equivalent to standard bilinear interpolation of four fitted endpoints. Exact recovery of those endpoints follows by construction and does not establish disentanglement, independent decoded effects, or improved interpolation quality.

  • The reported cross-method results do not support a controlled superiority claim. Appendix D acknowledges different evaluation samples and incompletely matched implementations, yet Section 4.2 interprets the numerical rankings as evidence of improved performance; the restriction of the main quantitative evaluation to one content object further limits generalization.

  • The experiments omit the comparisons most directly needed to test carrier-conditioned editing. Shared-residual transfer, direct color assignment, and geometry-fixed Gaussian-color optimization should be evaluated at comparable color attainment, while the interaction ablation changes the joint endpoint and therefore does not isolate interior control quality.

  • The method is insufficiently specified for reproduction from the manuscript. Appendix B.3 acknowledges missing definitions for several normalization, smoothing, projection, palette-assignment, and loss operations, while the VLM evaluation lacks a sufficiently specified scoring protocol.

Questions for authors
  1. Beyond global palette transfer, can the method support spatially varying albedo editing, such as changing the base color of a specific region while preserving texture, geometry, and illumination-dependent shading?

  2. Can the method handle transparent or translucent objects, or edit material properties such as transparency? What limitations does the underlying representation impose on these edits?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.