acceptodds
Under review as a conference paper at ICLR 2027

CrediTex: View-Level Credit Assignment for Human-Aligned 3D Texturing

Abstract

Despite advances in joint multi-view texture generation, individual views can still contain artifacts, residual illumination, and implausible details. A promising approach is to use post-training with human preference feedback, but existing asset-level rewards obscure quality differences across views and provide the same feedback to views of varying quality. To address this issue, we introduce CrediTex, a post-training framework for 3D texture generation with context-aware per-view rewards and fine-grained credit assignment. First, we construct a dataset of 48K human preference pairs between corresponding views to train the reward model. Given a target-view query, the reward model predicts a continuous scalar reward for the specified view using the reference image and complete view set as context. Subsequently, we compute relative advantages from these rewards across candidates at each fixed viewpoint to guide view-level reinforcement learning. These advantages provide distinct feedback to individual views of the same 3D asset, preserving view-level credit assignment throughout policy optimization. Experiments across multiple models and evaluation metrics demonstrate state-of-the-art performance in both reward modeling and 3D texture generation.

Primary area:
generative models
Keywords:
3D texture generation, multi-view diffusion, view-level credit assignment, reward model, preference alignment

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_yh4tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses a meaningful problem in the task of 3D texture generation by proposing a two-stage post-training framework. Specifically, existing post-training methods for 3D texture generation use 3D-asset-level rewards/signals to improve the texture of a 3D asset. However, the authors argue that this may not be the best way to reward texture quality because, in a view-by-view texture generation approach, each painted view may have a different quality, and rewarding the entire 3D asset cannot capture these differences. To address this issue, the authors propose evaluating the texture by assigning a credit/reward to each individual viewpoint that contributes to the final texture. To this end, they first construct a dataset of human preference pairs between corresponding views to train a reward model. Then, using this trained reward model, they leverage an existing online post-training approach, DiffusionNFT, to guide the policy model (the texture generation model) toward generating textures that achieve higher scores according to the view-level reward model. Through quantitative and qualitative experiments as well as a user preference study, they demonstrate the effectiveness of their approach and show that the proposed post-training framework can generate textures with more detailed and clearer appearances for 3D assets compared to existing methods.

Critical strengths
  • The problem addressed by this paper is a critical and important one in texture generation: evaluating texture quality during post-training based only on the final textured/painted asset cannot properly capture the quality differences between the views used to generate the final texture. Yet, many existing post-training approaches for texture generation rely on 3D-asset-level evaluation.

  • The paper is well-written and easy to follow.

  • The constructed human preference texture dataset and the trained reward model could be valuable to the 3D community for further research, even beyond the online post-training approach leveraged in this work, given that the authors make the dataset and reward model checkpoint publicly available.

  • The method verifies the effectiveness of such per-view constructed and trained reward models by applying them to an existing online post-training method.

Critical weaknesses
  • In the contribution bullet points in the Introduction, evaluating the reward model and the post-training framework is listed as one of the contributions of this work. While this evaluation is valuable and I appreciate the authors’ efforts in conducting it, I do not consider it a contribution in itself. In fact, evaluation is a required component of any research paper as it allows the community and audience to assess the paper’s position compared to existing work in the field. Therefore, I strongly recommend removing this item from the contribution bullet points.

  • Throughout the paper, especially in Sec. 3.3, the proposed post-training framework is presented as if DiffusionNFT were proposed by this paper. However, based on my understanding of Sec. 3.3, almost all the mathematical formulations and the overall approach appear to be identical to those presented in the original DiffusionNFT paper (with the exception, of course, of the per-view formulation of DiffusionNFT). The authors should clearly explain the main modifications and contributions they make to the DiffusionNFT approach. Otherwise, the current presentation gives the impression that all the mathematical formulations in Sec. 3.3 are proposed by this paper, which is not the case. I understand that one of the main contributions of this work is applying the DiffusionNFT post-training approach to view-level-credited post-training for 3D texture generation, rather than image generation, which was the primary focus of DiffusionNFT. However, even this distinction is not sufficiently clear in the current version of Sec. 3.3. Overall, I recommend presenting the mathematical formulations and methodology inherited from DiffusionNFT in a preliminary section, followed by a clear explanation of the modifications and how DiffusionNFT is adapted to the 3D texture generation task. This would help readers clearly distinguish between the contributions of the original DiffusionNFT paper and those of the current work.

  • Although the paper is properly evaluated and analyzed through quantitative experiments and a user preference study, and the existing qualitative examples in Figure 4 are mostly convincing, I believe it would be worthwhile to include more qualitative examples and comparisons with the baselines, either in the main paper or in the appendix. These examples could demonstrate the differences across a wider range of mesh geometries, topologies, and texture patterns.

  • Minor comments:

    • In the Introduction and Appendix, the Bradley-Terry model is mentioned but not cited. I recommend adding the appropriate citation, along with a brief explanation of the intuition and mathematical formulation behind the Bradley-Terry objective function, which is widely used in RLHF.

    • In Equation (1), $ \mathcal{F}_{\theta} $ is not defined and is never used again. I assume this refers to the multi-view 3D texture generation model, which is also considered the policy model in the DiffusionNFT approach, particularly in Equation (14). Is this correct? If so, please use consistent notation throughout the paper.

Questions for authors
  • Is there any specific theoretical or practical contribution to DiffusionNFT that was necessary to make the proposed approach work? To be precise, as I also pointed out in the weaknesses above, I understand that the proposed method applies DiffusionNFT to the 3D texturing task using the proposed per-view reward model. However, I am specifically asking whether any theoretical or practical novelty has been introduced to the post-training optimization framework itself. I would appreciate clarification from the authors on this point as it would help reveal the actual novelty of the proposed method on the post-training optimization side.

  • Could the authors clarify how the reference images and input mesh geometries shown in Figure 4 were obtained and what their sources are? I could not find this information in the paper. Based on my observations, it seems that the reference images might be renderings of textured meshes from a frontal view (obtained from a texture generation model). I may be wrong about this, but in all the examples in Figure 4, the geometric details of the objects appear to be identical in both the reference images and the input meshes. If this is the case, what is the rationale behind using such reference images? More importantly, what would happen if the reference image were generated by an image-generation model and the input mesh came from a different source? In this scenario, the geometry of the object depicted in the reference image might differ from that of the actual mesh to be textured. I would expect such a discrepancy to potentially cause the proposed method to fail or produce suboptimal results. I may be mistaken in this assumption and so I would appreciate clarification from the authors on this point particularly regarding the method's robustness to geometric discrepancies between the reference image and the input mesh.

8 comments

Thank you for recognizing the importance of view-level texture evaluation and the potential value of our dataset and reward model. We have revised the contribution statement, separated the inherited DiffusionNFT formulation from our modifications, and added qualitative comparisons and a reference–geometry mismatch study. Below, we provide clarifications and additional results for each of your comments and questions in turn.


W1. Evaluation as a contribution. We agree and have removed evaluation as a standalone contribution. The revised list focuses on three components: the human view-level preference dataset, the context-aware per-view reward model, and the view-level credit-assignment scheme built on DiffusionNFT. Experiments are presented as evidence supporting these components. The revised contributions are reproduced below.

  • We construct a large-scale dataset of human view-level preferences for 3D texture generation. We establish task-specific criteria covering reference fidelity, material plausibility, and albedo cleanliness to guide same-view comparisons within the full multi-view context.
  • We develop a context-aware per-view texture reward model that scores each queried view using the reference image and the complete multi-view context. This design combines global context with local evaluation, providing fine-grained quality feedback that captures local differences obscured by asset-level scores.
  • We introduce a view-level DiffusionNFT post-training scheme that derives advantages from comparisons among candidates at the same viewpoint and assigns corresponding feedback to each view, enabling fine-grained credit assignment within a shared multi-view generation policy to accelerate convergence.

In the revised manuscript:

  • Introduction, p. 3, lines 124–134: Reorganized the contributions around the preference dataset, reward model, and view-level post-training scheme, removing evaluation as a standalone contribution.

W2. Distinguishing inherited DiffusionNFT components from our modifications. Thank you for pointing out the ambiguity in the original presentation. We have added Sec. 3, “Preliminaries: DiffusionNFT,” to introduce and attribute the underlying framework before presenting our method. We explicitly retain DiffusionNFT's positive/negative branch construction and regression mechanism. The EMA sampling-policy update and timestep-weighting mechanism are also inherited rather than claimed as new contributions.

Sec. 4.3 now focuses on our modification: rewards are normalized across candidates independently at each fixed viewpoint, and the resulting weights are applied to the corresponding view losses. The generator still processes the full multi-view input jointly. Thus, a candidate no longer has to receive the same preference feedback at every view.

We also added Eq. (16), which decomposes the view-weighted objective into a shared-weight NFT counterpart and a term capturing within-candidate view-weight deviations. This makes the difference explicit without reattributing the underlying optimizer to our work.

In the revised manuscript:

  • Sec. 3, p. 4, Eqs. (1)–(3): Introduced the inherited DiffusionNFT formulation in a dedicated preliminary section.
  • Sec. 4.3, pp. 6–7, Eqs. (13)–(16): Clarified view-level normalization, loss weighting, and the objective decomposition.

W3. Broader qualitative comparisons. We have added six examples comparing MV-Adapter before and after CrediTex post-training across a wider range of geometries, topologies, and texture patterns. Each example shows the reference image and all six output views, with red boxes highlighting local differences.

In the illustrated cases, CrediTex reduces misplaced patterns and color artifacts on less-visible surfaces, such as the chair's underside and the back of the interlocking rings, while better preserving the dragon's wing colors and the rope's braided pattern. These examples supplement the stylized and realistic characters, text-bearing objects, and architecture already shown in Fig. 4.

In the revised manuscript:

  • Sec. 5.3, p. 9: Described the range of qualitative examples and added a reference to the supplementary comparisons.
  • Appendix J, p. 20: Added analysis of further geometries, topologies, and texture patterns.
  • Fig. 14, p. 23: Presented six additional examples with complete six-view comparisons before and after post-training.

M1. Bradley–Terry citation and intuition. We have added Bradley and Terry (1952) to the references and cited it in the Introduction and alongside the preference-learning objective in Sec. 4.2. We also explain that σ(sw−sl)\sigma(s_w-s_l) models the probability of choosing the preferred candidate over the rejected one. Minimizing its negative log-likelihood encourages a higher score for the preferred candidate, allowing learning from pairwise judgments without requiring absolute quality ratings.

In the revised manuscript:

  • Introduction, p. 2: Added the Bradley–Terry citation.
  • Sec. 4.2, pp. 5–6, Eq. (11): Cited the original work and explained the preference probability and negative log-likelihood objective.
  • References, p. 11: Added the bibliographic entry for Bradley and Terry (1952).

M2. Definition of the generator notation. Thank you for pointing this out. Fθ\mathcal F_\theta denotes the multi-view diffusion generator parameterized by θ\theta, which is the generation policy optimized during post-training. We have added this definition immediately after the generation equation, now Eq. (4), on p. 4. The notation vθv_\theta used in the DiffusionNFT formulation denotes this policy's prediction at a noisy input and timestep, whereas Fθ\mathcal F_\theta denotes the overall conditional generation mapping. The sampling policy uses the corresponding parameters θold\theta_{\mathrm{old}}, as clarified in Sec. 4.3.

In the revised manuscript:

  • Sec. 4.1, p. 4, Eq. (4) and line 196: Defined the multi-view diffusion generator Fθ\mathcal F_\theta.
  • Secs. 3 and 4.3, pp. 4 and 6–7: Specified the trainable and sampling policies used in post-training.

Q1. Contribution on the optimization side. Our contribution is a view-level credit-assignment modification within DiffusionNFT, rather than a new branch construction or a new optimization algorithm. A per-view reward model alone does not ensure per-view supervision during optimization: aggregating its outputs into one trajectory-level advantage still broadcasts the same preference feedback to every view. We instead normalize rewards across candidates at each fixed viewpoint and retain the resulting view-specific weights in the regression objective.

We have added an objective decomposition in Sec. 4.3 (p. 7, Eq. (16)) to explain how view-level weighting differs from assigning a shared weight to all views. Writing the branch losses as ℓg,v±=∥vθ,g,v±−ug,v∥22\ell_{g,v}^{\pm}=\lVert v_{\theta,g,v}^{\pm}-u_{g,v}\rVert_2^2 and the mean view weight as rˉg=1N∑v=1Nrg,v\bar r_g=\frac{1}{N}\sum_{v=1}^{N}r_{g,v}, we express the objective as

Lview=Et,ε[w(t)GN∑g,v(rˉgℓg,v++(1−rˉg)ℓg,v−)]⏟LNFT: shared view weights+Et,ε[w(t)GN∑g,v(rg,v−rˉg)(ℓg,v+−ℓg,v−)]⏟view-level weight deviations.\mathcal{L}_{\mathrm{view}} = \underbrace{\mathbb{E}_{t,\boldsymbol{\varepsilon}}\left[\frac{w(t)}{GN}\sum_{g,v}\left(\bar r_g\ell_{g,v}^{+}+(1-\bar r_g)\ell_{g,v}^{-}\right)\right]}_{\mathcal{L}_{\mathrm{NFT}}\text{: shared view weights}} + \underbrace{\mathbb{E}_{t,\boldsymbol{\varepsilon}}\left[\frac{w(t)}{GN}\sum_{g,v}(r_{g,v}-\bar r_g)(\ell_{g,v}^{+}-\ell_{g,v}^{-})\right]}_{\text{view-level weight deviations}}.

Here, LNFT\mathcal L_{\mathrm{NFT}} is a shared-weight counterpart constructed using rˉg\bar r_g, with the remaining terms unchanged. The additional term retains within-candidate differences and modulates how each view contributes to the shared parameter update. It vanishes when all view weights within a candidate are equal.

The existing ablation in Sec. 5.5 and Fig. 6(a) compares this formulation with trajectory-level aggregation and advantage averaging. Over the 250-step training run shown in Fig. 6(a), both shared-feedback alternatives exhibit a sharp reward decline in the later stages, whereas view-level weighting continues to improve.

In the revised manuscript:

  • Sec. 3, p. 4: Separated the inherited DiffusionNFT components from our modifications.
  • Sec. 4.3, pp. 6–7, Eqs. (14)–(16): Presented view-specific weighting and an algebraic decomposition explaining its effect.
  • Sec. 5.5 and Fig. 6(a), p. 10: Provided the supporting ablation of view-level weighting, trajectory-level aggregation, and advantage averaging.

Q2. Input sources and robustness to reference–geometry mismatch. Thank you for asking us to clarify the input sources. The reference images in Fig. 4 are generated by image-generation models. The input meshes are subsequently reconstructed from those images using a pretrained image-to-3D geometry generator that was not publicly available at submission. Thus, the reference images are not obtained by rendering the final textured meshes. Their geometric similarity arises from conditioning mesh reconstruction on the reference image, which provides a coherent image–shape pair for the texturing task.

We have also added a mismatch study in Appendix K (p. 21) that keeps the original crowned-lion texture reference fixed and pairs it with three meshes: a decorated lion with local structural differences, a lion without accessories, and a standing tiger. These meshes are reconstructed from modified images using the same geometry generator.

Fig. 15 compares MV-Adapter before and after CrediTex post-training. In the illustrated examples, CrediTex reduces misplaced colors on the back of the cape and spurious decorative patterns on the soles, showing that its benefits persist under the tested discrepancies. These results provide qualitative evidence that CrediTex retains its benefits under the tested reference–geometry mismatches.

In the revised manuscript:

  • Sec. 5.3, p. 9: Clarified how the reference images and meshes in Fig. 4 were obtained.
  • Appendix K, p. 21: Described the construction and analysis of the reference–geometry mismatch cases.
  • Fig. 15, p. 24: Compared MV-Adapter before and after CrediTex under three mismatch settings.
Reviewer_cX9ZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces CrediTex, a 3D texture generation framework with per-view rewards and fine-grained credit assignment. The authors construct a dataset of human preference pairs to train reward model. The reward model is further used to compute advantages to guide view-level reinforcement learning. Experiments evaluate reward prediction and texture generation across three generation backbones, with quantitative comparisons, ablations, and a human preference study.

Critical strengths
  • Coherent motivation and an effective method. The paper connects the limitations of asset-level feedback with a method that preserves view-level supervision by reward modeling and policy optimization. Experiments including a user study show the effectiveness of CrediTex. These results support the practical value of the overall design.
  • Substantial effort in constructing human preference data for 3D texturing. Collecting and curating 48K human preference pairs from multi-view cases represents a substantial data-generation and annotation effort. This provides supervision for learning texture-quality judgments, adding a valuable empirical contribution alongside the optimization method.
Critical weaknesses
  1. The preference dataset is insufficiently documented. Section 3.2 specifies the comparison criteria and retained pair counts, but do not identify the model used for candidate generation. The sources of the reference images and geometries are unclear. The human annotation details are also missing. How many annotators and independent judgments were used per view-level comparison? How were disagreements and ties handled?
Questions for authors
  1. Clarification of preference-data construction and annotation. Which models and checkpoints generated the preference candidates, and where did the reference images and geometries originate? How many annotators participated, and how many independent judgments were collected per view-level comparison? How were judgments aggregated into win/loss/tie labels, particularly when annotators disagreed, and what annotation-quality checks were performed?

  2. Clarification of reward training. The Bradley–Terry objective learns from binary preferences, whereas Eqs. (11)–(13) use continuous score spacings to determine diffusion-training weights. As a suggestion for additional validation, a comparison with rank-based advantages could help evaluate sensitivity to score spacings.

2 comments

Thank you for recognizing the motivation, practical value, and data-collection effort of CrediTex. We have expanded the documentation of preference-data construction and annotation and added the suggested comparison with rank-based advantages. Below, we provide clarifications and additional results for each of your comments and questions in turn.


W1 and Q1. Preference-data construction and annotation. These comments concern the same documentation gap, so we address them together. We have added Appendix B with the following details.

  • Candidates and sources. For each case, the Qwen-Image-Edit SFT checkpoint used in our experiments generates two six-view candidates with random seeds 0 and 42 under the same reference image and mesh. Reference images are collected from multiple websites or produced by image-generation models. The conditioning meshes are produced by a pretrained image-to-3D geometry generator that was not publicly available at submission and remain fixed during texture optimization. Appendix B also describes reference-image screening.
  • Annotators and presentation. Fifteen annotators complete training and trial annotation. Each view-level comparison receives two independent judgments. Annotators see the reference image and both complete six-view sets but judge only the designated target view. A/B ordering is randomized, model identities are hidden, and the available responses are A, B, and tie.
  • Aggregation. A clear preference is retained only when both annotators select the same winner. Any explicit tie or disagreement is assigned a tie label and excluded from the default reward-model training data.
  • Quality control. We organize the data into batches of approximately 3,000 view-level comparisons and sample 50 cases from each batch for expert review. Two experts divide the sampled cases between them. We accept a batch only when view-level agreement with expert judgments exceeds 90%. Otherwise, we reassign it for reannotation.
  • Data splitting. We retain 47,714 clear preference pairs from 18,343 cases and randomly split the cases into training and validation sets at a 95:5 ratio. We keep all candidates and view-level comparisons from the same case in the same split. The 1,024 unseen cases used for reward-model evaluation form a separate test set and are not part of these 18,343 cases.

In the revised manuscript:

  • Appendix B, pp. 14–15: Documented data sources, candidate generation, image screening, annotation, quality control, and case-level train–validation splitting.
  • Sec. 5.1, p. 7: Added a reference to the detailed data and annotation protocol.

Q2. Sensitivity to reward score spacings. We agree that pairwise preference learning does not by itself establish calibrated absolute quality scores. Our optimization uses rewards normalized across candidates at each fixed viewpoint, so the relevant question is whether retaining relative score spacings helps beyond the ordering alone.

Following your suggestion, we added a rank-based variant. At each viewpoint, we sort the eight candidates by their reward scores, replace the scores with uniformly spaced values 0,1/7,…,10,1/7,\ldots,1, and apply the same normalization, clipping, and branch-weight mapping. Both variants use training seed 43. This is a separate paired comparison from the seed-42 run reported in the main table.

MV-Adapter variantFID_CLIP ↓CMMD ↓CLIP-I ↑LPIPS ↓PSNR ↑SSIM ↑
Original18.1260.58040.91700.420516.9140.4463
CrediTex, rank-based, seed 4316.6820.54440.92270.396917.9090.4721
CrediTex, score-based, seed 4316.6500.53170.92340.393318.0300.4784

Both variants improve over the base model, showing that optimization remains effective after removing the original score spacings. Score-based advantages achieve better results on all six metrics in this paired experiment. Fig. 11 shows an overall upward reward trend for both variants over 250 training iterations with seed 43. Score-based advantages generally achieve higher rewards in the later stages, with the plotted curves ending at approximately 7.25 for score-based advantages and 6.92 for rank-based advantages. Together, these results suggest that the relative score spacings provide useful additional information in this setting.

In the revised manuscript:

  • Appendix H, p. 19: Described the score-based and rank-based comparison using training seed 43.
  • Table 7, p. 19: Reported the six quantitative metrics for the two variants and the base model.
  • Fig. 11, p. 20: Compared the reward curves over.

Thank you for the response, the explanation and the experiment have addressed my concerns. I retain my score at 3.

Reviewer_BR9QOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses preference alignment for multi-view texture generation. Regarding data, the authors constructed 48,000 view-level human preference pairs and used them to fine-tune a Qwen3.6-27B model with a Bradley-Terry loss, resulting in a reward model capable of scoring individual views. The model takes as input a composite image of six views, a reference image, and a prompt specifying the target view; it derives a score from the logit difference between "yes" and "no" responses, with KV cache reuse employed to significantly accelerate inference. During training, the model calculates an advantage value for each candidate view, which then drives the view-specific loss in DiffusionNFT. The approach yields comprehensive improvements in quality metrics across three backbone models: MV-Adapter, Qwen-Image-Edit, and FLUX.2-klein. The alignment between the reward model and human judgment increased from 71.3% (single-view) to 78.7%, with the most notable improvement seen in the reduction of "Janus" artifacts (erroneous multi-face/multi-front representations) in front-facing views.

Critical strengths
  1. The "global context combined with local evaluation" approach offers practical value. In the absence of a complete set of views, scoring models often misidentify valid front-facing details as "Janus" artifacts. By incorporating the full 6-view context, the accuracy for front-facing views improved from 61.2% to 75.5%, and the performance gap across different views narrowed from 15.2 percentage points to 4.6 percentage points.
  2. The dataset is a significant contribution; the open-sourcing of 48k view-level preference pairs is likely to foster further research in this field.
Critical weaknesses
  1. The effectiveness of the reward model was tested exclusively on diffusion-based models. Its potential impact on autoregressive multi-view image generation models, such as MV-AR [1], remains unclear.
  2. There is a lack of comparative experiments. Although methods such as MVC-ZigAL (RL using a shared multi-view reward) and Carve3D (RL for multi-view consistency) are mentioned in the "Related Work" section, no empirical comparisons are provided.
  3. The user study lacks specific details.

[1] Auto-Regressively Generating Multi-View Consistent Images. ICCV 2025.

Questions for authors

Tie cases account for a large portion of the data: 57%. Could this high proportion of ties negatively affect the training of the reward model?

4 comments

Thank you for recognizing the practical value of combining global context with per-view evaluation and the potential value of our preference dataset. We have added autoregressive validation, comparisons with the two suggested post-training methods, a tie-inclusive reward-model experiment, and a more detailed user-study protocol. Below, we provide clarifications and additional results for each of your comments and questions in turn.


W1. Applicability to autoregressive generators. Thank you for this suggestion. We have added an experiment on NAMVIS, an autoregressive multi-view generator, using the CrediTex reward model to guide GRPO post-training. We use albedo renderings from GObjaverse and evaluate the original model and the checkpoints at steps 200 and 500 on 200 test objects. This experiment evaluates the applicability of our reward model beyond diffusion-based generators.

NAMVIS checkpointFID_CLIP ↓CMMD ↓CLIP-I ↑LPIPS ↓PSNR ↑SSIM ↑
Original34.682.1750.7890.23816.050.856
GRPO, step 20033.782.1180.7930.23516.190.859
GRPO, step 50031.292.0590.7970.22916.910.869

All six metrics improve at step 200 and improve further at step 500. The training reward also increases from −15.22 at the beginning of training to −0.68 at step 500, despite fluctuations. On the fixed validation set, the reward increases from −13.40 to 0.30 over the same training period. Qualitative comparisons show cleaner color regions and fewer dark patches. Together, these results support the applicability of the CrediTex reward model to autoregressive multi-view generation through GRPO post-training.

In the revised manuscript:

  • Appendix G, pp. 18–19: Experimental setup and analysis of the NAMVIS post-training results.
  • Table 6, p. 18: Quantitative comparisons of the original NAMVIS model and the step-200 and step-500 checkpoints.
  • Fig. 10, p. 19: Training reward and fixed-set validation reward curves.
  • Fig. 13, p. 22: Qualitative comparisons before and after post-training.

W2. Comparisons with MVC-ZigAL and Carve3D. Thank you for this suggestion. We have adapted MVC-ZigAL and Carve3D to our texture-generation task using MV-Adapter. Table 2 now includes comparisons with both methods.

MV-Adapter variantFID_CLIP ↓CMMD ↓CLIP-I ↑LPIPS ↓PSNR ↑SSIM ↑
Original18.130.5800.9170.42116.910.446
+ Carve3D, adapted18.020.5750.9180.42016.970.449
+ MVC-ZigAL, adapted18.070.5860.9170.41916.940.447
+ CrediTex17.030.5510.9220.40217.370.456

CrediTex achieves better results on all six reported metrics under these adaptation and training settings. Appendix F provides implementation details and an analysis of the adapted objectives. Carve3D uses reconstruction consistency, which does not directly measure reference fidelity or human texture preferences. For MVC-ZigAL, Fig. 9 plots the standard and zigzag rewards over the first 100 training steps. The curves remain close for both PickScore and HyperScore, suggesting limited average reward gains from zigzag refinement in this setting. These results suggest that CrediTex provides effective feedback for optimization while accounting for the specific requirements of texture generation.

In the revised manuscript:

  • Table 2, p. 8: Added quantitative comparisons with adapted MVC-ZigAL and Carve3D.
  • Sec. 5.2, p. 9: Discussed the comparative results.
  • Appendix F, pp. 17–18: Specified baseline adaptations and training configurations.
  • Fig. 9, p. 18: Compared standard and zigzag PickScore and HyperScore rewards over the first 100 training steps.

W3. User-study details. We have expanded the user-study description in Sec. 5.2. Six evaluators with relevant professional expertise each assess all 50 cases, yielding 300 individual judgments. For each case, they inspect the reference image and the two methods' complete six-view outputs, randomly assigned to A and B. They select A, B, or a tie based on reference fidelity, material plausibility, and albedo cleanliness, using the same criteria as the preference annotation. For each case, the most frequently selected option determines the final outcome, with any tie for the highest vote count recorded as a tie. This gives 29 wins for CrediTex (58%), 14 wins for the base model (28%), and 7 ties (14%) across the 50 cases, as reported in Fig. 5.

In the revised manuscript:

  • Sec. 5.2, p. 9: Expanded the user-study protocol, including evaluators, presentation, assessment criteria, and case-level aggregation.

Q1. Effect of the high tie rate. Thank you for raising this question. We would like to clarify that we train our default reward model only on clear preference pairs, without tie data. We exclude both explicit ties and comparisons on which the two annotators disagree. Finally, we retain 47,714 clear preference pairs to train and validate the reward model.

To examine the effect of tie supervision, we have also trained a reward model with ties included. We define the score difference as Δ=sA−sB\Delta=s_A-s_B and optimize the following loss

L=−E[ylog⁡σ(Δ)+(1−y)log⁡σ(−Δ)],\mathcal L=-\mathbb E\left[y\log\sigma(\Delta)+(1-y)\log\sigma(-\Delta)\right],

where y=1y=1 or 00 represents a clear preference and y=12y=\tfrac12 represents a tie. Tie supervision encourages equal scores within a pair rather than requiring both scores to be zero.

Reward-model trainingAccuracy ↑Equal-score rate ↓Reward mean ± std.
Without ties78.73%0.73%5.394 ± 10.022
With ties71.79%16.16%0.115 ± 0.664

We observe that including ties reduces preference accuracy from 78.73% to 71.79% and increases the equal-score rate on clear preference pairs from 0.73% to 16.16%. We also compare the reward distributions in Fig. 12 and find that training with ties concentrates scores near zero, with the mean changing from 5.394 to 0.115 and the standard deviation decreasing from 10.022 to 0.664. These results support our choice to train the default reward model on clear preferences only. During subsequent reinforcement learning, more frequent equal scores can also remove useful preference distinctions. If all candidates in a group receive the same score, their normalized advantages are zero.

In the revised manuscript:

  • Appendix B, pp. 14–15: Clarified how explicit ties and annotation disagreements are handled.
  • Appendix I, pp. 19–20, Eqs. (17)–(18): Added the tie-inclusive training setup and objective.
  • Table 8, p. 20: Compared preference accuracy, equal-score rates, and reward statistics.
  • Fig. 12, p. 21: Visualized reward distributions with and without tie-inclusive training.

Thank you to the author for the detailed response. It has fully addressed my concerns. I will raise my score to 3 (weak accept).

Thank you for your thoughtful feedback and for taking the time to review our revisions. We are glad that our responses have addressed your concerns. Your suggestions helped us strengthen both the experiments and the presentation of our work. We wish you all the best!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Contributions and Revisions

Dear Reviewers and ACs,

Thank you for your time and constructive feedback throughout the review process. Your suggestions have helped us strengthen the experiments and explain our method and contributions more clearly.

Our work focuses on generating plausible textures for surfaces that lack direct reference-image guidance while preserving clear, coherent details across views, which remains a key challenge in reference-guided 3D texture generation. To address this challenge, we introduce a human view-level preference dataset, a context-aware per-view reward model, and a post-training method that assigns feedback to individual view losses. Using this framework, we post-train multiple backbones and evaluate the resulting improvements through quantitative experiments and extensive visual comparisons.

We appreciate the reviewers' recognition of our work. Reviewer BR9Q highlighted the practical value of combining global context with local evaluation and identified the preference dataset as a significant contribution. They subsequently confirmed that our responses had fully addressed their concerns and raised their rating to weak accept. Reviewer cX9Z recognized the clear motivation, the effectiveness of the approach, and the substantial effort devoted to constructing the preference dataset. Reviewer yh4t recognized the importance of the problem and the value of our dataset, per-view reward model, and post-training approach.

We have strengthened the manuscript in the following areas:

  • Broader validation: Added autoregressive experiments and comparisons with additional baselines.
  • Reward learning: Added ablations on tie-inclusive training and rank-based advantages.
  • Data and user study: Expanded the documentation of data sources, annotation, quality control, and the user study protocol.
  • Method formulation: Clarified our contributions beyond DiffusionNFT and explained the role of view-level weighting.
  • Qualitative evaluation: Added extensive comparisons across backbones, a reference–geometry mismatch study, and failure cases. We also grouped observed errors by type to make the comparisons easier to inspect.

Thank you again for your time and consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.