acceptodds
Under review as a conference paper at ICLR 2027

Delighting 3D Assets via Reinforcement Learning

Abstract

Recently, 3D generation has made progress. However, unwanted highlights in generated textures reduce realism while limiting the editability of 3D assets. Despite this progress, 3D delighting remains an underdeveloped research direction, with existing methods only mitigating lighting as a generation by-product without dedicated evaluation systems or optimization paradigms. In this paper, we highlight the lighting issue in 3D and redefine the delighting paradigm in a comprehensive manner. We first construct several lighting datasets and develop three structurally distinct evaluation models for lighting assessment. Together, they establish a unified evaluation benchmark for 3D delighting optimization. Then, we propose an adaptive decoupled multi-objective Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) joint architecture: SFT injects basic delighting capability, while RL further enhances performance. In experiments, our method improves delighting performance by 18.7% while reducing FID by 55.0%, demonstrating substantially better content consistency. To the best of our knowledge, this paper is the first to introduce an RL paradigm and reward models to the 3D delighting tasks. Code and model weights will be made public after publication.

Primary area:
learning on time series and dynamical systems
Keywords:
Shadow Removal

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_M1Z2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the problem of residual lighting and shadows in textures generated for 3D assets and proposes a delighting framework combining Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL). The method first trains a texture generation model using multi-view normal maps and reference images to establish basic delighting capabilities. The authors then develop reward models that evaluate both delighting quality and content consistency, and further optimize the generated results through reinforcement learning. In addition, the paper introduces an exploration reward and a stage-aware multi-objective optimization strategy, along with a dedicated dataset and evaluation framework. Experimental results demonstrate significant improvements in delighting performance and image fidelity.

Critical strengths

(1) The research problem has practical significance. Residual lighting in generated 3D assets can negatively affect subsequent relighting and scene editing. This paper addresses this issue with a dedicated optimization framework, offering clear practical value. (2) The reward design is well motivated. The authors separately model delighting quality and content consistency, accounting for the trade-off between shadow removal and preservation of the original textures. The ablation studies also demonstrate the effectiveness of consistency constraints in mitigating color shifts and texture degradation.

Critical weaknesses

(1) The technical contribution is not sufficiently clear. The framework combines multi-view SFT, Flow-GRPO, reward modeling, and stage-aware optimization, but the main methodological novelty beyond task-specific integration is not clearly established. The exploration reward is relatively simple, and the stage-aware scheduling appears largely heuristic.

(2) The necessity of reinforcement learning is not sufficiently demonstrated. Although the paper shows clear improvements over the SFT baseline, the RL stage also introduces additional reward supervision. It is therefore difficult to determine how much of the improvement specifically comes from online reinforcement learning.

(3) The generalization of the proposed framework could be further validated. The experiments are mainly conducted with a Hunyuan-3D-style texture generation pipeline and a relatively fixed asset/rendering setting. It would be valuable to evaluate whether the proposed RL strategy and reward design can generalize to other 3D generation pipelines, more diverse asset categories, complex materials, and different lighting conditions. The current results are promising, but broader evaluation would strengthen the practical significance of the method.

Questions for authors

No

7 comments
AuthorsOpenReview ↗

We Sincerely Appreciate Reviewer M1Z2’s Positive Assessment and Support

We sincerely thank Reviewer M1Z2 for the positive evaluation of our work and for recognizing the practical significance, reward design, and experimental effectiveness of our method. We have made every effort to address all of your concerns in detail, and if you find our responses satisfactory, we would greatly appreciate your consideration of a higher score; if you have any further questions, please feel free to raise them and we will respond promptly.

AuthorsOpenReview ↗

### W1. Methodological contribution and the exploration/scheduling mechanisms

Reviewer comment (excerpt):

The technical contribution is not sufficiently clear. The framework combines multi-view SFT, Flow-GRPO, reward modeling, and stage-aware optimization, but the main methodological novelty beyond task-specific integration is not clearly established.

Thank you for asking us to clarify the methodological contribution. Our design addresses two optimization issues in 3D delighting: exploring beyond the SFT solution without relying solely on stronger sampling noise, and balancing illumination removal against visual-quality preservation.

(1) Dedicated exploration

Flow-GRPO introduces stochasticity through SDE sampling. Our exploration reward provides an additional incentive based on latent-state variation, rather than using increased noise as the sole mechanism for exploration. The reward supplies a complementary optimization signal without requiring another learned evaluator. Figure 5 compares Flow-GRPO + R_de with our method at low, medium, and high noise levels, including both training and evaluation curves. It illustrates the tradeoff that higher noise can improve training reward without a corresponding evaluation gain. Figure 1 and supplementary Figures 16–20 also compare the full method with w/o exploration, whose displayed trajectories retain residual lighting.

(2) Stage-aware objective scheduling

A fixed weighted sum exposes the model to both objectives simultaneously, even when their optimization effects differ across generation stages. Our scheduling uses a stage-dependent signal to select the objective, with CLIP variation serving as a proxy for the transition in the generation process. The intent is to balance delighting and visual quality through the timing of their optimization.

(3) Additional Fixed-Sch ablation

The original Figure 6 already compares our method with single-objective optimization and weighted MORL under both PickScore and Aesthetic Score. We have now also evaluated Fixed-Sch. Its fixed switching point is the mean switching point observed over the complete adaptive training run, and the same value is used in every epoch. Figure 6(a,b) shows the extended comparison. In the PickScore setting, Ours reaches a higher PickScore while attaining a similar late-stage delighting reward. In the Aesthetic Score setting, Ours reaches higher late-stage aesthetic and delighting scores than the tested fixed schedule. These results support the value of adaptive scheduling relative to that fixed-switching configuration.

The two mechanisms have complementary roles: exploration supports optimization beyond the initial solution, while scheduling balances delighting and visual quality. The ablations and Fixed-Sch comparison support these roles.

Revision location: Section 4.3 (scheduling ablation analysis) and Figure 6(a,b) (fixed schedule comparison); existing Figures 1 (exploration ablation trajectories), 5 (sampling noise comparison), and 16–20 (exploration–consistency ablations).

AuthorsOpenReview ↗

### W2. Online RL versus additional reward supervision

Reviewer comment (excerpt):

Although the paper shows clear improvements over the SFT baseline, the RL stage also introduces additional reward supervision. It is therefore difficult to determine how much of the improvement specifically comes from online reinforcement learning.

The RL stage performs reward-based post-training for delighting. SFT provides the initial policy, and the reward supplies feedback for illumination removal and appearance preservation.

(1) Motivation for reward-based post-training

This formulation follows established work on optimizing diffusion and flow-matching models with rewards, including DDPO and Flow-GRPO. DDPO also compares policy optimization with reward-weighted likelihood methods in its image-generation tasks, providing relevant motivation for our choice of optimization framework.

(2) Questions addressed by the comparisons

Our SFT–Ours comparison evaluates the complete reward-based post-training stage. SSIM increases from 0.8465 to 0.9063, PSNR from 22.34 to 27.48, and L2 decreases from 0.1128 to 0.0514. Figure 3 additionally compares our method with Flow-GRPO + R_de, which also uses the delighting reward, evaluating our optimization design with the same task reward.

These comparisons address two questions: SFT–Ours measures the complete reward-based post-training gain, and Flow-GRPO + R_de–Ours evaluates the optimization design under the same task reward. Which alternative reward-use scheme would the reviewer consider most relevant for further comparison?

Revision location: Section 2.2 (reward-based post-training), Section 4.2 (two comparison settings), Figure 3 (same-reward optimization comparison), Table 1 (post-training over SFT).

Reviewer_NHCGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a reinforcement learning framework for 3D texture delighting. Instead of applying an independent delighting model to reference images before texture generation, the authors integrate delighting directly into the texture generation process. The proposed approach first uses supervised fine-tuning (SFT) to establish basic delighting capabilities, followed by reinforcement learning with dedicated reward models. The authors further introduce an exploration reward and adaptive multi-objective optimization to improve training effectiveness. Experiments demonstrate strong performance in removing baked-in highlights and shadows while preserving the original texture details.

Critical strengths
  1. Well-motivated framework. Integrating delighting into texture generation is a reasonable direction. It avoids potential error accumulation from independently performing image delighting and texture generation.

  2. Effective RL design. The proposed RL framework demonstrates promising improvements, and the ablation studies support the effectiveness of its key components.

  3. Strong experimental results. The proposed method achieves impressive quantitative and qualitative results, demonstrating effective removal of baked-in illumination while maintaining texture consistency and visual quality.

Critical weaknesses
  1. Insufficient details of the user study. Although the paper reports human preference results, important details of the evaluation protocol are missing, such as the number of participants, the number of evaluated samples, and the statistical significance of the results.

  2. Lack of limitation and failure-case analysis.

Questions for authors

What motivates the use of the SAM 3D image encoder instead of more general-purpose visual features, such as DINOv2? Have the authors conducted an ablation study comparing different image encoders?

8 comments
AuthorsOpenReview ↗

We Sincerely Appreciate Reviewer NHCG’s Support and Positive Assessment

We sincerely thank you for your positive evaluation and support of our work. We have made every effort to address all of your concerns in detail, and if you find our responses satisfactory, we would greatly appreciate your consideration of a higher score; if you have any further questions, please feel free to raise them and we will respond promptly.

AuthorsOpenReview ↗

### W1. Human-study protocol and statistical uncertainty

Reviewer comment (excerpt):

Insufficient details of the user study. Although the paper reports human preference results, important details of the evaluation protocol are missing, such as the number of participants, the number of evaluated samples, and the statistical significance of the results.

Thank you for highlighting the need to make the user-study protocol explicit. The study contains 20 objects. The analysis includes 10 human raters, all interns with computer-related backgrounds, providing 10 valid evaluations per object. Participants compared the two methods’ three-view results side by side and selected the preferred result for delighting instruction following and appearance consistency. Overall is computed from these component rankings.

(1) External-method preferences

We have also extended the evaluation beyond internal variants through pairwise comparisons with MaterialMVP and UniTEX. Human preferences favor Ours in 70% / 65% of cases for delighting / consistency against MaterialMVP, and 75% / 70% against UniTEX.

Revision location: Section 4.1 (participants and samples), Appendix E (evaluation protocol), Section 4.4 (preference analysis), and Table 3 (external baseline preferences).

AuthorsOpenReview ↗

### W2. Limitations and failure cases

Reviewer comment (excerpt):

Lack of limitation and failure-case analysis.

We have added Section 5: Limitations and retained the expanded discussion in Appendix F. The main focus is the tradeoff between illumination removal and preservation of complex colors and fine textures: stronger preservation can leave residual lighting, whereas more aggressive removal can change original colors or details. The revised captions of Figures 10, 13, and 15 identify concrete limitations of the complete method.

(1) Residual illumination

First, lighting removal can remain incomplete. In Figure 10, the table-lamp result retains smooth intensity variations on the lampshade and base compared with the more uniform reference. This illustrates that the method can reduce strong shadows while still retaining residual appearance variations associated with illumination.

(2) Local color and texture discrepancies

Second, preserving the overall appearance does not guarantee exact recovery of fine texture and color. In Figures 13 and 15, the patterned lunch bag and ceramic vase retain their overall identity and surface patterns, but local colors or fine-scale details differ from the reference.

These full-method examples illustrate the remaining tradeoff between illumination removal and local appearance preservation. Figures 16–20 additionally show the effects of removing exploration or consistency in the ablation study.

Revision location: Section 5 (delighting–appearance tradeoff), Appendix F (extended limitations), and Appendix C Figures 10 (lamp residual illumination), 13 (bag appearance discrepancies), and 15 (vase texture discrepancies).

Reviewer_9hKQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduced a RL(Reinforcement Learning) based method for 3D assets delighting. The main contributions are: 1) this paper introduced a new pipeline for 3D mesh delighting with two-stage training that combines SFT (supervised fine-tuning) and RL. 2) It proposed a new benchmark for quantitative evaluation of the 3D delighting problem. Through qualitative and quantitative experiments, it shows better results compared to previous methods.

However, it needs more validation for its contribution:

  1. The paper misses an important research direction: PBR texture generation, which is directly related to the proposed 3d delighting. Since PBR generation explicitly estimates the albedo separated from the illumination, it actually addresses the problem from a different angle.

  2. The proposed benchmark metrics were fully built based on VLM evaluation, whose reliability is questionable, as VLMs can be unreliable, especially for ambiguous tasks.

  3. Lots of details of the experiments are missing, which makes it hard to replicate. For example, the exact checkpoint variants of the VLM, the optimizer settings for the fine-tuning, and the fine-tuned parameters. What's more, the details of the human evaluation in Fig. 7 should also be added, e.g., the number of participants, where they are from, and what the survey instructions were.

Critical strengths
  1. This paper first introduces RL methods to solve the illumination problem in 3D asset generation.
  2. It establishes a new benchmark that enables quantitative evaluation for 3d delightning problem.
Critical weaknesses
  1. The paper misses an important research direction: PBR texture generation, which is directly related to the proposed 3d delighting. Since PBR generation explicitly estimates the albedo separated from the illumination, it actually addresses the problem from a different angle. For example, "Material Anything: Generating Materials for Any 3D Object via Diffusion, CVPR 2025" and "MaterialMVP: Illumination-Invariant Material Generation via Multi-view PBR Diffusion, ICCV 2025". Especially the second one, it applied a Consistency-Regularized Training strategy to enforce stability across varying viewpoints and illumination conditions, ensuring illumination-invariant results. Thus, the discussion and comparison with this line of methods are necessary for validating the contribution, e.g., adding discussion in the related work and comparison in Table 1.

  2. The proposed benchmark metrics were fully built based on VLM evaluation, whose reliability is questionable, as VLMs can be unreliable, especially for ambiguous tasks. An analysis of the overlap between human evaluation and the proposed metrics could partially validate them.

Questions for authors

None

13 comments
AuthorsOpenReview ↗

Our Gratitude to Reviewer 9hKQ for the Feedback

We would like to take this opportunity to convey our sincere appreciation to Reviewer 9hKQ for the careful consideration and feedback provided on our work. We genuinely hope that these clarifications will help resolve any remaining uncertainties and provide a clearer understanding of the merits of our work. If you find our responses convincing and satisfactory, we would be sincerely grateful if you could consider updating your assessment and raising your score accordingly. We also highly value the opportunity for continued discussion and would be delighted to address any additional questions or suggestions you may have.

AuthorsOpenReview ↗

### W1. PBR texture generation and comparisons

Reviewer comment (excerpt):

The paper misses an important research direction: PBR texture generation, which is directly related to the proposed 3d delighting. Since PBR generation explicitly estimates the albedo separated from the illumination, it actually addresses the problem from a different angle.

We agree that PBR material generation is directly relevant to 3D delighting. In particular, its albedo output seeks to represent appearance without illumination, while additional material channels support rendering under new lighting. Our work focuses on generating multi-view delighted RGB that removes illumination while retaining the reference appearance.

(1) Relationship to PBR generation

We have therefore evaluated MaterialMVP and Material Anything, together with UniTEX. MaterialMVP uses consistency-regularized training to encourage illumination-invariant material generation. Material Anything estimates materials progressively across views and refines them in UV space. Their illumination-invariant appearance outputs make them relevant comparisons for our reward-based optimization approach.

(2) Additional comparison results

In the additional comparison, MaterialMVP obtains FID 19.76, PSNR 25.62, and VLM-based R_de 0.8676; Material Anything obtains 25.63, 24.12, and 0.6348. Ours obtains 12.38, 27.48, and 0.9162, respectively. Table 1 gives all seven measures, and the four-view qualitative comparison shows the resulting surface appearance.

We evaluate the shared albedo/base-color outputs for illumination removal and appearance preservation. Material Anything takes Paint3D-textured assets as input; MaterialMVP and UniTEX use mesh geometry and a reference image. Section 4.1 specifies these settings.

Revision location: Section 1 (PBR task relationship), Appendix A.1.4 (PBR related work), Table 1 (PBR baseline comparison), Section 4.1 (baseline input settings), and Figure 8 (multiview visual comparisons).

AuthorsOpenReview ↗

### W2. Reliability of benchmark metrics

Reviewer comment (excerpt):

The proposed benchmark metrics were fully built based on VLM evaluation, whose reliability is questionable, as VLMs can be unreliable, especially for ambiguous tasks.

We have added a human alignment study of the learned VLM evaluator, examining both its scores and scoring rationales.

(1) Alignment of VLM and human scores

We evaluate 20 test samples with 10 human raters. Raters independently score illumination removal and appearance preservation before viewing the VLM rationales. Results across the 200 human–VLM score pairs are reported in Table 5:

Evaluation dimensionSpearman correlation ↑Mean absolute difference ↓
Illumination removal0.3150.753
Appearance preservation0.3110.835

VLM and human scores are positively correlated in both dimensions.

(2) Human assessment of scoring rationales

Raters assess the VLM rationales as reasonable, neutral, or unreasonable. Table 6 reports:

Evaluation dimensionReasonableNeutralUnreasonable
Illumination removal~80%~15%~5%
Appearance preservation~80%~15%~5%

In each dimension, approximately 80% of rationales are rated reasonable, indicating that raters found most explanations consistent with the image content and the associated scores.

(3) Relation to the evaluation framework

This study directly examines alignment between the learned VLM evaluator and human judgments. FID, SSIM, PSNR, L2, and the human and GPT-5.5 preference evaluations provide complementary assessments of delighting and appearance preservation.

Revision location: Appendix E.4 (human evaluator alignment), Table 5 (score alignment results), and Table 6 (rationale quality assessment).

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

The paper studies removal of baked-in lighting from 3D textures while preserving appearance. It constructs paired data and learned reward models, initializes a multiview flow model through supervised fine-tuning, and applies Flow-GRPO with a latent-displacement bonus and adaptive scheduling between delighting and aesthetic rewards.

Critical strengths

The task is practically useful, separating shadow removal from appearance preservation is well motivated. Table 1 shows substantial improvements over SFT in FID, SSIM, and PSNR. The consistency ablation demonstrates a useful constraint against destructive edits, while multiview visualizations and the human study provide additional evidence for the full system.

Critical weaknesses
  1. Table 1 includes Paint3D and Hunyuan3D, but omits relevant alternatives such as MaterialMVP and UniTEX. They cover illumination-invariant material generation and explicit delighting within texture generation, respectively, making them directly relevant to the paper's motivation.

  2. The Human and VLM rankings in Fig.7 compares only internal variants, without external baselines. Details including the number of participants/objects and uncertainty are also missing.

  3. Contradicting clarifications of the labeling model: GPT in the main text while Gemini in the appendix, while GPT is also used in evaluations.

Questions for authors

Were the reward-training labels produced by GPT-5.5 (Section 3.2) or Gemini (Appendix A.2)? If GPT-5.5 supplied the labels, how did you assess the potential bias from using it again for final VLM evaluation (Section 4.5)?

7 comments
AuthorsOpenReview ↗

Sincere Thanks to Reviewer uHUD

First and foremost, on behalf of all the authors, we would like to express our deepest gratitude and sincere respect for the time and effort you have devoted to reviewing our manuscript, as well as for your thoughtful and valuable comments. We have made every effort to thoroughly address all your concerns, and we sincerely hope that our responses will help clarify our work and provide a better understanding of its contributions. If you find that our responses have satisfactorily resolved your concerns, we would be truly grateful if you could kindly reconsider your evaluation and consider a more favorable score. Should you have any further questions or concerns, we would warmly welcome them and would be more than happy to provide detailed clarifications promptly.

AuthorsOpenReview ↗

### W1. Comparisons with MaterialMVP and UniTEX

Reviewer comment (excerpt):

Table 1 includes Paint3D and Hunyuan3D, but omits relevant alternatives such as MaterialMVP and UniTEX. They cover illumination-invariant material generation and explicit delighting within texture generation, respectively, making them directly relevant to the paper's motivation.

Thank you for identifying these relevant methods. We have added MaterialMVP, UniTEX, and Material Anything to the revised main comparison in Table 1 and expanded the four-view qualitative comparison in Figure 8.

(1) Task relationship and comparison scope

These methods address an important shared objective: generating appearance with reduced baked-in illumination. MaterialMVP generates PBR materials, including albedo and metallic–roughness maps, while UniTEX includes a separate delighting stage in its texture-generation pipeline. Our model learns to generate multi-view delighted RGB through SFT and reward-based post-training.

(2) Quantitative and visual comparisons

Selected rows from revised Table 1 (p. 6). Ours denotes the VLM-reward-trained variant. The full paper table also retains the other Ours variants.

MethodFID ↓SSIM ↑PSNR ↑L2 ↓R_de: ResNet ↑R_de: ViT ↑R_de: VLM ↑
SFT36.280.846522.340.11280.85270.70460.6813
Hunyuan3D27.540.864323.750.09350.89120.74280.7719
MaterialMVP19.760.847425.620.05930.82420.85740.8676
UniTEX32.760.851222.860.10670.87650.82530.8391
Material Anything25.630.871424.120.08790.79160.65210.6348
Ours12.380.906327.480.05140.91360.90680.9162

The VLM-reward-trained Ours outperforms the three additional methods on all seven measures. The revised PDF also includes additional visual comparisons in Figure 8, showing illumination removal and appearance preservation from multiple viewpoints.

(3) Human and VLM preferences

We have also added pairwise preference comparisons against MaterialMVP and UniTEX, evaluated by humans and GPT-5.5, with 20 objects per comparison.

Table 3 (p. 10). Percentage of cases in which Ours is preferred.

ComparisonEvaluatorPF ↑Consistency ↑Derived Overall ↑
Ours vs. MaterialMVPHuman70%65%70%
Ours vs. MaterialMVPGPT-5.575%70%75%
Ours vs. UniTEXHuman75%70%75%
Ours vs. UniTEXGPT-5.570%75%70%

Both humans and GPT-5.5 favor Ours in all reported comparisons. This directly extends the subjective evaluation to external methods.

Revision location: Table 1 (external baseline metrics), Figure 8 (multiview visual comparisons), Table 3 (external baseline preferences), Section 4.1 (experimental settings), Section 4.2 (quantitative comparison analysis), and Section 4.4 (preference analysis).

AuthorsOpenReview ↗

### W2. Human-evaluation setup and ranking protocol

Reviewer comment (excerpt):

The Human and VLM rankings in Fig.7 compares only internal variants, without external baselines. Details including the number of participants/objects and uncertainty are also missing.

(1) Participants and evaluated objects

The original study in Figure 7 evaluates Ours, Base, SFT, and Flow-GRPO on 20 objects. The analysis includes 10 valid human raters, all interns with computer-related backgrounds, providing 10 human evaluations per object. GPT-5.5 provides the VLM preference evaluation along the same dimensions.

(2) Presentation and evaluation dimensions

Participants viewed two methods’ three-view results side by side and selected their preference separately for delighting instruction following (PF) and non-lighting appearance consistency. Pairwise selection provides a direct comparison of illumination removal and appearance preservation.

(3) Ranking aggregation and Overall

For the original four-method ranking, each rater’s ranks 1–4 receive 4, 3, 2, and 0 points, respectively. Points are summed across raters for each object, and methods are ranked in descending order of total points for PF and consistency. Overall is computed as:

= \mathrm{PF}_{\mathrm{rank}} \times \left(0.4 + 0.6\,\mathrm{Consistency}_{\mathrm{rank}}\right).$$ The resulting scores are re-ranked in ascending order; lower component ranks indicate stronger preference. Section 4.1 and Appendix E now specify the participants, evaluation unit, presentation, ranking aggregation, and Overall rule. The added external-method preference results are reported in W1. **Revision location:** Section 4.1 (human evaluation setup), Appendix E (complete evaluation protocol), and existing Figure 7 (four-method preference rankings).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.