acceptodds
Under review as a conference paper at ICLR 2027

Flow3D-OPD: Multi-Teacher On-Policy Distillation for 3D Geometry Generation with Flow-Matching Diffusion Transformer

Abstract

Recent image-to-3D generation models built on flow-matching diffusion Transformers (DiT) can produce high-fidelity meshes, yet their post-training strategy remains largely unexplored. There exist several critical bottlenecks in reinforcement learning: the inherent difficulty of defining comprehensive rewards for 3D geometric quality, and the training instability that arises when jointly optimizing heterogeneous objectives. Inspired by the practicability of on-policy distillation (OPD) in large language models and image generation, we propose Flow3D-OPD, a two-stage post-training framework that introduces multi-teacher distillation into 3D geometry generation. In the first stage, we utilize the semi-policy to enhance the foundational capability of the pretrained model and then design an agentic verifier for 3D geometric quality evaluation. Based on the verifier, we cultivate domain-specialized teacher models via direct preference optimization (DPO). In the second stage, we consolidate heterogeneous expertise into a unified student model through on-policy distillation with a mixed divergence optimization strategy, which enhances training stability in joint optimization. Without relying on elaborate modifications, our straightforward yet effective design achieves consistent improvements across various geometric dimensions and surpasses all teacher models in the overall metric. Extensive experiments demonstrate that our approach provides an effective paradigm for reinforcement learning and on-policy distillation in 3D generation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D generation, Flow-matching, On-policy distillation

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.20 over 5 reviews, above 49% of the others.
Reviewer_bo7vOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Flow3D-OPD, a two-stage post-training framework for improving 3D geometry generation. The authors first employ semi-policy fine-tuning and train three specialized teachers for orientation, pose, and geometric consistency using agentic rewards and DPO. Subsequently, multi-teacher on-policy distillation with mixed divergence optimization is used to integrate their capabilities into a unified student model.

The experiments demonstrate improvements across multiple geometric quality dimensions compared with existing methods.

Critical strengths

1. Meaningful problem and practical reward design.

The paper addresses an important problem in 3D generation. The agentic reward system, incorporating multi-view rendering, zoom-in inspection, and case memory, provides a practical approach to evaluating fine-grained geometric quality.

2. Encouraging experimental results.

The proposed framework achieves improvements across several evaluation dimensions. The final student demonstrates more balanced performance than the individual specialized teachers.

Critical weaknesses

1. Limited methodological novelty.

My main concern is that the proposed method largely combines existing techniques, including DPO, multi-teacher on-policy distillation, and gradient accumulation. In particular, prior works such as Flow-OPD have already explored multi-teacher distillation for flow-matching models.

While extending these techniques to 3D generation is useful, the paper does not sufficiently clarify what fundamentally new methodological contributions are introduced beyond combining existing approaches.

I would appreciate further clarification of the specific novelty and how the proposed framework differs from existing OPD methods.

2. Insufficient justification for the complex training framework.

The proposed method involves multiple training stages, including semi-policy fine-tuning, three specialized teacher models, and student distillation. Although the overall framework achieves promising results, it remains unclear whether such a complex design is necessary.

In particular, the proposed Mixed Divergence Optimization appears closely related to conventional task-specific routing and gradient accumulation. The improvement over the simpler multi-step sequential strategy is relatively modest (61.0% vs. 60.4%).

Could simpler approaches, such as single-model multi-objective DPO or conventional multi-task distillation, achieve comparable performance? Stronger comparisons under matched computational budgets would help justify the necessity of the proposed framework.

3. Evaluation reliability.

The main multi-dimensional evaluation relies on an agentic verifier closely related to the reward system used during training. Although the authors provide additional GPT-5.5 and human evaluations, the improvement over the strongest teacher under GPT-5.5 is relatively small (51.4% vs. 50.9%).

Additional independent evaluations or statistical significance analysis would help strengthen the conclusions.

Questions for authors

The paper addresses a meaningful problem and presents a practical framework with encouraging empirical results. However, my main concerns are the limited methodological novelty and the insufficient justification for combining multiple existing techniques into a relatively complex training pipeline.

Although the complete framework performs well, the current experiments do not convincingly establish why this particular combination is necessary or what new methodological insights it provides beyond existing approaches.

I therefore lean toward rejection in the current form, but would be willing to reconsider my rating if the authors can adequately address these concerns during the rebuttal.

5 comments
AuthorsOpenReview ↗

Response to W1 & Q1

Our contribution is not limited to the final OPD stage. Rather, we propose a complete framework for post-training in 3D generation that addresses three challenges: the base model cannot roll out high-quality samples, no reliable reward is available, and multiple rewards must be properly integrated into a single student model. Specifically, we fine-tune the model with a semi-policy strategy in the first stage, which enables the base model to roll out high-quality samples and benefits the subsequent training. As shown in Table 1 in the paper, OPD from the pretrained model reaches only 57.3% B-wins, while ours reaches 61.0%. Second, for teacher training, we design an agentic VLM verifier that is aligned with how human experts judge 3D meshes, which is the key to training the teacher models effectively. Most importantly, in the OPD stage, previous methods do not assign each sample its own most suitable teacher. Their input data or text prompts are not clearly separated by the samples' own characteristics, so semantically similar samples may learn from multiple conflicting teacher velocity fields. In contrast, our method matches each input image with its preferred and aligned teacher: samples prone to orientation errors are supervised by the orientation teacher, and samples with complex human poses are distilled by the pose teacher. This allows each sample to learn the trajectory of its most suitable teacher, which is necessary in 3D generation. The following Table R6 compares our method with learning the velocity fields of all three teachers simultaneously on each sample. The results show that the previous strategy is inferior to our method in both B-wins and Uni3D-I, and its user-study score is even lower than that of FT+DPO. The relevant experiments have been added in section 5.3 of the revised version.

Table R6. Comparison with various alternatives: single-model multi-objective DPO (FT+DPO), previous mixture distillation strategy like DiffusionOPD, and our method.

A-wins ↓\downarrowB-wins ↑\uparrowUni3D-I ↑\uparrowUser study ↑\uparrow
FT + OPD6.657.20.3161.90
mixture distillation strategy5.758.80.3181.87
Ours4.961.00.3212.23
AuthorsOpenReview ↗

Response to W2

First, the overall metric in Table 1 of the paper shows that each module of our design progressively improves the generation quality. Moreover, it is not reasonable to consider our method unnecessary merely because its B-wins gain over the multi-step sequential strategy is modest, since that strategy itself already adopts our FT initialization and gradient-accumulated distillation. Regarding training complexity, our only additional time cost compared with FT+DPO is the distillation stage, which takes only about 1/3 of the time needed to train a single teacher because it requires no verifier calls. Regarding comparisons with simpler approaches, "FT + DPO" in Table 1 of the paper corresponds to single-model multi-objective DPO, and it shows a clear gap to Flow3D-OPD (48.3% vs. 61.0% B-wins). In addition, we implement conventional multi-task distillation, whose results are reported in Table R6 in our response to W1 and further demonstrate the effectiveness of our method.

AuthorsOpenReview ↗

Response to W3

The independent evaluation uses GPT-5.5 as the judge model. We found it to be relatively conservative: for every dimension, the spread of its A-wins and B-wins scores is smaller than that of Gemini-3-Flash. Nevertheless, our method still outperforms the best expert model on the overall metric. We further add comparisons with more judge models (e.g., GLM-5.3) and with geometric metrics, namely reconstruction metrics such as Chamfer Distance (CD) and F-score with a threshold of 0.001 on the GSO subset-200, as shown in the following Table R7. We also add a statistical significance analysis under Gemini-3-Flash. Our method achieves the best results under all of these evaluations. The relevant experimental details and explanations have been added in sections 4.2 and 4.3 of the Appendix.

Table R7. Additional independent evaluations. (1) A-wins / B-wins (%) by the agentic verifier with GLM-5.3 as the judge; (2) Reconstruction metrics against ground-truth meshes on the GSO subset-200; (3) Uncertainty analysis of Gemini-3-Flash over 5 repeated evaluations with different seeds.

MethodGLM-5.3 (A/B wins)CD ↓\downarrow (×10−4\times 10^{-4})F1-score (×10−4\times 10^{-4}) ↑\uparrowGemini-3-Flash (A/B wins, mean ±\pm std)
Orient-Teacher24.6 / 65.16942.7010.5 ±\pm 0.4 / 52.7 ±\pm 0.5
Pose-Teacher25.1 / 67.97672.5615.9 ±\pm 0.6 / 48.2 ±\pm 0.7
Consistency-Teacher22.7 / 69.06962.625.4 ±\pm 0.2 / 56.2 ±\pm 1.1
Flow3D-OPD (Ours)20.6 / 71.26432.794.8 ±\pm 0.3 / 60.6 ±\pm 0.6
Reviewer_QxfnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a post-training pipeline for improving orientation, human pose, and reference-image consistency in Hunyuan3D-2.1. Initial fine-tuning combines supervised velocity fitting with preferences between ground-truth and generated meshes. Three copies of this model then specialize through DPO, using preferences supplied by a multimodal verifier that can inspect additional mesh views and close-ups. A single student matches the specialists along its own denoising trajectories, with task losses accumulated before each update. The experiments compare the student with the specialists and other training configurations on three task datasets, supplemented by a separate model evaluator and human rankings.

Critical strengths
  • I find the verifier's targeted inspection the most convincing part of the 3D adaptation. A preference about pose or orientation can depend on a part of the mesh that is poorly exposed in the initial renders; requesting another view or a close-up lets the evaluator examine that geometry before choosing between candidates. The reported expert comparison favors adding these tools even after multiview input and case memory are available (Figure 3a), supporting their use in collecting geometric preferences for teacher training.

  • The student preserves most of the specialists' gains while substantially reducing the consistency losses of the orientation and pose teachers. Its win rate against the pretrained baseline exceeds FT + DPO on each of the three evaluation datasets (Table 1). These improvements are available in one generator with the original mesh decoding pipeline, without requiring the teachers or verifier at inference. Each specialist retains a higher win rate on its own criterion, so the practical achievement is a model that covers the three tested task distributions with fewer cross-task regressions.

Critical weaknesses
  • The paper does not establish what is new in its mixed divergence optimization. Flow-OPD [1] already routes each task to a specialist queried along the student's own trajectory, and DiffusionOPD [2] averages losses over a complete task cycle before updating the student. Eq. 10 uses that same accumulation procedure. The introduction's broader claim that RL post-training is absent from 3D generation also needs narrowing: DeepMesh [3] applies DPO to autoregressive mesh generation.

  • Routing each example to a single teacher does not ensure conflict-free gradients, because all task losses still update the same student parameters. When task gradients point in opposing directions, their average can cancel or increase one task's loss. The empirical results also give limited support to conflict reduction as the explanation for the improvement. In Table 4, adding mixed task routing improves B-wins by only 0.6 percentage points with multi-step accumulation, while it decreases B-wins with single-step updates. Other metrics improve in the latter comparison, and the full configuration performs best. This supports using the complete configuration, but the comparison does not establish that routing reduces conflict between task gradients.

  • The initial FT stage accounts for much of the pose and consistency improvement, but the experiment does not separate what it learns from ground-truth supervision and from preference optimization. Eq. 4 combines supervised velocity fitting with preferences between ground-truth meshes and policy samples. A supervised-only control using the same meshes would establish what the proposed semi-policy preference term adds. Assessing this stage also requires knowing what ground-truth supervision it receives: the source and size of the FT mesh set are unspecified, and Table 7 lists three task datasets without explaining their relation to those meshes.

  • It is unclear whether FT + DPO provides a comparison against direct learning of all three preferences under comparable conditions. The full student improves over both FT and FT + DPO on each task, but the paper does not explain how FT + DPO samples tasks or combines their preferences. This leaves the benefit of training separate teachers before distillation unresolved by that comparison.

  • Independent evaluation provides limited support for the student's quality advantage over the teachers. This matters because the main win rates use the same verifier that supplies training preferences. The ULIP and Uni3D results in Table 3 support improved global similarity over the pretrained generator, but provide neither teacher comparisons nor separate orientation or pose judgments. The preference evaluations in Table 2 have ambiguous model definitions: “FT + OPD” and “Flow3D-OPD (Ours)” appear as separate methods even though Table 1 identifies FT + OPD as the proposed pipeline. Taking “Ours” as the intended model, the GPT-5.5 judge gives it only a 0.5-percentage-point advantage over the consistency teacher in win rate against the pretrained baseline. Without uncertainty estimates, this provides little evidence of a reliable improvement over that teacher. The human ranking study favors the student, but its sample size and rank-to-score rule are unreported. The expert comparison could help validate the training verifier itself, yet it omits how annotation cases were constructed and separated from case memory, leaving its performance on new cases unclear.

  • The SDE loss weight has a sign error that affects the interpretation of the sampler comparison. The paper defines the denoising time increment as negative, but the weight following Eq. 9 and the second expression in Eq. 19 multiply that increment by a squared quantity. The resulting coefficient is negative, so minimizing the written loss would favor increasing the student–teacher velocity difference. The first expression in Eq. 19 has the correct positive sign. It is unclear which expression was implemented in the SDE experiment in Table 5. Until that is specified, the poorer SDE result cannot securely support the explanation that stochastic diffusion interferes with teacher alignment.

References:

[1] Flow-OPD: On-Policy Distillation for Flow Matching Models. Zhen Fang et al. arXiv, 2026. https://arxiv.org/abs/2605.08063

[2] DiffusionOPD: A Unified Perspective of On-Policy Distillation in Diffusion Models. Quanhao Li et al. SIGGRAPH Asia, 2026. https://arxiv.org/abs/2605.15055

[3] DeepMesh: Auto-Regressive Artist-mesh Creation with Reinforcement Learning. Ruowen Zhao et al. ICCV, 2025. https://arxiv.org/abs/2503.15265

Questions for authors
  1. What distinguishes “FT + OPD” from “Flow3D-OPD (Ours)” in Table 2?
  2. How are the three tasks combined in FT + DPO, and what sampling difference separates Seq. from MTR in Table 4?
  3. What are the sample sizes and annotation protocols for the expert comparison and human study, and how is case memory separated from evaluation?
  4. Which SDE loss weight was implemented for the comparison in Table 5?
8 comments
AuthorsOpenReview ↗

Response to W1

There are two existing problems of multi-teacher on-policy distillation in 3D generation: the limited capability of the base model and the gradient conflicts in distillation.

(1) For the first problem, we train the student starting from a semi-policy FT model rather than the pretrained model as in DiffusionOPD, which ensures the basic denoising capability of the student.

(2) For the second problem, in previous OPD image generators, the input content or text prompts are not clearly separated by the samples' own characteristics, which causes similar samples to learn from conflicting teacher velocity fields. In contrast, our method matches each input image with its specific teacher according to its characteristics: samples prone to orientation errors are supervised by the orientation teacher, and samples with complex human poses are distilled by the pose teacher. As shown in Table R8 below, we further compare our method with distilling multiple teachers on the similar sample, which is clearly worse than ours across various metrics. The relevant experiments have been added in section 5.3 of the revised version.

As for DeepMesh, it is essentially a mesh reconstruction task that generates meshes from point clouds, which falls outside the scope of our DiT-based image-to-3D generation task. Since its overall pipeline and key challenges differ from ours, we believe the introduction does not overstate our claim.

Table R8. Comparison between mixture distillation strategy and our routed distillation.

A-wins ↓\downarrowB-wins ↑\uparrowUni3D-I ↑\uparrowUser study ↑\uparrow
Mixture distillation strategy5.758.80.3181.36
Ours4.961.00.3211.64
AuthorsOpenReview ↗

Response to W2

We think the cancellation or increase of task losses mainly occurs when a single sample is matched with multiple denoising velocity fields of the teacher. Specifically, if the output vθv_\theta of the student is trained to minimize the average of ∥vθ−vt1∥22\Vert v_\theta - v_{t_1}\Vert_2^2, ∥vθ−vt2∥22\Vert v_\theta - v_{t_2}\Vert_2^2, and ∥vθ−vt3∥22\Vert v_\theta - v_{t_3}\Vert_2^2 on the same sample, its optimal target becomes (vt1+vt2+vt3)/3(v_{t_1} + v_{t_2} + v_{t_3})/3. If the three velocities fully conflict, e.g., vt1+vt2+vt3=0v_{t_1} + v_{t_2} + v_{t_3} = 0, this training target is obviously unreasonable.

However, under our hard teacher-routing mechanism, this issue could be alleviated because the student only needs to learn the velocity field of the corresponding teacher for every single sample. Within a mini-batch, the student minimizes ∥vθ1−vt11∥22+∥vθ2−vt22∥22+∥vθ3−vt33∥22\Vert v^1_\theta - v^1_{t_1}\Vert_2^2 + \Vert v^2_\theta - v^2_{t_2}\Vert_2^2 + \Vert v^3_\theta - v^3_{t_3}\Vert_2^2, where the vθ1,vθ2,vθ3v^1_\theta, v^2_\theta, v^3_\theta denote the student's velocity field in different samples. Even if the target velocities vt11v^1_{t_1}, vt22v^2_{t_2}, and vt33v^3_{t_3} point in opposite directions, this does not prevent the student from optimizing vθ1v^1_\theta, vθ2v^2_\theta, and vθ3v^3_\theta toward their respective targets. A comparison of the results for our task routing and multi-teacher hybrid distillation is provided in the response to W1. Furthermore, in Table 4 of the paper, we demonstrate that mixed task-routing sampling combined with multi-step updates is the superior strategy, which aligns with our proposed method.

AuthorsOpenReview ↗

Response to W3

The dataset for the fine-tuning (FT) stage consists of 3,164 samples (carefully collected and filtered by our human experts), which are kept strictly separate from the samples used in our DPO and evaluation stages.

In the FT stage, the two loss functions serve distinct purposes: the MSE loss guides each sampling step toward the GT velocity field, while the semi-policy loss enables the model to effectively distinguish between the velocity fields of "winning" and "losing" trajectories, thereby establishing a foundational capability for subsequent policy-based reinforcement learning (RL).

We adjust the relative weights, λsemi\lambda_{semi} and λmse\lambda_{mse}, to determine the respective contribution of each term. As shown in Table R9, the model achieves its best performance under λsemi=0.2\lambda_{semi}=0.2 and λmse=0.8\lambda_{mse}=0.8, which surpasses the supervised-only control (λmse\lambda_{mse}=1.0). The result confirms that the integration of these two loss functions is well-justified. The above comparison has been provided in Appendix 4.5.

Table R9. Effect of the weight of λsemi\lambda_{\mathrm{semi}} and λmse\lambda_{\mathrm{mse}}.

λsemi:λmse\lambda_{\mathrm{semi}} : \lambda_{\mathrm{mse}}ULIP-I ↑\uparrowULIP-T ↑\uparrowUni3D-I ↑\uparrowUni3D-T ↑\uparrow
1.0:0.01.0 : 0.00.1210.0730.3100.237
0.8:0.20.8 : 0.20.1270.0780.3160.246
0.2:0.80.2 : 0.80.1320.0820.3190.250
0.0:1.00.0 : 1.00.1340.0800.3160.248
Reviewer_Gk9QOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Flow3D-OPD is a multi-teacher post-training framework for image-to-3D geometry generation. The authors first fine-tune Hunyuan3D-2.1 with ground-truth supervision and semi-policy DPO. An agentic verifier with multi-view inspection and tool use then provides task-specific preferences to train three teachers through DPO for orientation, human pose, and geometric consistency. A student combines their capabilities through on-policy distillation with mixed divergence optimization, using velocity-field supervision along its own trajectories, task routing, and averaged losses accumulated across tasks.

Critical strengths
  • The agentic verifier can change viewpoints, zoom in, and inspect differences, matching the need for multi-view inspection of 3D objects. It may capture fine-grained errors missed by global similarity metrics such as ULIP.

  • 3D generation must balance orientation, human body structure, and reference-image consistency. Learning these capabilities separately and combining them into one model is well motivated and practically relevant.

  • The experiments demonstrate the method's effectiveness. Against the same baseline in Table 1, each teacher has higher B-wins than FT on its target dimension, but some other dimensions deteriorate. The student achieves a better overall balance, with additional support from an independent VLM and human evaluation.

Critical weaknesses
  • The 3D-specific designs mainly lie in the task definitions and agentic verifier. The paper's related work indicates that velocity-field distillation along student trajectories, the SDE-based KL derivation, and multi-teacher integration have precedents. Beyond applying these ideas to 3D, the new technical challenge addressed remains unclear. Much of the paper discusses mechanisms that may not be novel.

  • Mixed divergence optimization accumulates and averages losses across tasks, which may reduce frequent changes in update direction. How does this compare with the original training scheme using a smaller learning rate?

  • Shared FT, DPO for three teachers, and student distillation involve five trained model versions, raising concerns about complexity, runtime, and error accumulation across stages. Hunyuan3D-2.1 uses one DiT; extending the procedure to both DiTs in a two-stage model such as TRELLIS could further increase complexity. Do the gains justify this overhead?

Questions for authors
  1. Table 1 labels the method “FT+OPD (Ours),” while Table 2 lists both “FT+OPD” and “Flow3D-OPD (Ours)” with different results. How do their configurations differ? Please clarify and use consistent names.

  2. The paper defines Δtj=tj+1−tj<0\Delta t_j=t_{j+1}-t_j<0, but the rightmost expression in Eq. (19) and the weight following Eq. (9) use +Δtj+\Delta t_j, inconsistent with the earlier −Δtj-\Delta t_j term and KL non-negativity. Is a minus sign missing? Which expression was used in the SDE experiments?

  3. My main concerns are the technical novelty and the cost of the multi-teacher pipeline. A clearer explanation of what the method contributes beyond adapting existing OPD techniques to 3D, together with evidence that its gains justify the additional training cost over simpler alternatives such as joint DPO, would help address these concerns. I would consider raising my score if the rebuttal makes a convincing case on both points.

7 comments
AuthorsOpenReview ↗

Response to W1 & W2 & Q3

Our contribution is a post-training framework for 3D generation that addresses three challenges: low-quality rollouts of the base model, the lack of a reliable reward, and the integration of multiple rewards into a single student. First, semi-policy fine-tuning enables the base model to roll out high-quality samples for subsequent training. Second, an agentic VLM verifier aligned with human expert judgement makes effective teacher training possible. More importantly, in the OPD stage, we curate the training images by their typical failure modes and match each image with its specific teacher. This is necessary for OPD in 3D generators, because if similar images appear in the training data of different teachers, the student will receive conflicting learning targets. We additionally compare the performance of FT+DPO, the previous mixture distillation strategy, and our final setting in Table R5. The user-study score of the previous mixture distillation strategy even falls below that of FT+DPO, which confirms that assigning each sample its most suitable teacher is necessary. The relevant experiments have been added in section 5.3 of the revised version

The comparison with the original training scheme using a smaller learning rate is also shown in Table 4 of the paper, denoted as ``Single-step + MTR''. This scheme achieves only 6.6% A-wins and 58.5% B-wins, which is clearly inferior to our method (4.9% A-wins and 61.0% B-wins).

Table R5. Comparison with various alternatives: single-model multi-objective DPO (FT+DPO), previous mixture distillation strategy, and our method.

A-wins ↓\downarrowB-wins ↑\uparrowUni3D-I ↑\uparrowUser study ↑\uparrow
FT+OPD6.657.20.3161.90
mixture distillation strategy5.758.80.3181.87
Ours4.961.00.3212.23
AuthorsOpenReview ↗

Response to W3 & Q3

We thank the reviewer for the concerns about runtime and training complexity, and clarify them as follows. (1) Although our pipeline involves five models, only one student is deployed, and its inference cost is identical to the base model (a single DiT with 50 NFE). (2) Training the three teachers independently with DPO costs about the same as training a single model sequentially with multi-reward DPO. In both cases, the main cost comes from pair rollouts and agentic verifier calls (about 7 minutes per step). Our only additional cost is distilling the teachers into the student. This stage requires no verifier calls, and thanks to the dense supervision and the mixed divergence optimization, it takes only about 1/3 of the time needed to train a single teacher. The corresponding clarification has been provided in Appendix 1.1. (3) Errors do not accumulate across stages: the teachers and the student are all initialized from the same FT model, and during OPD the teachers are frozen, and each supervises only its own dimension. For example, the orientation teacher degrades consistency (30.8% A-wins), but the student does not inherit this degradation (10.0% A-wins). (4) The gains are clear: compared with FT+DPO, Flow3D-OPD improves B-wins by +3.8 (Gemini) and +6.7 (GPT-5.5), and the user-study score from 3.62 to 3.91.

AuthorsOpenReview ↗

Response to Q1

We apologize that ‘FT+OPD' in the second-to-last row of Table 2 is a typo and should be revised to ’FT+DPO'. The last row, `Flow3D-OPD (Ours)' is the exact FT+OPD setting. We have corrected this in the revision.

Reviewer_XTZNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Flow3D-OPD combines semi-policy fine-tuning, geometry-specific teacher training, and multi-teacher on-policy distillation for image-to-3D generation. A tool-augmented verifier evaluates orientation, pose, and consistency. Experiments report improved aggregate performance, supported by embedding metrics, independent judging, and a user study.

Critical strengths
  1. The multiview, tool-augmented verifier provides useful geometry-specific supervision, with improved agreement with expert judgments.
  2. The student achieves a better balance across quality dimensions than individual teachers, with improvements supported by several evaluation protocols.
Critical weaknesses
  1. The methodological distinction from DiffusionOPD is unclear, particularly task-wise loss accumulation. The ablations do not sufficiently isolate the proposed optimization contribution.
  2. Independent evaluation shows only a small B-win advantage over the strongest teacher. Missing uncertainty estimates and human-study details weaken the evidence for consistent superiority.
Questions for authors
  1. How does Equation (10) differ from DiffusionOPD, and are the Table 4 comparisons matched for effective batch size and training budget?
  2. What distinguishes “FT + OPD” from “Ours” in Table 2? Please clarify evaluation sample sizes, human-study scoring, and uncertainty.
  3. Since (\Delta t_j<0), should the SDE weight in Equation (19) use (-\Delta t_j)? Which expression was implemented?
6 comments
AuthorsOpenReview ↗

Response to W1

Response to W1: Considering that the meshes generated in 3D generation are of relatively low quality, and that the types of mesh defects are strongly correlated with the object categories in the input image, our differences from DiffusionOPD lie in the following aspects. (1) Where to distill. DiffusionOPD initializes the student from the pretrained model, whereas we additionally use a semi-policy strategy to fine-tune the model, which enables the basic model to rollout the high-quality samples and benefits the following training. As shown in Table 1, OPD from the pretrained model reaches only 57.3% B-wins, while ours reaches 61.0%. (2) How to distill. The input content or text prompts in image generators like DiffusionOPD are not clearly separated, so similar samples may learn from conflicting teacher velocity fields. In contrast, our method matches each input image with its preferred and aligned teacher: samples prone to orientation errors are supervised by the orientation teacher, and samples with complex human poses are distilled by the pose teacher. The following Table R3 compares the 3D generation metric between FT+DPO, distillation strategies like DiffusionOPD, and our setting. Without such specific sample-teacher matching, the user-study score of the Diffusion strategy even falls below that of FT+DPO, which confirms that assigning each sample its most suitable teacher is necessary in 3D generation. We have added relevant experiments in section 5.3 of the revised version.

Table R3. Comparison with various alternatives: single-model multi-objective DPO (FT+DPO), previous mixture distillation strategy like DiffusionOPD, and our method.

A-wins ↓\downarrowB-wins ↑\uparrowUni3D-I ↑\uparrowUser study ↑\uparrow
FT + DPO6.657.20.3161.90
mixture distillation strategy5.758.80.3181.87
Ours4.961.00.3212.23
AuthorsOpenReview ↗

Response to W2

(1) The independent evaluation uses GPT-5.5 as the judge model. We found it to be relatively conservative: for every dimension, the spread of its A-wins and B-wins scores is smaller than that of Gemini-3-Flash. Nevertheless, our method still outperforms the best expert model on the overall metric. We further add comparisons with more judge models (e.g., GLM-5.3) and more geometric metrics (reconstruction metrics such as Chamfer Distance (CD) and F-score with a threshold of 0.001 on the GSO subset-200 [1]), as shown in Table R4. (2) For the user study, our evaluation criteria and protocol are more comprehensive and stricter. Specifically, the evaluation is conducted by 3D research experts along five dimensions: macro-structure adherence, geometric detail adherence, smoothness, rationality of inferred structure, and multi-view consistency. The meshes generated by all N evaluated models are shuffled so that the experts conduct a blind test. The experts then rank the generated meshes from best to worst and assign descending scores from N to 1. Finally, we calculate the average score of each model. (3) In the right of Table R4, we further report uncertainty estimates for the Gemini-3-Flash evaluation. Specifically, the agentic verifier scores the outputs on the test set 5 times with different random seeds and sampling temperatures, and we report the mean and standard deviation. Our student model still achieves the best mean score across the repeated evaluations. The relevant experimental details and explanations have been added in sections 4.2 and 4.3 of the Appendix.

Table R4. Additional independent evaluations. (1) A-wins / B-wins (%) by the agentic verifier with GLM-5.3 as the judge; (2) Reconstruction metrics against ground-truth meshes on the GSO subset-200; (3) Uncertainty analysis of Gemini-3-Flash over 5 repeated evaluations with different seeds.

MethodGLM-5.3 (A/B wins)CD ↓\downarrow (×10−4\times 10^{-4})F1-score (×10−4\times 10^{-4}) ↑\uparrowGemini-3-Flash (A/B wins, mean ±\pm std)
Orient-Teacher24.6 / 65.16942.7010.5 ±\pm 0.4 / 52.7 ±\pm 0.5
Pose-Teacher25.1 / 67.97672.5615.9 ±\pm 0.6 / 48.2 ±\pm 0.7
Consistency-Teacher22.7 / 69.06962.625.4 ±\pm 0.2 / 56.2 ±\pm 1.1
Flow3D-OPD (Ours)20.6 / 71.26432.794.8 ±\pm 0.3 / 60.6 ±\pm 0.6

Reference

[1] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B. McHugh, and Vincent Vanhoucke. Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items. In \emph{2022 International Conference on Robotics and Automation (ICRA)}, pp. 2553–2560, 2022.

AuthorsOpenReview ↗

Response to Q1

The differences between our method and DiffusionOPD are detailed in our response to W1. In Table 4, all settings use the same batch size, number of forward passes, and total number of training samples to ensure a fair comparison. The only differences are the backward frequency and the data sampling order, which are exactly the factors our ablation is designed to study.

Reviewer_R85HOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a novel 3D post-training framework (Flow3D-OPD), which effectively addresses the challenges in the current 3D generation field, such as difficulty in quality assessment, instability in optimization, and weak foundation models. It provides innovative and experimentally validated methods. The ablation experiment design is comprehensive, including independent evaluation to prevent overfitting, and combines human-computer interaction research.

Critical strengths
  1. First to introduce multi-teacher on-policy distillation (OPD) into a two-stage post-training framework for 3D geometry generation, where the first stage utilizes domain-specific teacher model training and the second stage employs multi-teacher on-policy distillation.
  2. Addressing the weak generative capability of the base model, semi-policy fine-tuning is introduced at the beginning of the first stage. This design ensures the target velocity field is optimal, requiring only a single denoising process to obtain negative samples, effectively enhancing the capability of the base model.
  3. Demonstrating comprehensive and robust metric advantages across multiple evaluation paradigms, focusing on fine-grained optimization for specific dimensions.
Critical weaknesses
  1. The paper proposes mixed divergence optimization to mitigate instability in multi-objective optimization. However, this is essentially a heuristic engineering strategy, and it does not provide a proof that heterogeneous objectives (orientation, pose, and consistency) do not conflict with one another in parameter space. Simply accumulating and averaging these divergence values may merely mask gradient conflicts, causing the student model to converge only to a sub-optimal solution.
  2. The current framework relies on exactly three domain-expert teachers. Whether there is an upper bound on the gains from adding more teachers, which could lead to a decline in overall performance, is not discussed in the paper.
  3. Although the agentic verifier employs multi-view rendering and difference overlay techniques, a VLM fundamentally evaluates the plausibility of 2D pixels rather than true 3D geometric properties. As long as defects are not visible in the rendered 2D views, correct evaluation is impossible, and there will inevitably be a gap between the 2D-rendering-based agent reward and the true 3D geometric quality.
Questions for authors
  1. [Evaluation Gap] To what extent does the agentic verifier evaluate genuine 3D geometric integrity versus 2D rendering plausibility? Given that VLMs inherently process 2D projections, how does the framework mitigate the risk of the model optimizing for 2D pixel-level coherence rather than achieving true 3D structural correctness?
  2. [Gradient Conflicts and Scalability] What empirical or theoretical mechanisms are in place to guarantee that the heterogeneous objectives of multiple teacher models do not induce gradient conflicts during joint distillation? Furthermore, does this multi-teacher paradigm exhibit a performance saturation point, where the introduction of additional specialized teachers might induce negative transfer and degrade the student model's overall performance?
  3. [Computational Overhead] Given the reliance on an agentic verifier for pairwise quality assessment during the training phase (iterative invocation for multi-round reasoning and tool usage), what is the computational overhead of this approach?
5 comments
AuthorsOpenReview ↗

Response to W1

In our method, each sample learns the velocity field of only one teacher rather than those of all three teachers simultaneously. As described in Section 4.2.2 of the paper, this is the core of how we address conflicts in parameter space. The divergences of different samples are then accumulated within a mini-batch before a single update, so that each update carries clear guidance from all dimensions. In contrast, forcing the student to learn the average of the three teachers' velocities on the same sample would genuinely cause gradient confusion. As shown in Table R1 below, we further compare our method with per-sample multi-teacher averaging, which is clearly worse than ours across various metrics. We have added relevant experiments in section 5.3 of the revised version.

Table R1. Comparison between per-sample mixture averaging distillation strategy and our routed distillation.

A-wins ↓\downarrowB-wins ↑\uparrowUni3D-I ↑\uparrowUser study ↑\uparrow
mixture distillation strategy5.758.80.3181.36
Ours4.961.00.3211.64
AuthorsOpenReview ↗

Response to W2 & Q2

When the rewards of different teachers do not contradict each other, adding a teacher raises the upper bound of the student's capability. To verify this, we introduce a new reward that focuses on the geometric details of the meshes, train a corresponding teacher with this reward, and include it in the distillation. As shown in Table R2, adding a new and reasonable teacher still brings further gains. This is because, under our algorithmic strategy, adding a teacher amounts to adding new training data. Since each sample is supervised only by its corresponding teacher, the student's upper bound on each dimension depends only on that dimension's teacher. We have added relevant explanations in Appendix 4.4.

Table R2. Multi-dimension and user study evaluation with four teacher models distillation.

MethodOrient-Eval (133)Pose-Eval (111)Consist-Eval (100)Detail-Eval (100)User study ↑\uparrow
Orient-Teacher2.7 / 22.70.8 / 89.030.8 / 53.275 / 253.07
Pose-Teacher10.1 / 14.30.0 / 93.741.7 / 41.747 / 533.02
Consistency-Teacher6.5 / 15.91.0 / 83.58.0 / 82.070 / 302.52
Detail-Teacher5.2 / 16.30.9 / 92.414.0 / 72.028 / 723.10
Flow3D-OPD (Ours)4.9 / 19.20.0 / 93.79.0 / 81.031 / 693.29
AuthorsOpenReview ↗

Response to W3 & Q1

In the image-to-3D generation task, the generated 3D mesh can only be judged against the input RGB image, since there is no ground-truth 3D mesh that allows a fully accurate comparison, which remains an open problem in this field. Previous methods such as Uni3D and ULIP measure this correspondence only with a global similarity score. In contrast, our VLM judge automatically renders new views, which allows it to evaluate unseen views, and its memory mechanism keeps its geometric judgements aligned with human expert experience. Notably, human experts also evaluate 3D meshes based on rendered 2D visual information, by rotating and zooming the model in 3D software like Blender. Moreover, we compare the accuracy of the VLM judge on human-annotated pairs under different settings, as shown in Fig. 3(a) of the paper. The single-view judge reaches only 66.7% accuracy, which rises to 75.8% with multi-view inputs and to 84.3% once the case-memory and tool-call mechanisms are added. This shows that each component of our verifier brings its judgements closer to those of human experts.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.