Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Nabla-R2D3: Effective and Efficient 3D Diffusion Alignment with 2D Rewards

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read 2D rewards suffice to align 3D diffusion models

desk verdict Worth a serious review, but the abstract overclaims and the actual loss is not the one with the proved unbiasedness; both are fixable. read the letter →

arxiv 2506.15684 v1 pith:X2AUL54G submitted 2025-06-18 cs.GR cs.CVcs.LG

Qingming Liu , Zhen Liu , Dinghuai Zhang , Kui Jia This is my paper
classification cs.GRcs.CVcs.LG
keywords 3DdiffusionmodelsrewardfinetuningGFlowNetshumanpreferencealignment2Drewardsmulti-viewrenderingtext-to-3Dnormal-mapconsistency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Finetuning 3D-native diffusion models to follow human preferences has been held back by the lack of 3D reward models. This paper claims that 2D reward models—which are abundant—are enough, provided the 3D reward is defined as the average of a 2D reward over randomly rendered camera views, a standard lifting assumption from text-to-3D work. The authors adapt the Nabla-GFlowNet gradient-consistency objective to this setting and show that within a few hundred finetuning steps it raises reward values while preserving the pretrained distribution better than policy-gradient (DDPO) and direct backprop (ReFL, DRaFT) baselines, without the multi-view inconsistency artifacts those baselines exhibit. A simplified loss that drops the learned correction term is reported to work well, which is both the method's practical edge and its main assumption. If right, this makes RLHF-style alignment of 3D generators practical using existing 2D preference and geometry signals.

What carries the argument

The carrying object is the Nabla-GFlowNet gradient-consistency loss, a score-matching objective that aligns the gradient of the log-density ratio between the finetuned and base models with β times the gradient of the log-reward, plus a learned correction term h_phi. The paper adapts it to 3D by defining the 3D reward as the average of a 2D reward over rendered camera views, and then uses the simplified forward loss L_approx, which keeps only the score-matching term with a stop-gradient reward-gradient estimate and drops h_phi. This loss is applied along the denoising trajectory of a 3D-native diffusion model using low-rank adaptation (LoRA) parameter updates.

What would settle it

Run the same finetuning with the full Nabla-GFlowNet objective, including the reverse-direction loss and the learned correction term h_phi, under identical seeds and hyperparameters; the simplified loss should match its reward-versus-FID tradeoff, and the learned h_phi should be small. A more direct check is to estimate the per-view reward distribution of samples from the finetuned model and compare it against the base model tilted by R^β: a large mismatch would show that the educated-guess assumption failed.

Watch

Extended reading notes

Core claim

Nabla-R2D3 claims to be the first method that effectively aligns 3D-native diffusion models with 2D-only reward signals. It builds the 3D reward as the expectation over camera poses of a differentiable 2D reward applied to rendered views, and finetunes the 3D model with the Nabla-GFlowNet loss that matches the log-density-ratio score to the reward gradient, regularized by a stop-gradient reward-gradient guess and a prior-preservation term. The paper's key empirical claim is that an approximate version of that loss, which discards the learned correction term and the reverse-direction loss, suffices: it yields higher rewards, lower multi-view FID, and higher text-object CLIP similarity than DDPO, ReFL, DRaFT, and 3D-SDS across appearance rewards (Aesthetic Score, HPSv2) and geometry rewards (normal-map consistency, depth-normal consistency), on both DiffSplat and GaussianCube base models.

Load-bearing premise

The load-bearing assumption is that the guessed reward-gradient (the stop-gradient estimate through the predicted clean sample) is accurate enough that the learned correction term can be dropped; if that guess is biased, the finetuning objective no longer samples from the intended reward-weighted distribution.

Editorial extensions

If this is right

  • 3D-native diffusion models can be aligned to human preferences without any 3D-labeled reward data; the reward model only needs to score rendered 2D views.
  • Because the generative model itself is finetuned, alignment is amortized: generating a well-aligned asset takes seconds, compared with minutes per object for SDS-style lifting approaches.
  • Geometry-only signals such as normal-map consistency and depth-normal consistency can steer 3D generation, suggesting that alignment can target structure rather than appearance alone.
  • In the paper's experiments, the simplified loss yields a better reward-versus-prior-preservation tradeoff than vanilla finetuning baselines, so reward finetuning of 3D models does not require a learned correction network.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the educated-guess assumption transfers elsewhere, the same simplified loss could make gradient-informed RL finetuning cheaper for other generative models wherever a differentiable reward-gradient estimate is available.
  • A testable extension would weight camera views by predicted reward variance instead of sampling uniformly, potentially reducing the reported variance and further stabilizing 3D alignment.
  • The paper's failure cases show the 2D normal estimator can hallucinate geometry; combining appearance and geometry rewards in one finetuning run, which the paper does not explore, could mitigate such hallucinations.
  • Because the 3D reward inherits the 2D reward's biases, per-view fairness auditing would be a natural next check; the paper does not report per-view reward distributions.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces Nabla-R2D3, an adaptation of the Nabla-GFlowNet framework to finetune 3D-native diffusion models using 2D reward signals. The method defines a 3D reward as an expectation of 2D rewards over camera views, derives forward and reverse losses, and then in practice replaces them with a single approximate forward loss (Eq. 9). Experiments on DiffSplat and GaussianCube with Aesthetic Score, HPSv2, normal-map, and depth-normal-consistency rewards are reported, with comparisons to DDPO, ReFL, and DRaFT.

Significance. If the claims are correct, the method provides a sample-efficient way to align 3D generative models with human preferences using only 2D rewards, addressing a practical gap in 3D asset generation. The experimental scope is broad, covering two base models and four reward types, and the paper includes qualitative and quantitative evaluations. However, the two central pillars of the contribution—the principled derivation and the empirical superiority—are weakened by the gap between the derived loss and the actually optimized loss, and by the overstatement of the reward and prior-preservation results. The paper has the potential to be a useful contribution after a major revision.

major comments (3)
  1. [Abstract, §5.2, Tables 1–2] The abstract and Section 5.2 state that Nabla-R2D3 'consistently achieves higher rewards' than vanilla finetuning baselines, but this is contradicted by the paper's own quantitative results. In Table 1, DRaFT achieves a higher HPSv2 reward (32.65) than Nabla-R2D3 (27.85). In Table 2, DRaFT achieves higher Aesthetic Score rewards on both DiffSplat-SD1.5 (6.40 vs 6.02) and GaussianCube (6.04 vs 5.92). The claim of 'reduced prior forgetting' is also not uniformly supported: DDPO has lower (better) multi-view FID than Nabla-R2D3 in all three settings of Table 1 (172.95 vs 217.89, 69.59 vs 131.38, 63.93 vs 104.45) and on DiffSplat-SD1.5 in Table 2 (54.79 vs 154.61). Please temper these claims to match the reported numbers or provide a secondary criterion (e.g., Pareto optimality or statistical significance) that reconciles them.
  2. [§4.1, Eq. (9); Appendix A] The derivation in Appendix A applies to the full Nabla-GFlowNet loss in Eq. (7), which includes the learned correction term h_phi and the reverse loss. In contrast, all experiments use L_approx from Eq. (9), which drops both h_phi and the reverse loss. The paper asserts that the educated guess of the reward gradient is accurate, but this is an unsupported assumption: no proof or ablation is provided to show that Eq. (9) targets the same tilted reward distribution or preserves the convergence guarantees of the full loss. This is load-bearing because the paper's claim that the method is 'built upon Nabla-GFlowNet in a principled manner' refers to an objective that is not the one optimized.
  3. [Appendix A] The proof labeled 'Proof of Unbiasedness' does not actually establish unbiasedness. It shows that the gradient of the loss in Eq. (12) equals the gradient of L_forward, which is an algebraic identity rather than a statement about the estimator's expectation. To prove unbiasedness in the sense claimed, the derivation would need to relate Eq. (12) or Eq. (9) to the gradient of a well-defined target objective, such as the KL divergence to the tilted reward distribution. As written, the proof is circular and provides no theoretical support for the method beyond restating the loss.
minor comments (5)
  1. [Figure 4 caption] The caption in Figure 4 says 'Quantitative comparison' but the figure shows qualitative renderings with per-object reward values; it should be 'Qualitative comparison'.
  2. [Eq. (9)] Equation (9) contains an extra closing parenthesis in the reward-gradient term: logR(g(ˆzθ(zt−1), c))) should have two closing parentheses after c, not three.
  3. [§5.1, base models] The model name is spelled inconsistently: 'DiffSplat-PixArt-Σ' and 'DiffSplat-Pixart-Σ' are used in different places; please standardize.
  4. [§5.2, 3D-SDS comparison] In the paragraph comparing with 3D-SDS, the phrase 'wherear inference with the finetuned model takes only around 8 seconds' contains the typo 'wherear'; it should read 'whereas'.
  5. [Appendix C] The implementation details state that training takes no more than one day but do not report the number of finetuning steps, the number of trajectories, or the evaluation checkpoints; please include these details for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Nabla-R2D3 is a direct application of the authors' prior Nabla-GFlowNet loss to a defined multi-view 2D reward, with the actual training loss being an explicitly acknowledged simplification rather than a disguised restatement of the inputs.

full rationale

The derivation chain in this paper is not circular. The 3D reward is defined in Eq. (5) as an expectation of 2D rewards over camera views, and Eqs. (7)-(8) are obtained by substituting that reward into the Nabla-GFlowNet loss from the cited prior work [25]. This is a straightforward application of a published method, not a reduction of the conclusion to the premise. The paper then states in Eq. (9) that it uses an approximate loss that drops the learned correction term h_phi, explicitly saying 'we assume that our educated guess of the reward gradient is accurate and therefore there is no need to learn the correction term h_phi anymore.' This is an openly stated modeling assumption and a potential correctness gap, but it is not circular: the approximate loss is not claimed to follow from the unbiasedness proof, and the paper does not disguise it as a fitted parameter or a prediction derived from the earlier equations. The empirical evaluation is also independent of the self-citation: results are compared against external baselines (DDPO, ReFL, DRaFT) using external reward models (Aesthetic Score, HPSv2, StableNormal) and external metrics (FID, CLIP-Sim). The disagreement between the abstract's word 'consistently' and some table entries is an overstatement of results, not a circularity. The paper's own limitations section and failure cases further indicate that the claims are not made true by construction. No circular step meeting the evidentiary bar is present.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The ledger is small: the method introduces no new entities and reuses a 3D reward defined as an expectation over pretrained 2D rewards. The main burden is the unproven L_approx approximation and the hand-chosen temperature and regularization strengths.

free parameters (2)
  • beta (reward temperature) = 2e6 (HPSv2), 1e7 (Aesthetic), 1e6 (geometry)
    Hand-chosen per reward model; controls the trade-off between reward and prior. Sensitivity is shown in Fig. 14, confirming the choice affects convergence and prior preservation.
  • lambda (gradient regularization weight) = 3e3 (Aesthetic), 5e3 (HPSv2), 1e4 (geometry)
    Hand-chosen weight for Lreg that keeps the updated noise prediction close to the previous step; no selection procedure is given, and results may depend on it.
assumptions (4)
  • domain assumption 3D reward equals expectation of 2D rewards over sampled camera views (Eq. 5)
    Common in text-to-3D lifting; assumes multi-view rendering captures all relevant 3D preferences and ignores 3D-only attributes such as physical plausibility or inner structure.
  • domain assumption The renderer g is differentiable and gradient of the 2D reward propagates meaningfully into the 3D latent z
    Used in Eqs. 7-9; for Gaussian splatting and NeRF, approximate volumetric rendering is needed, and the resulting gradient bias is not quantified.
  • ad hoc to paper The stop-gradient reward-gradient estimate in L_approx is accurate enough to discard the learned correction term h_phi
    Explicit assumption in Sec. 4.1 after Eq. 9; no proof or ablation isolating h_phi is provided, and this is the weakest link in the derivation.
  • domain assumption Pretrained 2D models (LAION Aesthetic, HPSv2, StableNormal) are valid reward proxies for 3D content
    The paper adopts these models without validation on 3D assets; failure cases in Sec. H show the normal estimator can hallucinate geometry, undermining the reward signal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Nabla-R2D3: Effective and Efficient 3D Diffusion Alignment with 2D Rewards." pith.science (2026). https://pith.science/paper/X2AUL54G

@misc{pith2026250615684,
  author       = {Pith},
  title        = {Pith review of: Nabla-R2D3: Effective and Efficient 3D Diffusion Alignment with 2D Rewards},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X2AUL54G}},
  note         = {Machine review of arXiv:2506.15684}
}
read the original abstract

Generating high-quality and photorealistic 3D assets remains a longstanding challenge in 3D vision and computer graphics. Although state-of-the-art generative models, such as diffusion models, have made significant progress in 3D generation, they often fall short of human-designed content due to limited ability to follow instructions, align with human preferences, or produce realistic textures, geometries, and physical attributes. In this paper, we introduce Nabla-R2D3, a highly effective and sample-efficient reinforcement learning alignment framework for 3D-native diffusion models using 2D rewards. Built upon the recently proposed Nabla-GFlowNet method, which matches the score function to reward gradients in a principled manner for reward finetuning, our Nabla-R2D3 enables effective adaptation of 3D diffusion models using only 2D reward signals. Extensive experiments show that, unlike vanilla finetuning baselines which either struggle to converge or suffer from reward hacking, Nabla-R2D3 consistently achieves higher rewards and reduced prior forgetting within a few finetuning steps.

Figures

Figures reproduced from arXiv: 2506.15684 by the authors.

Figure 1
Figure 1. Our Nabla-R2D3 can efficiently and robustly finetune 3D-native diffusion models with differentiable reward models learned from human preferences on appearance, geometry and many other attributes. Abstract Generating high-quality and photorealistic 3D assets remains a longstanding chal￾lenge in 3D vision and computer graphics. Although state-of-the-art generative models, such as diffusion models, have made significan… view at source ↗
Figure 2
Figure 2. Convergence curves of metrics for different finetuning methods on Aesthetic Score. Our method [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison of 3D assets produced by models finetuned with different methods on Aesthetic [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (16 more)
Figure 4
Figure 4. Figure 4: Quantitative comparison on HPSv2 [45]. For each object we present the front and back views. A three-tier cake with icing and sprinkles. A Chinese Sakura bonsai tree sculpture with pink flowers on a wooden base. Prompt Pretrained Model Ours DiffSplat-SD1.5 GaussianCube …
Figure 5
Figure 5. Figure 5: Finetuning results on two base models, DiffSplat-SD1.5 [ [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: 360◦ visualization of 3D shapes generated by models finetuned with Nabla-R2D3 and DRaFT. DRaFT-finetuned model is prone to overfitting and suffers from the Janus problem. Prompt ReFL 91.26 DRaFT 91.59 Ours 93.81 A blue character, possibly a bear or robot, standing on a…
Figure 7
Figure 7. Figure 7: Quantitative comparison of different finetuning methods on the normal estimator reward (Eqn. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 9
Figure 9. Figure 9: Comparison with 2D-SDS-based align￾ment method [50]. DNC reward. In [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: Visual evolution of the generated object with [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 12
Figure 12. Figure 12: Qualitative comparison with MVDream. For MVDream, we show the multi-view images directly [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]
Figure 13
Figure 13. Figure 13: Qualitative comparison with pretrained model on DNC reward. [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]
Figure 14
Figure 14. Figure 14: The relationship between the temperature parameter [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]
Figure 15
Figure 15. Figure 15: Convergence curves of metrics on different learning rate. [PITH_FULL_IMAGE:figures/full_fig_p017_15.png]
Figure 16
Figure 16. Figure 16: Trade-offs among reward maximization, prior preservation, and text-object alignment for various [PITH_FULL_IMAGE:figures/full_fig_p017_16.png]
Figure 17
Figure 17. Figure 17: More qualitative results on Aesthetic Score. [PITH_FULL_IMAGE:figures/full_fig_p018_17.png]
Figure 18
Figure 18. Figure 18: More qualitative results on HPSv2. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_18.png]
Figure 19
Figure 19. Figure 19: More qualitative results on the normal estimator reward. [PITH_FULL_IMAGE:figures/full_fig_p020_19.png]
Figure 20
Figure 20. Figure 20: Failure cases of the normal estimator reward. Left: rendered RGB image, Middle: rendered normal [PITH_FULL_IMAGE:figures/full_fig_p021_20.png]
Figure 21
Figure 21. Figure 21: Failure cases of the Aesthetic Score (first row) and HPSv2 (second row). The left column shows [PITH_FULL_IMAGE:figures/full_fig_p021_21.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Edit in 2D, Verify in 3D: Reinforcement Learning for Multi-view Consistent Scene Editing

    cs.CV 2026-03 conditional novelty 6.0 of 10

    RL3DEdit fine-tunes FLUX-Kontext with GRPO using VGGT confidence and pose rewards to produce multi-view consistent 3D scene edits in a single pass.

Reference graph

Works this paper leans on

63 extracted references · 48 canonical work pages · cited by 1 Pith paper

  1. [1]

    GFlowNet foundations.Journal of Machine Learning Research, 2023

    Yoshua Bengio, Salem Lahlou, Tristan Deleu, Edward J Hu, Mo Tiwari, and Emmanuel Bengio. GFlowNet foundations.Journal of Machine Learning Research, 2023. 3

    work page 2023

  2. [2]

    Training diffusion models with reinforcement learning

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. InICLR, 2024. 2, 3, 5

    work page 2024

  3. [3]

    PixArt- Σ: Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation

    Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt- Σ: Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation. InECCV, 2024. 5, 15

    work page 2024

  4. [4]

    Kevin Clark, Paul Vicol, Kevin Swersky, and David J. Fleet. Directly fine-tuning diffusion models on differentiable rewards. InICLR, 2024. 2, 5

    work page 2024

  5. [5]

    Objaverse: A universe of annotated 3d objects

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. In CVPR, 2023. 5

    2023

  6. [6]

    Carles Domingo-Enrich, Michal Drozdzal, Brian Karrer, and Ricky T. Q. Chen. Adjoint matching: Fine- tuning flow and diffusion generative models with memoryless stochastic optimal control. InICLR, 2025. 5, 14

    work page 2025

  7. [7]

    Gpld3d: Latent diffusion of 3d shape generative models by enforcing geometric and physical priors

    Yuan Dong, Qi Zuo, Xiaodong Gu, Weihao Yuan, Zhengyi Zhao, Zilong Dong, Liefeng Bo, and Qixing Huang. Gpld3d: Latent diffusion of 3d shape generative models by enforcing geometric and physical priors. InCVPR, 2024. 3

    work page 2024

  8. [8]

    Get3d: A generative model of high quality 3d textured shapes learned from images

    Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In NeurIPS, 2022. 5

    work page 2022

Show all 63 references
  1. [9]

    Reinforcement learning with deep energy-based policies

    Tuomas Haarnoja, Haoran Tang, Pieter Abbeel, and Sergey Levine. Reinforcement learning with deep energy-based policies. InICLR, 2017. 3, 4

    2017

  2. [10]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. InNeurIPS, 2017. 5

    2017

  3. [11]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. InNeurIPS, 2020. 3

    2020

  4. [12]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. InICLR, 2022. 9, 15

    2022

  5. [13]

    A3d: Does diffusion dream about 3d alignment? InICLR, 2025

    Savva Victorovich Ignatyev, Nina Konovalova, Daniil Selikhanovych, Oleg V oynov, Nikolay Patakin, Ilya Olkov, Dmitry Senushkin, Alexey Artemov, Anton Konushin, Alexander Filippov, Peter Wonka, and Evgeny Burnaev. A3d: Does diffusion dream about 3d alignment? InICLR, 2025. 3

    2025

  6. [14]

    Zero-shot text-guided object generation with dream fields

    Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. InCVPR, 2022. 3

    2022

  7. [15]

    Dreammotion: Space-time self-similar score distillation for zero-shot video editing

    Hyeonho Jeong, Jinho Chang, Geon Yeong Park, and Jong Chul Ye. Dreammotion: Space-time self-similar score distillation for zero-shot video editing. InECCV, 2024. 3

    2024

  8. [16]

    Test-time alignment of diffusion models without reward over-optimization

    Sunwoo Kim, Minkyu Kim, and Dongmin Park. Test-time alignment of diffusion models without reward over-optimization. InICLR, 2025. 3

    2025

  9. [17]

    Laion aesthetic score predictor

    LAION. Laion aesthetic score predictor. https://laion.ai/blog/laion-aesthetics/, 2024. Ac- cessed: 2024-09-27. 5

    2024

  10. [18]

    Diffsplat: Repurposing image diffusion models for scalable 3d gaussian splat generation

    Chenguo Lin, Panwang Pan, Bangbang Yang, Zeming Li, and Yadong Mu. Diffsplat: Repurposing image diffusion models for scalable 3d gaussian splat generation. InICLR, 2025. 5, 7, 8, 15

    2025

  11. [19]

    Flow matching for generative modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. InICLR, 2023. 14

    2023

  12. [20]

    Flow-grpo: Training flow matching models via online rl.arXiv preprint arXiv:2505.05470,

    Jie Liu, Gongye Liu, Jiajun Liang, Yangguang Li, Jiaheng Liu, Xintao Wang, Pengfei Wan, Di Zhang, and Wanli Ouyang. Flow-grpo: Training flow matching models via online rl.arXiv preprint arXiv:2505.05470,

    arXiv

  13. [21]

    Black, Adrian Weller, and Bernhard Schölkopf

    Weiyang Liu, Zeju Qiu, Yao Feng, Yuliang Xiu, Yuxuan Xue, Longhui Yu, Haiwen Feng, Zhen Liu, Juyeon Heo, Songyou Peng, Yandong Wen, Michael J. Black, Adrian Weller, and Bernhard Schölkopf. Parameter-efficient orthogonal finetuning via butterfly factorization. InICLR, 2024. 9

    2024

  14. [22]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. InICLR, 2023. 14

    2023

  15. [23]

    Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu

    Zhen Liu, Yao Feng, Michael J. Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffu- sion: Score-based generative 3d mesh modeling. InICLR, 2023. 2

    2023

  16. [24]

    Black, and Bernhard Schölkopf

    Zhen Liu, Yao Feng, Yuliang Xiu, Weiyang Liu, Liam Paull, Michael J. Black, and Bernhard Schölkopf. Ghost on the shell: An expressive representation of general 3d shapes. InICLR, 2024. 2

    2024

  17. [25]

    Xiao, Weiyang Liu, Yoshua Bengio, and Dinghuai Zhang

    Zhen Liu, Tim Z. Xiao, Weiyang Liu, Yoshua Bengio, and Dinghuai Zhang. Efficient diversity-preserving diffusion alignment via gradient-informed gflownets. InICLR, 2025. 2, 4, 5, 6, 14

    2025

  18. [26]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.NeurIPS, 2022

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.NeurIPS, 2022. 5

    2022

  19. [27]

    Baking symmetry into gflownets

    George Ma, Emmanuel Bengio, Yoshua Bengio, and Dinghuai Zhang. Baking symmetry into gflownets. arXiv preprint arXiv:2406.05426, 2024. 3

    2024 arXiv

  20. [28]

    Inference-time scaling for diffusion models beyond scaling denoising steps.arXiv preprint arXiv:2501.09732, 2025

    Nanye Ma, Shangyuan Tong, Haolin Jia, Hexiang Hu, Yu-Chuan Su, Mingda Zhang, Xuan Yang, Yandong Li, Tommi Jaakkola, Xuhui Jia, et al. Inference-time scaling for diffusion models beyond scaling denoising steps.arXiv preprint arXiv:2501.09732, 2025. 3

    2025 arXiv

  21. [29]

    Gradient imbalance in direct preference optimization.arXiv preprint arXiv:2502.20847, 2025

    Qinwei Ma, Jingzhe Shi, Can Jin, Jenq-Neng Hwang, Serge Belongie, and Lei Li. Gradient imbalance in direct preference optimization.arXiv preprint arXiv:2502.20847, 2025. 3

    2025 arXiv

  22. [30]

    Stochastic generative flow networks.UAI, 2023

    Ling Pan, Dinghuai Zhang, Moksh Jain, Longbo Huang, and Yoshua Bengio. Stochastic generative flow networks.UAI, 2023. 3

    2023

  23. [31]

    Dreamfusion: Text-to-3d using 2d diffusion

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. InICLR, 2023. 3, 4, 5, 8

    2023

  24. [32]

    Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d

    Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. InCVPR, 2024. 5

    2024

  25. [33]

    Controlling text-to-image diffusion by orthogonal finetuning

    Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Schölkopf. Controlling text-to-image diffusion by orthogonal finetuning. InNeurIPS, 2023. 9

    2023

  26. [34]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InICML, 2021. 5

    2021

  27. [35]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. InNeurIPS, 2023. 3

    2023

  28. [36]

    Texture: Text-guided texturing of 3d shapes

    Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. InACM SIGGRAPH 2023 conference proceedings, 2023. 3

    2023

  29. [37]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. InCVPR, 2022. 5

    2022

  30. [38]

    Meshgpt: Generating triangle meshes with decoder-only transformers

    Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. InCVPR, 2024. 2

    2024

  31. [39]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. InICLR, 2021. 5

    2021

  32. [40]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. InICLR, 2021. 14

    2021

  33. [41]

    Lgm: Large multi-view gaussian model for high-resolution 3d content creation.arXiv preprint arXiv:2402.05054, 2024

    Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation.arXiv preprint arXiv:2402.05054, 2024. 15 11

    2024 arXiv

  34. [42]

    Fine-tuning of continuous-time diffusion models as entropy-regularized control.arXiv preprint arXiv:2402.15194, 2024

    Masatoshi Uehara, Yulai Zhao, Kevin Black, Ehsan Hajiramezanali, Gabriele Scalia, Nathaniel Lee Diamant, Alex M Tseng, Tommaso Biancalani, and Sergey Levine. Fine-tuning of continuous-time diffusion models as entropy-regularized control.arXiv preprint arXiv:2402.15194, 2024. 2

    2024 arXiv

  35. [43]

    Mvreward: Better aligning and evaluating multi-view diffusion models with human preferences

    Weitao Wang, Haoran Xu, Yuxiao Yang, Zhifang Liu, Jun Meng, and Haoqian Wang. Mvreward: Better aligning and evaluating multi-view diffusion models with human preferences. InAAAI, 2025. 3

    2025

  36. [44]

    Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation

    Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. InNeurIPS, 2023. 2, 8

    2023

  37. [45]

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis.arXiv preprint arXiv:2306.09341, 2023

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis.arXiv preprint arXiv:2306.09341, 2023. 5, 7

    2023 arXiv

  38. [46]

    Structured 3d latents for scalable and versatile 3d generation.arXiv preprint arXiv:2412.01506, 2024

    Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation.arXiv preprint arXiv:2412.01506, 2024. 2

    2024 arXiv

  39. [47]

    Imagereward: learning and evaluating human preferences for text-to-image generation

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: learning and evaluating human preferences for text-to-image generation. InNeurIPS, 2023. 2, 5

    2023

  40. [48]

    Dancegrpo: Unleashing grpo on visual generation.arXiv preprint arXiv:2505.07818, 2025

    Zeyue Xue, Jie Wu, Yu Gao, Fangyuan Kong, Lingting Zhu, Mengzhao Chen, Zhiheng Liu, Wei Liu, Qiushan Guo, Weilin Huang, et al. Dancegrpo: Unleashing grpo on visual generation.arXiv preprint arXiv:2505.07818, 2025. 14

    2025 arXiv

  41. [49]

    Stablenormal: Reducing diffusion variance for stable and sharp normal.ACM Transactions on Graphics (TOG), 2024

    Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal.ACM Transactions on Graphics (TOG), 2024. 5

    2024

  42. [50]

    Dreamreward: Text-to-3d generation with human preference

    Junliang Ye, Fangfu Liu, Qixiu Li, Zhengyi Wang, Yikai Wang, Xinzhou Wang, Yueqi Duan, and Jun Zhu. Dreamreward: Text-to-3d generation with human preference. InECCV, 2024. 3, 8, 9

    2024

  43. [51]

    Learning to sample effective and diverse prompts for text-to-image generation

    Taeyoung Yun, Dinghuai Zhang, Jinkyoo Park, and Ling Pan. Learning to sample effective and diverse prompts for text-to-image generation. InCVPR, 2025. 3, 9

    2025

  44. [52]

    Rade-gs: Rasterizing depth in gaussian splatting.arXiv preprint arXiv:2406.01467, 2024

    Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting.arXiv preprint arXiv:2406.01467, 2024. 5

    2024 arXiv

  45. [53]

    Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling.arXiv preprint arXiv:2403.19655, 2024

    Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling.arXiv preprint arXiv:2403.19655, 2024. 5, 7, 8, 15

    2024 arXiv

  46. [54]

    Dinghuai Zhang, Ricky T. Q. Chen, Nikolay Malkin, and Yoshua Bengio. Unifying generative models with GFlowNets and beyond.arXiv preprint arXiv:2209.02606v2, 2022. 3

    2022 arXiv

  47. [55]

    Diffusion generative flow samplers: Improving learning signals through partial trajectory optimization

    Dinghuai Zhang, Ricky TQ Chen, Cheng-Hao Liu, Aaron Courville, and Yoshua Bengio. Diffusion generative flow samplers: Improving learning signals through partial trajectory optimization. InICLR,

  48. [56]

    Let the flows tell: Solving graph combinatorial problems with gflownets

    Dinghuai Zhang, Hanjun Dai, Nikolay Malkin, Aaron C Courville, Yoshua Bengio, and Ling Pan. Let the flows tell: Solving graph combinatorial problems with gflownets. InNeurIPS, 2023. 3

    2023

  49. [57]

    Generative flow networks for discrete probabilistic modeling

    Dinghuai Zhang, Nikolay Malkin, Zhen Liu, Alexandra V olokhova, Aaron Courville, and Yoshua Bengio. Generative flow networks for discrete probabilistic modeling. InICML, 2022. 3

    2022

  50. [58]

    Dinghuai Zhang, Ling Pan, Ricky T. Q. Chen, Aaron Courville, and Yoshua Bengio. Distributional GFlownets with quantile flows.Transactions on Machine Learning Research, 2024. 3

    2024

  51. [59]

    Susskind, Navdeep Jaitly, and Shuangfei Zhai

    Dinghuai Zhang, Yizhe Zhang, Jiatao Gu, Ruixiang Zhang, Joshua M. Susskind, Navdeep Jaitly, and Shuangfei Zhai. Improving GFlownets for text-to-image diffusion alignment.Transactions on Machine Learning Research, 2025. 5

    2025

  52. [60]

    Clay: A controllable large-scale generative model for creating high-quality 3d assets.ACM Transactions on Graphics (TOG), 2024

    Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets.ACM Transactions on Graphics (TOG), 2024. 2

    2024

  53. [61]

    Guided flows for generative modeling and decision making.arXiv preprint arXiv:2311.13443, 2023

    Qinqing Zheng, Matt Le, Neta Shaul, Yaron Lipman, Aditya Grover, and Ricky TQ Chen. Guided flows for generative modeling and decision making.arXiv preprint arXiv:2311.13443, 2023. 14

    2023 arXiv

  54. [62]

    Dreamdpo: Aligning text-to- 3d generation with human preferences via direct preference optimization.arXiv preprint arXiv:2502.04370,

    Zhenglin Zhou, Xiaobo Xia, Fan Ma, Hehe Fan, Yi Yang, and Tat-Seng Chua. Dreamdpo: Aligning text-to- 3d generation with human preferences via direct preference optimization.arXiv preprint arXiv:2502.04370,

    arXiv

  55. [2025]

    2) for random variable zt but with the 3D reward specified in Eqn

    3 12 Appendix Table of Contents A Proof of Unbiasedness of Nabla-R2D3 14 B Application to Flow Matching Models 14 C Implementation Details 15 D Comparison with Multi-view Generative Model 15 E Quantitative and Qualitative Results on Depth-Normal Consistency 16 F More Ablation ...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.