acceptodds
Under review as a conference paper at ICLR 2027

FLiTo: Fast High-Fidelity 3D Generation via Image Supervision

Abstract

Recent advances in latent 3D generative modeling have enabled high-quality 3D content generation from images. However, state-of-the-art models typically require many generation steps, resulting in substantial inference cost and limiting their use in interactive applications. While distillation can reduce the number of sampling steps, generation quality degrades significantly in the few-step regime, particularly for single-step generation. In this work, we study how 2D supervision can be effectively incorporated to improve single- and few-step latent 3D generation. We derive a variational objective that naturally unifies conditioning-image reconstruction with distribution matching against a teacher 3D generator. Building on this formulation, we incorporate complementary forms of 2D supervision, including feature-distribution matching, semantic consistency, and human preference. Through extensive ablations, we analyze the contribution of each supervision signal and show that 2D supervision is particularly effective in the challenging single-step regime. Our final single-step generator achieves state-of-the-art speed and generation quality among single-step 3D generators, reducing FID by 47% over the strongest baseline while running in 67.5 ms on an NVIDIA B200 GPU and 2.4 seconds on an MacBook Pro with M5 Max.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_VnxKOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FLiTo improves one- and few-step LiTo generation using image reconstruction, teacher distribution matching, and additional image-space losses.

Critical strengths
  1. The single-step results show a substantial improvement in both input fidelity and novel-view quality.
  2. Comparisons with MeanFlow and DMD on the same backbone support the value of image supervision.
  3. The paper also provides useful loss ablations and detailed runtime breakdowns.
Critical weaknesses
  1. The method is evaluated only on LiTo. Testing MeanFlow and DMD initializations does not establish whether the gains transfer to other 3D backbones or representations.
  2. The benefit of human-preference supervision is unclear: removing it improves every reported metric in Table 6, including HPS.
Questions for authors
  1. Why retain the HPS loss when removing it improves all metrics in Table 6? Does it help in longer training or human evaluation?
  2. Can you directly evaluate geometry for the full and no-KL variants? Teacher likelihood alone does not establish geometric quality.
  3. Which KL-gradient sign is used in the implementation, Eq. (3) or Eq. (18)?
4 comments
AuthorsOpenReview ↗

Author response — Part 1

Thank you for the careful review and the comments --- we do our best to answer your questions below. Please let us know if you have any remaining thoughts or concerns.

Scaling beyond LiTo. We agree that demonstrating our method on approaches beyond LiTo would make our work stronger. However, for our current framework to be most effective, it requires certain properties for the initial base model that most existing baselines do not satisfy. For one, it requires that the generated objects, when rendered from a consistent camera pose, produce an image aligned with the input image. This is not only required for the conditioning reconstruction loss, but also maximizes the performance for the other rewards, e.g., it allows for aligned ground truth and rendered novel views for feature-distribution matching and paired human-preference supervision. Unfortunately, many approaches output generated 3D assets in a canonical world frame rather than in the conditioning camera’s coordinate system, e.g., most TRELLIS-type approaches. Some approaches also output the pose of the conditioning camera (e.g., SAM3D), but we found their performance to be quite poor in the few-step setting --- in fact, in our evaluations, we had to run slow post-generation pose optimization for meaningful results (sec. C of the appendix). LiTo performed the best in our experiments, as it explicitly predicts the generated object in a consistent coordinate system with the conditioning view on the x-axis.

Secondly, a latent space for the generative model that is capable of replicating the real image distribution is desirable for our framework. In our experiments, we found most popular existing baselines suffered from various appearance artifacts. Public checkpoints of TRELLIS and SAM3D suffer from significant color shift, which can be attributed to the lossiness of their DINO-based framework (Singh et al., 2026). Mesh-based approaches, e.g., TRELLIS-2, output PBR materials, which not only require accurate lighting but also generally struggle to replicate real-world appearance. On the other hand, LiTo directly parameterizes the surface light field, which allows the base model to already well-replicate realistic appearance.

While overcoming these view-alignment and latent-space challenges is not impossible with sufficient engineering, given the limited paper (and rebuttal) timeline, we decided to primarily build upon LiTo for our experiments. We have updated Sec. 4.1 to describe these desired properties. We leave extending the ideas of our paper to other base models as interesting future work.

Why HPS? Even though removing HPS improved our metrics in our small-scale subtractive ablation, we opted to keep it for our full-size run for a few reasons. First, we observed that it was helpful in our additive ablation --- as shown in Table 2, it improves preference metrics and slightly improves novel-view metrics over a base model. Second, qualitatively when HPS was removed, we observed a slight “flattening” of generated assets that we found to be an indicator of later collapse in our earlier experiments (and does not typically present in the metrics). To better quantify this, we computed the ratio of the extent along the x-axis (i.e., depth direction) of generated objects relative to the mean of their y- and z-extents. On the test set, the average ground-truth ratio is 1.064. We found an average ratio of 1.078 for our full recipe vs 0.888 for the run without HPS, indicating consistent flattening.

Finally, from a more philosophical point-of-view, HPS fills an important gap relative to the other 2D rewards we propose. The conditioning reconstruction loss provides strong supervision on the front-side of an object. For novel views, feature-distribution matching provides coarse, distribution level supervision, while the semantic consistency loss provides a rough conceptual training signal. From that perspective, HPS provides complementary, low-level supervision for the visual quality of a novel view. We believe a human-preference model that is easier to backpropagate through than the current HPSv3 model would likely significantly improve the quality of our results.

To clarify these points, we have added a new Figure 12 illustrating the slight flattening without HPS to the appendix, and will refine the discussion in the paper to better highlight our conceptual reasoning for HPS. Unfortunately, we do not have the resources to complete a full training run without HPS within the time limits of the rebuttal period.

AuthorsOpenReview ↗

Author response — Part 2

KL and geometry. As suggested, to measure the quality of 3D geometry, we have calculated Uni3D metrics for a few informative runs (please see below), and will add complete numbers for all runs in Table 2 and 6. We also added new Figure 10 and 11 to visualize the effects of the KL regularization.

To start, in Table 2, when we finetune a base MeanFlow model with the conditioning reconstruction loss (+cr) and additionally with the KL regularization (+cr + kl), we observe that the model trained without KL produces worse novel view metrics and worse Uni3D score (0.343 vs 0.353). Without KL, the geometry deteriorates further with additional training --- with just twice the iterations, objects are often generated without a backside, as visualized in the new Figure 10(a); this is reflected in a Uni3D score of 0.302 without KL, versus 0.354 with. We hypothesize that when only the conditioning reconstruction loss is used, the backsides are sacrificed to maximize frontside performance due to the limited model capacity of few-step models. The model with KL regularization does not suffer from such artifacts.

Under our full recipe that has additional objective terms, the geometry-preserving aspect of the KL term is less prominent --- the model trained without KL has a Uni3D score of 0.367 and that with KL has 0.364, well within the standard deviation of 0.06. Nonetheless, other benefits of the KL term become more prominent.

First, we observe that the KL regularization helps preserve view-dependent appearances in the generated objects. Specifically, when we remove the KL regularization from our full recipe (i.e., -kl in Table 6), the model tends to bake view-dependent highlights into the diffuse appearance, as shown in the new Figure 10(b). To the best of our efforts, we could not find a metric that measures view-dependent appearance, and we used NLL as a proxy. To help quantify this effect, we added a new Figure 11(a), which shows the average energy of the spherical harmonics coefficients of different degrees of the Gaussians in the generated objects across the test set. As we can see, removing KL from the recipe results in a higher 0th-degree energy and lower higher-degree energy, indicating the loss of view-dependent effects. We appreciate any suggestions on a metric that might better capture this loss of view-dependence.

Second, we observe that without the KL term, the proposed 2D supervision can push generated latents out of the original 3D token distribution. This can degrade the performance of downstream decoders that are not utilized during our training. As shown in the new Figure 10(c) and 11(b), an illustrative example is the point cloud, flow-matching-based decoder of the LiTo representation. LiTo uses this decoder to supervise 3D geometry during tokenizer training --- however, to produce Gaussian splats during inference, they use an independently-trained occupancy decoder for efficiency. We similarly use this occupancy decoder for our decoded Gaussians in our training pipeline, and we do not use the point cloud decoder. Ideally, an occupancy grid inferred from the point cloud decoder should roughly match that of an occupancy decoder --- we find an average Intersection over Union (IoU) of 0.778 on the ground-truth tokenizer output from our test set. However, after training without the KL regularization, the two decoders begin to diverge from each other --- we observe an average IoU of 0.572 without the KL term, versus 0.783 with the KL term which is closer to the reference from the ground-truth tokenizer (Figure 11(b)). This visually manifests as a visible bias in the point cloud decoder output, which is visualized in a new Figure 10(c).

We have expanded the discussion of the subtractive ablation to include the above, and we will add the Uni3D, SH-degree and decoder IoU metrics for all runs in Table 2 and 6. We believe these observations support our derivation in Section 3.1 and the use of KL regularization.

AuthorsOpenReview ↗

Author response — Part 3

KL-gradient sign inconsistency. We apologize for the typo. The correct expression should be the one in the appendix, i.e., the negative of the original Equation (3), which is now fixed in the current manuscript. We use the correct expression in our implementation, as reflected by the findings above.

Other improvements to the paper. We also made a few other improvements to the paper compared to the original version. First, we have updated Table 1(b) with larger-scale checkpoints of our 4-step models, which results in further improved metrics. Second, in Table 3(b), we accidentally used an incorrect result for “ours (meanflow init)” --- the correct numbers correspond to better input view metrics and novel view FID. Third, we added a new Figure 3, which compares the results of our 1-step model with the 1-step MeanFlow base model and highlights the benefits of our recipe. Additionally, we have slightly expanded the related work, primarily in terms of past methods of 2D supervision for 3D generation. Finally, we have fixed typos and improved writing throughout the paper.

Thanks for the response, I will keep my rating.

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

FLiTo distills a pretrained image-conditioned 3D generator into a model requiring very few sampling steps. The formulation combines image-space supervision with a KL-based constraint to the 3D teacher, instantiated through distribution matching. Additional feature-distribution, semantic, and preference objectives improve the rendered outputs. The method is evaluated against several 3D generation and distillation baselines, with ablations of the training losses. The one-step model substantially improves rendered-image metrics over the reported one-step alternatives while retaining a large latent-generation speed advantage over the teacher.

Critical strengths
  • The paper tackles a useful bottleneck: obtaining strong 3D outputs from a one-step generator without relying only on matching the teacher's sampling trajectory. The results in Table 1 are substantial, particularly the novel-view FID improvement over the reported one-step alternatives.

  • The ablations give meaningful evidence for image supervision as the main source of improvement. Table 2 examines combinations of reconstruction, KL, feature-distribution, semantic, and preference terms, while Table 3 tests the approach with several distillation formulations and sampling budgets. This supports a useful training strategy rather than a result tied to only one sampler.

Critical weaknesses

The role of the KL term in protecting 3D quality is not yet independently established. This term is central to the interpretation that image supervision can improve appearance while retaining the teacher's 3D prior. However, the subtractive ablation in Appendix E, Table 6, shows that removing KL improves FID and LPIPS while worsening teacher-prior NLL. The paper acknowledges the associated appearance trade-off, and the ablation is valuable. The remaining issue is that teacher NLL measures agreement with the very prior being imposed; it does not independently establish better geometry or multi-view consistency. An otherwise matched comparison with and without KL using an independent 3D or cross-view diagnostic would determine whether this regularizer preserves a consequential property that the rendered-image metrics miss. Without it, the practical results remain strong, but the explanation of why the 3D regularization is needed is less convincing.

Questions for authors
  • What independently measurable 3D property deteriorates when KL is removed in Table 6? For example, can the matched models be compared for cross-view consistency, geometric accuracy where reference geometry is available, or a clearly defined view-dependent artifact measure?
  • The score difference in main-text Eq. (3) has the opposite sign from Appendix A, Eq. (18). Can you confirm the implemented minimization direction and reconcile these expressions?
2 comments
AuthorsOpenReview ↗

Author response — Part 1

Thank you for the careful review and the comments --- we do our best to answer your questions below. Please let us know if you have any remaining thoughts or concerns.

Role of the KL divergence term. To better illustrate the role of the KL divergence, we have added new Figures 10 and 11 where we compare models trained with and without the KL divergence both qualitatively and quantitatively. Specifically, there are three main artifacts that we have observed in models trained without KL divergence.

To start, in Table 2, when we finetune a base MeanFlow model with the conditioning reconstruction loss (+cr) and additionally with the KL regularization (+cr + kl), we observe that the model trained without KL produces worse novel view metrics and worse Uni3D score (0.343 vs 0.353). Without KL, the geometry deteriorates further with additional training --- with just twice the iterations, objects are often generated without a backside, as visualized in the new Figure 10(a); this is reflected in a Uni3D score of 0.302 without KL, versus 0.354 with. We hypothesize that when only the conditioning reconstruction loss is used, the backsides are sacrificed to maximize frontside performance due to the limited model capacity of few-step models. The model with KL regularization does not suffer from such artifacts.

Second, we observe significant loss in view-dependent appearance when the KL term is removed from our full recipe in the subtractive ablation. As can be seen in the new Figure 10(b), the model trained without the KL term tends to bake view-dependent highlights into the diffuse appearance. Notably, this view-dependent effect is difficult to measure, and to the best of our efforts, we could not find a 2D or 3D metric that was able to independently quantify it. Thus, we opted for teacher NLL, which is an imperfect proxy as you have already noted. To provide better evidence for this finding, we have added two new figures to the manuscript. First, we have added qualitative visualizations in the appendix (Figure 10 (b)) and the supplemental website (“KL divergence ablation”). Second, we have also added a plot of the strength of each Gaussian spherical harmonic averaged over the dataset in the new Figure 11(a) --- removing the KL term reduces the strength of higher order spherical harmonics, indicating the loss of view-dependent effects. We appreciate any suggestions on a metric that might be more sensitive to this loss of view-dependence.

Third, we observe that without the KL term, the proposed 2D supervision can push generated latents out of the original tokenizer distribution. This can degrade the performance of downstream decoders that are not utilized during our training. We have illustrated the effect in the new Figure 10(c) and 11(b). An illustrative example is the point cloud, flow-matching-based decoder of the LiTo representation. LiTo uses this decoder to supervise 3D geometry during the tokenizer training --- however, to produce Gaussian splats during inference, they actually use an independently-trained occupancy decoder for efficiency. We similarly use this occupancy decoder for our decoded Gaussians in our training pipeline, and we do not use the point cloud decoder. Ideally, an occupancy grid inferred from the point cloud decoder should roughly match that of an occupancy decoder --- we find an average IoU of 0.778 on the direct ground-truth tokenizer output from our test set. However, we observe that after training, the two decoders begin to diverge from each other if the KL term is not utilized --- we observe an average IoU of 0.572 without the KL term, versus 0.783 with the KL term (Figure 11(b)). This visually manifests as a visible bias in the point cloud decoder output, which is visualized in a new Figure 10(c).

We have expanded the discussion of the subtractive ablation to include the above, and we will add the Uni3D, SH-degree and decoder IoU metrics for all runs in Table 2 and 6. We believe these observations support our derivation in Section 3.1 and the use of KL regularization.

AuthorsOpenReview ↗

Author response – Part 2

KL-gradient sign inconsistency. We apologize for the typo. The correct expression should be the one in the appendix, i.e., the negative of the original Equation (3), which is now fixed in the current manuscript. We use the correct expression in our implementation, as reflected by the findings above.

Other improvements to the paper. We also made a few other improvements to the paper compared to the original version. First, we have updated Table 1(b) with larger-scale checkpoints of our 4-step models, which results in further improved metrics. Second, in Table 3(b), we accidentally used an incorrect result for “ours (meanflow init)” --- the correct numbers correspond to better input view metrics and novel view FID. Third, we added a new Figure 3, which compares the results of our 1-step model with the 1-step MeanFlow base model and highlights the benefits of our recipe. Additionally, we have slightly expanded the related work, primarily in terms of past methods of 2D supervision for 3D generation. Finally, we have fixed typos and improved writing throughout the paper.

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

FLiTo addresses the inference cost of image-to-3D generation through a variational framework combining input-view reconstruction with distribution matching to a 3D teacher. Training also incorporates feature-distribution matching, semantic consistency, and preference-model supervision. Evaluated on 1,000 Objaverse-XL test objects, the single-step model generates 3D latents in 67.5 ms on a B200 GPU, excluding decoding, and achieves a novel-view FID of 13.14, a 47% reduction relative to single-step MDT-dist. The four-step variant reaches an FID of 10.77, close to the 25-step teacher’s 10.06. Ablations support the complementary roles of reconstruction and KL regularization, as well as the benefits of feature-distribution matching and semantic supervision, while preference supervision shows mixed results. Evaluation relies mainly on rendered-image metrics rather than direct measurements of 3D geometry.

Critical strengths
  1. Clear speed and quality gains in single-step 3D generation. Single-step FLiTo achieves a novel-view FID of 13.14, approximately 47% lower than MDT-dist’s 24.92, while reducing latent generation time from 150.3 ms to 67.5 ms. These results demonstrate an improved speed–quality trade-off under the tested settings and provide the strongest empirical support for acceptance.

  2. A well-motivated supervision framework supported by targeted ablations. The method combines input-view reconstruction with teacher distribution regularization to balance input fidelity and 3D plausibility. For novel views, distribution-level and semantic supervision avoid imposing pixel-wise constraints on inherently ambiguous unseen regions. Ablations support the complementary roles of these signals, with improvements across different initialization schemes and in four-step generation, suggesting applicability across distillation settings.

Critical weaknesses

A sign inconsistency in the KL-gradient formula makes the method difficult to verify. Equation (3) in the main text uses the teacher score minus the student score, whereas Equation (18) in the appendix gives the opposite order. Since this term is central to preventing degeneration of the generated 3D representation, the actual optimization direction needs to be clarified to establish consistency between the derivation and implementation. Although this may simply be a typographical error, resolving it is important for verifying correctness and ensuring reproducibility.

Questions for authors

Which KL-gradient direction is used in the implementation? Equations (3) and (18) use opposite signs for the score difference. Please clarify which expression is correct and whether the discrepancy is merely a typographical error.

1 comment

Thank you for the careful review and the comments --- we do our best to answer your questions below. Please let us know if you have any remaining thoughts or concerns.

KL-gradient sign inconsistency. We apologize for the typo. The correct expression should be the one in the appendix, i.e., the negative of Equation (3), which is now fixed in the current manuscript. We use the correct expression in our implementation --- this is reflected by the worsened teacher NLL without KL in Table 6.

As further evidence for the correctness of our implementation, the new Figure 10 and 11 highlight the issues we observed when the KL term is removed from training. First, training with only a conditioning-reconstruction loss without KL results in a model that outputs low-quality object backsides, as supervision focuses on improving the conditioning view alone (Figure 10(a)). For a model trained with our full recipe except the KL term, it tends to bake view-dependent highlights into diffuse appearance, as shown in Figure 10(b) and 11(a). We also observe degraded performance in downstream decoders not used during training, indicating a distribution shift in generated latents as visualized in Figure 10(c) and 11(b). These problems do not arise when training with the KL term.

Other improvements to the paper. We also made a few other improvements to the paper compared to the original version. First, we have updated Table 1(b) with larger-scale checkpoints of our 4-step models, which results in further improved metrics. Second, in Table 3(b), we accidentally used an incorrect result for “ours (meanflow init)” --- the correct numbers correspond to better input view metrics and novel view FID. Third, we added a new Figure 3, which compares the results of our 1-step model with the 1-step MeanFlow base model and highlights the benefits of our recipe. Additionally, we have slightly expanded the related work, primarily in terms of past methods of 2D supervision for 3D generation. Finally, we have fixed typos and improved writing throughout the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.