acceptodds
Under review as a conference paper at ICLR 2027

FlowMorpher: Inversion-Free Textured 3D Morphing via Velocity Field Interpolation

Abstract

Textured 3D morphing aims to achieve smooth changes in geometry and appearance while recovering both endpoints. However, existing training-free approaches require access to model internals for feature matching and interpolation, and often exhibit abrupt transitions. Other methods require costly caching of intermediate latent states and attention features obtained through 3D inversion. To address these gaps, we introduce FlowMorpher, a training-free and inversion-free method that constructs sampling dynamics by interpolating conditional velocity fields of pretrained image-to-3D flow models without modifying their internals. FlowMorpher constructs sampling trajectories directly from shared Gaussian noise, eliminating the need for inversion. In addition, we define the sampling update rule by interpolating the model's conditional velocity predictions, without modifying internal attention operations. Under the stated regularity assumptions, our formulation guarantees Lipschitz continuity of the terminal latent with respect to the morphing weight within a fixed latent space. This analysis provides theoretical support for smooth transitions by establishing latent continuity under the stated assumptions. To evaluate FlowMorpher, we assemble MorphEdit3D from existing datasets and benchmarks, covering color, shape, and semantic changes. Our quantitative results improve over the strongest baseline on every reported metric, reducing KID by 11.2% and PDV by 21.4% while raising target-endpoint DINO-I by 0.039. Qualitative comparisons show intermediate assets with smooth transitions in the generated 3D sequences. These findings show that textured 3D morphing can be controlled through sampling dynamics defined by pretrained velocity predictions.

Primary area:
generative models
Keywords:
Image-to-3D Generation, Textured 3D Morphing

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_MAM7OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a new training-free morphing pipeline via a novel velocity field interpolation (VFI) for textured 3D morphing. They use a pretrained Trellis text-image-to-3D flow model without retraining any weights. Via direct latent manipulation using their novel VFI, they are able to perform morphing naturally without training. With this they are able to perform multiple types of morphing, such as color, shape, and basic semantics. They are able to surpass previous baselines and achieve a SoTA morphing benchmark. Additionally, they developed a novel benchmark for morphing, called MorphEdit3D.

Critical strengths
  • A newer way of evaluating morphing that is essential for the community, as the MorphEdit3D benchmark, as morphing evaluation is difficult and counterintuitive.
  • State-of-the-art results on their own evaluation dataset
  • Marginal numerical results on MorphEdit3D
  • Impressive 3D morphing results when compared to prior training-free schemas.
  • Technically sound and well-written paper with a simple pipeline for latent morphing.
Critical weaknesses

The biggest weaknesses of this pipeline and the historical approach to training free 3D morphing is that the morphing task is most commonly used to morph between objects that are distant from each other, for example, a human becoming a wolf, a car becoming a humanoid car, a human becoming a fish. So on, obviously, sometimes a human-to-human morph is also requested, but is less common. What I am saying here is that your pipeline does not behave well with highly different subjects (semantically), and that is a downside that has to be made clear in the main text, not only in the supplementary material.

  • Marginal numerical results on MorphEdit3D

  • Under 3D morphing-related works, some newer models can perform morphing with implicit neural networks [2] and also with flow models [1], which some are not "training-free" and perform well with semantically different inputs, but [2] has downsides of needing matching landmarks and [1] is not open-source; could you please compare their approach with yours, either numerically or in the related works.

  • The MorphEdit3D seems like a good dataset for evaluating 3D morphing. However, I would like to see some examples of it (input and output). Also, the most important part of morphing is the 0.5 time of interpolation. How do you evaluate that? I know that the evaluation is done at 1/3 and 2/3 of alpha, but 1/2 is the most challenging part. I'd also like to see an evaluation table at the 1/2 time point.

  • What is the comparison of computer time to generate a morphing FlowMorpher vs. baselines? An "average time" column on Table 1 would suffice for my question. A per-scene breakdown of time to generate the morph is also good to answer my question but would have less impact.

I am open to considering raising my scores if the authors can address my concerns.

[1] Wukong's 72 Transformations: High-fidelity Textured 3D Morphing via Flow Models, NeurIPS 2025 [2] FLOWING 🌊: Implicit Neural Flows for Structure-Preserving Morphing, NeurIPS 2025

Questions for authors

I do not have any questions.

4 comments
AuthorsOpenReview ↗

Response to Reviewer MAM7 (Part 1/2)

We sincerely thank you for your appreciation and recognition of our work. Below, we provide detailed responses to your concerns.

W1: The biggest weaknesses of this pipeline and the historical approach to training free 3D morphing is that the morphing task is most commonly used to morph between objects that are distant from each other, for example, a human becoming a wolf, a car becoming a humanoid car, a human becoming a fish. So on, obviously, sometimes a human-to-human morph is also requested, but is less common. What I am saying here is that your pipeline does not behave well with highly different subjects (semantically), and that is a downside that has to be made clear in the main text, not only in the supplementary material.

Thank you for pointing this out. We have revised the Introduction and Method to state explicitly which cases FlowMorpher handles and where it fails.

Specifically, FlowMorpher handles moderate changes, up to the semantic level of MorphEdit3D. For more distant categories (e.g., a human and a fish), local attributes such as color still change gradually, but the overall geometry jumps abruptly at an intermediate weight. We attribute this to the pretrained prior, which offers no plausible intermediate shapes between such categories.

W2: Marginal numerical results on MorphEdit3D

Thanks for your comment. From the Table 1 in manuscript, the margins over the strongest baselines are modest for FID and LPIPS_adj, and that the larger differences are in PDV and target-endpoint DINO-I. However, our qualitative comparisons show a clear gap in fidelity, with baselines exhibiting artifacts such as collage-like compositions. Therefore, to further compare FlowMorpher with the baselines, we ran a user study with 30 participants on 45 pairs randomly selected from MorphEdit3D.

Experimental setup: For each comparison, participants saw the results of FlowMorpher and one baseline side by side, anonymized and in random order, and rated their preference on a 7-point scale. Ratings were coded so that 1 means a strong preference for the baseline, 4 no preference, and 7 a strong preference for FlowMorpher.

Results: As shown in Table C1, the mean rating is above the neutral value of 4 for every baseline and criterion (smoothness, plausibility, endpoint fidelity, and overall). This user study provides complementary evidence that participants prefer FlowMorpher over the baselines.

Table C1: User study comparing FlowMorpher with baselines on MorphEdit3D.

FlowMorpher vs. BaselineSmoothnessPlausibilityEndpoint fidelityOverall
DiffMorpher6.276.136.306.33
VoxHammer6.175.605.136.17
FreeMorph6.335.936.536.13
Interp3D5.875.004.875.97
Nano3D6.305.175.735.87

W3: Under 3D morphing-related works, some newer models can perform morphing with implicit neural networks [2] and also with flow models [1], which some are not "training-free" and perform well with semantically different inputs, but [2] has downsides of needing matching landmarks and [1] is not open-source; could you please compare their approach with yours, either numerically or in the related works.

Thank you for pointing these works out. They are closely related and deserve a more explicit discussion. Like FLOWING [C1], FlowMorpher builds on a flow-based generative model, and Wukong [C2] is the closest prior work to ours, as it also performs training-free 3D morphing.

We have clarified the differences and positioned our method relative to them in detail in Related Work (Sec. 2). Specifically, the key differences are as follows:

  • FLOWING trains a flow-based implicit network for each pair to align source and target from landmark correspondences, and then blends them. FLOWING needs correspondences, which are undefined for pairs without shared structure. FlowMorpher needs neither correspondences nor per-pair training, and generates the intermediates from the 3D prior.
  • Wukong interpolates in condition space, avoids inversion and formulates morphing as an optimal transport barycenter problem with sequential initialization across frames. While FlowMorpher interpolates in velocity space and the generator is queried with the two real conditions, and the blending is applied to its outputs at a shared state. This requires no auxiliary optimisation, sequential initialisation strategy, or threshold, and the same rule applies to both stages.
AuthorsOpenReview ↗

Response to Reviewer MAM7 (Part 2/2)

W4: The MorphEdit3D seems like a good dataset for evaluating 3D morphing. However, I would like to see some examples of it (input and output). Also, the most important part of morphing is the 0.5 time of interpolation. How do you evaluate that? I know that the evaluation is done at 1/3 and 2/3 of alpha, but 1/2 is the most challenging part. I'd also like to see an evaluation table at the 1/2 time point.

Dataset examples: We have added the source and target input images to Figures 6–12, so that each qualitative result now shows the inputs next to the generated sequence.

Evaluation at α = 1/2: The midpoint was already one of the seven frames generated for each pair, and the smoothness metrics in Table 1 (LPIPS_adj and PDV) are computed over all six adjacent transitions, including the two around α = 1/2. Only FID and KID were evaluated at α = 1/3 and 2/3. We now also report them at α = 1/2 with the same protocol: the midpoint asset is rendered from 1,000 viewpoints and compared with the renders of the two endpoints.

Results: FlowMorpher remains the best method at the midpoint, and its margin is wider there. Its FID is 17% lower and its KID 33% lower than the strongest baseline at this point (Interp3D), compared with 3.5% and 11.2% over the strongest baseline at α = 1/3 and 2/3 (Nano3D, Table 1). This agrees with the point that the midpoint is the most demanding part of a morph. We have added these results to Table 1 in the revised manuscript.

Table C2: Quantitative comparison on MorphEdit3D at α = 1/2. KID values are multiplied by 10310^3.

MethodsFID@½ ↓KID@½↓
FreeMorph142.5991.88
VoxHammer83.3140.65
DiffMorpher81.2837.55
Nano3D61.1422.95
Interp3D58.3819.69
FlowMorpher (ours)48.4613.21

W5: What is the comparison of computer time to generate a morphing FlowMorpher vs. baselines? An "average time" column on Table 1 would suffice for my question. A per-scene breakdown of time to generate the morph is also good to answer my question but would have less impact.

Experimental setup: We measured all methods on a single RTX 4090 for a 7-frame sequence on MorphEdit3D. Setup covers every per-pair cost before the first frame (2D morphing, inversion, feature caching, LoRA tuning), and peak memory is the maximum allocated GPU memory.

Results: The efficiency results are reported in Table C3. FlowMorpher has no setup stage and the lowest total time. This is 20% less than Nano3D, 2.4× faster than Interp3D and 4.6× faster than VoxHammer, whose inversion alone takes 95 s.

Table C3: Efficiency comparison for generating a 7-frame sequence.

MethodInversion (MiB)Feature cache (MiB)Per-pair tuning3D maskSetup (s)Per-frame (s)Peak mem (GB)Total (s)
Image morphing + 3D generation
FreeMorph0.5—✗✗46.45.816.687.0
DiffMorpher262625✓✗59.15.715.999.2
3D editing adapted for morphing
VoxHammer†47,38361✗✓95.129.515.8301.4
Nano3D†——✗✓10.910.214.382.5
Native 3D morphing
Interp3D—2303✗✗17.120.716.9161.9
FlowMorpher (ours)——✗✗—9.4415.766.1

[C1] Bizzi A, Grynberg Portnoy M, Pereira Matias V, et al. Flowing: Implicit neural flows for structure-preserving morphing. NeurIPS, 2025.

[C2] Minghao Yin, Yukang Cao, and Kai Han. Wukong’s 72 Transformations: High-fidelity Textured 3D Morphing via Flow Models. NeurIPS, 2025.

You addressed all my concerns: the evaluation at α = 0.5, the computational efficiency comparison, the user study, and the clearer discussion of limitations regarding large semantic transformations.

The new experiments strengthen the empirical evidence supporting the proposed method, and the revised manuscript provides a more complete evaluation of its strengths and limitations. Although some limitations remain, I believe the authors have adequately addressed my major concerns, and the paper itself provides a technical contribution. I will wait for the other reviewers' discussions, but I am leaning towards improving my score.

Thank you very much for your thoughtful follow-up and for recognizing the technical contribution of our work. We are glad that the evaluation at α = 0.5, the efficiency comparison, the user study, and the clearer discussion of limitations have addressed your concerns.

Your suggestions have substantially improved the paper. We sincerely appreciate your positive assessment of our work.

Reviewer_otXcOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes FlowMorpher, a training-free method for textured 3D morphing built on pretrained rectified-flow 3D generators (mainly TRELLIS). For each morph weight, sampling starts from the same Gaussian noise used to generate both endpoints. At each step, the source- and target-conditioned velocities are evaluated at the same state and combined linearly. The same rule is applied to both TRELLIS stages. In the SLAT stage, noise is drawn from a coordinate-indexed field so that voxels shared across supports receive identical initial noise. They also introduce MorphEdit3D, 135 pairs covering color, shape, and semantic changes, and report improvements over multiple baselines.

Critical strengths

The method is simple, training-free, and leaves model internals unchanged. It can transfer across flow-based 3D generators.

The paper is carefully written. The scope of the theory is stated precisely, including what it does not guarantee. I appreciate this honesty.

Qualitative results are clean. The comparisons with the editing-based adaptation clearly illustrate the feature-collage failure mode.

Critical weaknesses

Here is my biggest concern. The core operation is a convex combination of two conditional velocities from shared noise. This is closely related to existing condition- and score-composition techniques, such as Composable Diffusion, CFG-style condition mixing, and prompt/condition interpolation in 2D morphing (e.g., MagicMix, AID, IMPUS). None of these are discussed.

The ablation suggests the SLAT-stage contribution is marginal.

Questions for authors

How much does the coordinate-indexed SLAT noise matter?

The natural controls for the claim “interpolating velocities is better” are simple alternatives under the same shared-noise setting. However, I didn't see such settings in the ablation experiment. Could the authors provide one?

3 comments
AuthorsOpenReview ↗

Response to Reviewer otXc (Part 1/2)

Thanks for your valuable review and constructive feedback. Below, we provide detailed responses to the concerns.

W1: Here is my biggest concern. The core operation is a convex combination of two conditional velocities from shared noise. This is closely related to existing condition- and score-composition techniques, such as Composable Diffusion, CFG-style condition mixing, and prompt/condition interpolation in 2D morphing (e.g., MagicMix, AID, IMPUS). None of these are discussed.

Thank you for pointing these works out. They are closely related and deserve a more explicit discussion. FlowMorpher shares its basic operator with existing condition- and score-composition techniques, including Composable Diffusion [B1] and CFG-style condition mixing, and our interpolation is closely related to prompt/condition interpolation methods for 2D morphing (MagicMix [B2], AID [B3], IMPUS [B4]).

We have clarified the differences and positioned our method relative to them in detail in Related Work (Sec. 2), and added a new paragraph discussing the convex combination operator in Method (Sec. 5.1). Specifically, the key differences are as follows:

  • Composable Diffusion and CFG-style condition mixing: CFG and its multi-condition extensions, such as Composable Diffusion, linearly combine conditional and unconditional predictions with fixed guidance weights to steer a single sample, which in Composable Diffusion means satisfying several concepts at once (a conjunction). FlowMorpher uses the same form for a different purpose: the convex weights (1−α, α) act as a continuous coordinate along a path between two objects, and the resulting field is integrated from a single shared noise sample. Our interpolation is also orthogonal to guidance: under a shared guidance strength, interpolating the guided velocities is equivalent to interpolating the conditional velocities and then applying CFG, so guidance sharpens each sample while α selects the position along the morph.
  • Mixing and interpolation methods:
    • MagicMix mixes at the level of the noisy latent state and splits the two concepts along the sampling trajectory: one concept supplies a noised layout as the starting state, and the other enters only through the text condition during the remaining denoising steps. The two inputs are thus treated asymmetrically.
    • AID interpolates inside the attention layers and selects the interpolation coefficients with a Beta prior. Its starting observation, that linear interpolation in the conditioning space gives inconsistent results, is consistent with our findings.
    • IMPUS interpolates in the text-embedding space and in the inverted Gaussian latent space, and requires per-pair optimization (endpoint text embeddings and low-rank adaptation) together with ODE inversion.

In summary, the differences lie in where the interpolation takes place and what it is for. FlowMorpher interpolates the output velocity, evaluated at the two original conditions, from a single shared noise sample. It treats the two inputs symmetrically and requires no inversion or per-pair optimization. Our contribution is to show that, for 3D morphing with rectified flow generators, velocity space is a more effective place to interpolate than condition space, and that shared noise matters (Table B1), and to analyze why this yields smooth, endpoint-faithful sequences (Sec. 5.1).

W2: The ablation suggests the SLAT-stage contribution is marginal.

Thank you for raising this. We have expanded the analysis in Sec. 6 to explain the SLAT-stage contribution.

The two stages control different factors. In TRELLIS [B5], the SS stage decides which voxels are occupied, i.e., the coarse structure, and the SLAT stage generates local latents on that fixed support, which encode fine geometry and appearance. Our metrics are computed on renders, where coarse structure accounts for most of the change between frames.

The contribution of each stage depends on the pair type. Table 3 is consistent with this difference: the gap between SS-only and SLAT-only is largest on shape-level pairs (FID 34.38 vs. 35.85, PDV 1.06 vs. 1.44), where the two objects differ in structure, and nearly vanishes on color-level pairs (FID 33.39 vs. 33.48, identical LPIPS_adj), where the structure is largely shared. On semantic-level pairs, SLAT-only also obtains better FID and KID than SS-only (38.41 vs. 39.21 and 9.04 vs. 10.22).

In summary, the SLAT-stage contribution depends on the pair type: the SS stage accounts for most of the gain when the two objects differ in structure, whereas SLAT-only matches SS-only on color-level pairs and achieves better per-frame quality (FID, KID) on semantic-level pairs. Therefore, FlowMorpher apply the same interpolation rule to both, achieving the best performance on MorphEdit3D across all three morphing levels.

AuthorsOpenReview ↗

Response to Reviewer otXc (Part 2/2)

Q1: How much does the coordinate-indexed SLAT noise matter?

Experimental setup: To measure the effect of the coordinate-indexed SLAT noise, we compared three ways of drawing the SLAT noise, keeping everything else fixed: (1) independent noise for each α; (2) the default TRELLIS sampler with the seed reset for each α; and (3) our coordinate-indexed field.

Results: Relative to the default sampler, coordinate indexing lowers FID by 9.6%, LPIPS_adj by 12%, and PDV by 20%. This is because the default sampler assigns noise by row order in the active-voxel list. When the support changes with α, the same voxel receives different noise in neighboring frames even though the seed is fixed, whereas indexing the noise by coordinate keeps it identical on all voxels shared between frames.

Table B1: Ablation on MorphEdit3D over the SLAT noise strategy.

SLAT noiseFID↓KID↓LPIPS_adj↓PDV↓
Independent44.6111.940.1242.09
TRELLIS default (seed reset)36.719.450.0951.38
Coordinate-indexed (ours)33.198.440.0841.11

Q2: The natural controls for the claim “interpolating velocities is better” are simple alternatives under the same shared-noise setting. However, I didn't see such settings in the ablation experiment. Could the authors provide one?

Experimental setup: We conducted a 3×2 factorial ablation on MorphEdit3D, using the same TRELLIS backbone, sampler, and evaluation protocol, to examine how FlowMorpher compares with simpler alternatives under both shared and independent initial noise.

  • Interpolation level: The first factor is where the interpolation is applied and we interpolate on the input images, on the DINOv2 condition tokens, or on the velocity predictions (velocity field interpolation, VFI).
  • Initial noise: The second factor is the initial noise, which is either shared across all α or drawn independently for each α.

Results: The quantitative results are reported in Table B2. Two observations follow. (1) First, with shared noise, VFI improves over condition-token interpolation on every metric. Pixel blending is far behind. (2) Second, the two factors play different roles. Sharing the noise is what makes the sequence smooth as it roughly halves LPIPS_adj for both token interpolation and VFI. The space at which interpolation is applied governs the quality of the intermediates.

We attribute the gap between condition and velocity to the fact that VFI queries the network only at the two real image conditions, whereas a blend of patch tokens from two different images is a condition the model has not seen during training.

Table B2: Factorial ablation on MorphEdit3D over the initial noise strategy and the interpolation level.

NoiseInterpolationFID↓KID↓LPIPS_adj↓PDV↓
IndependentImage60.4425.930.1833.17
IndependentCondition tokens42.2111.820.1642.78
IndependentVelocity39.3910.500.1612.48
SharedImage49.5018.490.1182.16
SharedCondition tokens39.5411.280.0901.68
SharedVelocity (Ours)33.198.440.0841.11

[B1] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B. Tenenbaum. Compositional Visual Generation with Composable Diffusion Models. ECCV, 2022.

[B2] Liew, Jun Hao and Yan, Hanshu and Zhou, Daquan and Feng, Jiashi. Magicmix: Semantic mixing with diffusion models. arXiv, 2022.

[B3] He, Qiyuan and Wang, Jinghao and Liu, Ziwei and Yao, Angela. Aid: Attention interpolation of text-to-image diffusion. NeurIPS, 2024.

[B4] Yang, Zhaoyuan and Yu, Zhengyang and Xu, Zhiwei and Singh, Jaskirat and Zhang, Jing and Campbell, Dylan and Tu, Peter and Hartley, Richard. Impus: Image morphing with perceptually-uniform sampling using diffusion models. ICLR, 2024.

[B5] Xiang, Jianfeng and Lv, Zelong and Xu, Sicheng and Deng, Yu and Wang, Ruicheng and Zhang, Bowen and Chen, Dong and Tong, Xin and Yang, Jiaolong. Structured 3d latents for scalable and versatile 3d generation. CVPR, 2025.

AuthorsOpenReview ↗

Kind reminder: has our response addressed your concerns?

Dear Reviewer otXc,

We would like to thank you again for your insightful feedback. As the discussion period is coming to an end, we are wondering if the two major concerns about Relation to Composition Techniques and SLAT-Stage Contribution have been addressed.

  • Relation to Composition Techniques (W1, Q2): We now explicitly discuss the convex combination operator and its relation to Composable Diffusion, CFG-style condition mixing, MagicMix, AID and IMPUS in the revised Sec. 2 and Sec. 5.1. Our new ablation (Table B2) shows that, under shared noise, interpolating velocities outperforms interpolating conditions on all metrics.
  • SLAT-Stage Contribution (W2, Q1): We clarify in the revised Sec. 6 that the contribution of the SLAT stage depends on the pair type, and achieves better sequence smoothness on semantic-level pairs. Our new ablation (Table B1) shows that the coordinate-indexed SLAT noise improves all metrics over the default TRELLIS sampler.

If there is anything that remains unclear or if you have further questions or comments, please don't hesitate to let us know. If our responses have addressed your concerns, we would be grateful if you could update your assessment accordingly.

Thank you very much.

Best regards,

Submission 35725 Authors

Reviewer_BL8pOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FlowMorpher, a training-free and inversion-free method for textured 3D morphing with pretrained image-to-3D flow models. It constructs sampling trajectories directly from shared Gaussian noise, and define the sampling update rule by interpolating the model’s conditional velocity predictions. On TRELLIS, the rule is applied in both the sparse-structure (SS) and structured-latent (SLAT) stages. Experiments demonstrate promising results.

Critical strengths
  1. Simple and broadly applicable. The method needs only conditional forward passes at a shared state, which I appreciate.
  2. Clear analysis of a natural baseline. Sec. 4 traces the failure of morphing-by-editing to its source-anchored initialization and reference state, with the target entering only through conditioning.
Critical weaknesses
  1. Missing key baselines and ablations. a) shared noise with interpolated condition tokens (e.g., DINOv2 features) in TRELLIS; b) shared noise with TRELLIS applied to interpolated or blended input images.
  2. Efficiency. Avoiding "costly" inversion and caching is a main motivation, yet the paper reports no runtime or memory comparison.
  3. Related-work positioning. Although the paper includes a related-work section, it does not sufficiently discuss the works closest to its core idea. The paper should discuss Composable Diffusion[1], which similarly combines conditional predictions during sampling, and Wukong[2], which performs inversion-free 3D morphing through condition interpolation. Clarifying these differences would better establish the novelty.

References

[1] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B. Tenenbaum. Compositional Visual Generation with Composable Diffusion Models. ECCV, 2022.

[2] Minghao Yin, Yukang Cao, and Kai Han. Wukong’s 72 Transformations: High-fidelity Textured 3D Morphing via Flow Models. NeurIPS, 2025.

Questions for authors

See weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer BL8p (Part 1/2)

Thanks for your efforts and valuable suggestions. We address your concerns as follows.

W1: Missing key baselines and ablations. a) shared noise with interpolated condition tokens (e.g., DINOv2 features) in TRELLIS; b) shared noise with TRELLIS applied to interpolated or blended input images.

Experimental setup: We conducted a 3×2 factorial ablation on MorphEdit3D, using the same TRELLIS backbone, sampler, and evaluation protocol, to examine how FlowMorpher compares with simpler alternatives under both shared and independent initial noise.

  • Interpolation level: The first factor is where the interpolation is applied: on the input images, on the DINOv2 condition tokens, or on the velocity predictions (velocity field interpolation, VFI).
  • Initial noise: The second factor is the initial noise, which is either shared across all α or drawn independently for each α.

Results: The quantitative results are reported in Table A1. Two observations follow. (1) With shared noise, VFI improves over condition-token interpolation on every metric. Image blending is far behind. (2) The two factors play different roles. Sharing the noise is what makes the sequence smooth, it roughly halves LPIPS_adj for both token interpolation and VFI. The level at which interpolation is applied governs the quality of the intermediates.

We attribute the gap between condition-token and velocity interpolation to the fact that VFI queries the network only at the two real image conditions, whereas a blend of patch tokens from two different images is a condition the model has not seen during training.

Table A1: Factorial ablation on MorphEdit3D over the initial noise strategy and the interpolation level.

NoiseInterpolationFID↓KID↓LPIPS_adj↓PDV↓
IndependentImage60.4425.930.1833.17
IndependentCondition tokens42.2111.820.1642.78
IndependentVelocity39.3910.500.1612.48
SharedImage49.5018.490.1182.16
SharedCondition tokens39.5411.280.0901.68
SharedVelocity (Ours)33.198.440.0841.11

W2: Efficiency. Avoiding "costly" inversion and caching is a main motivation, yet the paper reports no runtime or memory comparison.

Experimental setup: We measured all methods on a single RTX 4090 for a 7-frame sequence on MorphEdit3D. Setup covers every per-pair cost before the first frame (2D morphing, inversion, feature caching, LoRA tuning), and peak memory is the maximum allocated GPU memory.

Results: The efficiency results are reported in Table A2. FlowMorpher has no setup stage and the lowest total time. Its total time is 20% lower than Nano3D’s, and it is 2.4× faster than Interp3D and 4.6× faster than VoxHammer, whose inversion alone takes 95 s.

Table A2: Efficiency comparison for generating a 7-frame sequence.

MethodInversion (MiB)Feature cache (MiB)Per-pair tuning3D maskSetup (s)Per-frame (s)Peak mem (GB)Total (s)
Image morphing + 3D generation
FreeMorph0.5—✗✗46.45.816.687.0
DiffMorpher262625✓✗59.15.715.999.2
3D editing adapted for morphing
VoxHammer47,38361✗✓95.129.515.8301.4
Nano3D——✗✓10.910.214.382.5
Native 3D morphing
Interp3D—2303✗✗17.120.716.9161.9
FlowMorpher (ours)——✗✗—9.415.766.1
AuthorsOpenReview ↗

Response to Reviewer BL8p (Part 2/2)

W3: Related-work positioning. Although the paper includes a related-work section, it does not sufficiently discuss the works closest to its core idea. The paper should discuss Composable Diffusion[1], which similarly combines conditional predictions during sampling, and Wukong[2], which performs inversion-free 3D morphing through condition interpolation. Clarifying these differences would better establish the novelty.

Thanks for your suggestion. Both papers are closely related and deserve a more explicit discussion. Composable Diffusion [A1] shares its basic operator with FlowMorpher, and Wukong [A2] is the closest prior work to ours, as it also performs inversion-free 3D morphing. We have added a dedicated discussion of both in Related Work (Sec. 2) and Method (Sec. 5) of the revised manuscript. The key differences are summarized below.

  • Composable Diffusion: The difference lies in what the combination is for. Composable Diffusion uses fixed guidance weights to produce a single sample that satisfies all concepts at once (a conjunction). FlowMorpher instead uses convex weights (1−α, α) as a continuous coordinate along a path between two objects, and integrates the resulting field from a single shared noise sample.
  • Wukong: The difference is where the interpolation happens. Wukong interpolates in condition space, avoids inversion and formulates morphing as an optimal transport barycenter problem with sequential initialization across frames. In contrast, FlowMorpher interpolates in velocity space: the generator is queried with the two real conditions, and the blending is applied to its outputs at a shared state. This requires no auxiliary optimisation, sequential initialisation strategy, or threshold, and the same rule applies to both stages.

Our contribution is therefore to show that velocity space is the right place to interpolate for morphing (see Table A1), and to analyse why this yields smooth, endpoint-faithful 3D sequences.


[A1] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B. Tenenbaum. Compositional Visual Generation with Composable Diffusion Models. ECCV, 2022.

[A2] Minghao Yin, Yukang Cao, and Kai Han. Wukong’s 72 Transformations: High-fidelity Textured 3D Morphing via Flow Models. NeurIPS, 2025.

The authors' rebuttal has addressed my main concerns. I will maintain my score and lean toward acceptance.

Thank you very much for your positive feedback. We are glad that our responses have addressed your main concerns. Your suggestions have substantially strengthened the manuscript. We sincerely appreciate your support for our paper.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

Dear Reviewers,

We appreciate all the reviewers for their insightful and constructive reviews of our manuscript. We are glad that the reviewers found that:

  • The method is simple but broadly applicable, and it is both training-free and inversion-free (Reviewer BL8p,otXc, andMAM7).
  • The analysis is clear and rigorous: Section 4 explains why morphing-by-editing fails, and the scope of the theory is stated precisely (Reviewer BL8p and otXc).
  • The morphing results are clean and impressive compared with prior training-free methods, and the comparisons clearly illustrate the feature-collage failure mode of existing approaches (Reviewer otXc and MAM7).
  • The paper is well-written, clear and easy to follow (Reviewer otXc and MAM7).
  • MorphEdit3D is a new benchmark that is essential for the community, as evaluating morphing is difficult (Reviewer MAM7).

Based on all the reviews, here we provide a general response to the concerns raised by multiple reviewers. The individual responses, with the full tables, are posted below each review.

  • Regarding questions about experiments, we have addressed the concerns as follows:
    • We add a 3×2 factorial ablation over where the interpolation is applied (input images, DINOv2 condition tokens, or velocities) and the initial noise (shared or independent). (Reviewer BL8p W1 and otXc Q2)
    • We add an efficiency comparison covering setup time, per-frame time, peak GPU memory and total time. (Reviewer BL8p W2 and MAM7 W5)
    • We add an ablation on the SLAT noise, comparing our coordinate-indexed noise with the default TRELLIS sampler and with independent noise. (Reviewer otXc Q1)
    • We add an evaluation at α = 1/2. (Reviewer MAM7 W4)
    • We add a user study comparing FlowMorpher with baselines on MorphEdit3D. (Reviewer MAM7 W2)
  • Regarding questions about novelty and related work, we have addressed the concerns as follows:
    • We clarify how FlowMorpher relates to Composable Diffusion and CFG-style condition mixing (Reviewer BL8p W3 and otXc W1), to 2D mixing and interpolation methods such as MagicMix, AID and IMPUS (Reviewer otXc W1), and to recent 3D morphing methods, Wukong and FLOWING (Reviewer BL8p W3 and MAM7 W3).
  • Regarding questions about scope and analysis, we have addressed the concerns as follows:
    • We state the clarification on semantically distant pairs explicitly in the main text. (Reviewer MAM7 W1)
    • We clarify the contribution of the SLAT stage relative to the SS stage in Sec. 6. (Reviewer otXc W2)
    • We add the source and target input images to Figures 6–12. (Reviewer MAM7 W4)

We have also uploaded a revised version of our submission, with major changes highlighted in [purple]. The changes are summarized as follows:

  • Introduction: We clarify the scope of FlowMorpher (Reviewer MAM7 W1) and state the ablation against simple alternatives under the same shared-noise setting (Reviewer BL8p W1 and otXc Q2).
  • Sec. 2 (Related Work): We substantially extend the positioning of FlowMorpher with respect to related work, covering methods that share its basic operator (Composable Diffusion and CFG-style condition mixing), 2D mixing and interpolation methods (MagicMix, AID and IMPUS), and the closest 3D morphing methods (Wukong and FLOWING). (Reviewer BL8p W3, otXc W1 and MAM7 W3)
  • Sec. 5.1: We add a new paragraph discussing the convex combination operator. (Reviewer BL8p W3 and otXc W1)
  • Sec. 6 (Experiments):
    • We update Table 1 with the results at α = 1/2 and the average generation time. (Reviewer MAM7 W4 and W5)
    • We add Table 2 and a new paragraph reporting the 3×2 factorial ablation over the initial noise and the interpolation level. (Reviewer BL8p W1 and otXc Q2)
    • We clarify the statement on the contribution of the SLAT stage in the ablation study. (Reviewer otXc W2)
    • We add an ablation on the SLAT noise strategy, with the full results reported in Table 4 (Appendix A.1). (Reviewer otXc Q1)
  • Appendix:
    • A.2.4: We add the full efficiency comparison (Table 5) and a new paragraph describing it. (Reviewer BL8p W2 and MAM7 W5)
    • A.2.5: We add a user study (Table 6) comparing FlowMorpher with the baselines. (Reviewer MAM7 W2)
  • Figures 6–12: We add the source and target input images. (Reviewer MAM7 W4)

We look forward to further feedback and discussion. Please feel free to let us know if further details or explanations would be helpful.

Once again, we thank all reviewers for their thoughtful feedback and valuable guidance in improving our work.

Best regards,

Submission 35725 Authors

AuthorsOpenReview ↗

Summary Response to ACs

Dear all ACs and Reviewers,

We sincerely thank all reviewers for their efforts and constructive feedback throughout the discussion period. Below, we provide a concise summary of the reviewers' recognitions and our corresponding responses during the rebuttal phase.


Reviewers' Positive Feedback:

Reviewers BL8p and MAM7 both responded promptly and positively. We are encouraged by their two positive ratings of 3 (weak acceptance): Reviewer BL8p leans toward acceptance, and Reviewer MAM7 is leaning towards raising the score. Their feedback highlights:

  • Concerns addressed by the rebuttal: Reviewer BL8p confirmed that our rebuttal "has addressed my main concerns". Reviewer MAM7 confirmed that we "addressed all my concerns".
  • Stronger empirical evidence: The new experiments, including the evaluation, the efficiency comparison and the user study, strengthen the empirical evidence supporting the proposed method (Reviewer MAM7).
  • A more complete evaluation of strengths and limitations: The revised manuscript gives a more complete picture of the method, including a clearer discussion of its limitations on large semantic transformations (Reviewer MAM7).
  • Technical contribution: Reviewer MAM7 recognized that the paper provides a technical contribution.

Responses to Reviewers' Concerns:

To address the reviewers' concerns, we conducted additional experiments and analyses, including:

  • Positioning with respect to related work (Reviewer BL8p W3, otXc W1, MAM7 W3): We extended Sec. 2 and Sec. 5.1 to clarify how FlowMorpher differs from existing work.
  • Ablation against simple alternatives (Reviewer BL8p W1, otXc Q2): A 3×2 factorial ablation over the interpolation level (image / condition token / velocity) and the initial noise (shared / independent) shows that velocity interpolation outperforms both alternatives on all metrics.
  • SLAT stage and SLAT noise (Reviewer otXc W2, Q1): We clarified that the SLAT contribution depends on the pair type, and showed that coordinate-indexed noise improves all metrics over the default TRELLIS sampler.
  • Efficiency comparison (Reviewer BL8p W2, MAM7 W5): FlowMorpher has no setup stage and the lowest total time.
  • Evaluation at α = 1/2 and user study (Reviewer MAM7 W2, W4): FlowMorpher remains the best method at the midpoint, with a wider margin, and is preferred over all baselines on every criterion in a 30-participant user study.
  • Scope and limitations (Reviewer MAM7 W1, W4): We moved the discussion of semantically distant pairs into the main text and added the input images to Figures 6–12.

All corresponding numerical experiments, visualizations and explanations are included in our discussion responses and the revised manuscript.


Discussion Status:

Reviewers BL8p and MAM7 both responded promptly and positively during the discussion phase, confirming that our rebuttal resolved their concerns:

  • Reviewer BL8p stated that the rebuttal "has addressed my main concerns", maintains the positive rating, and leans toward acceptance.
  • Reviewer MAM7 stated that we "addressed all my concerns", and they indicated that they are leaning towards raising the score.

We have not yet received a response from Reviewer otXc. For the AC's convenience, we briefly summarize how their concerns were addressed:

  • W1 (main concern, relation to composition and interpolation techniques): This concern overlaps closely with the positioning concerns of Reviewers BL8p (W3) and MAM7 (W3). Composable Diffusion in particular was raised by both otXc and BL8p. Our revised Sec. 2 and the new paragraph in Sec. 5.1 address it, and both BL8p and MAM7 have confirmed that their concerns were resolved.
  • W2 (SLAT-stage contribution): We stated that the SLAT contribution depends on the pair type, and achieves better sequence smoothness on semantic-level pairs. We have clarified the analysis in Sec. 6.
  • Q1 (SLAT-stage's coordinate-indexed noise): We added a new ablation (Table B1) shows that the coordinate-indexed SLAT noise improves all metrics over the default TRELLIS sampler.
  • Q2 (controls under the same shared noise setting): This is the same factorial ablation that Reviewer BL8p requested (W1), and BL8p confirmed that it addressed their concern. The factorial ablation (Table B2) answers the question empirically, where interpolating velocities outperforms interpolating conditions on all metrics.

By now, the overall ratings remain (3, 3, 2), unchanged from the initial ratings, and Reviewer MAM7 has indicated that they are leaning towards raising the score.


We hope this summary gives the AC a clear picture of our work and of how the reviewers' concerns were resolved. Thank you again for your time and effort. If any questions remain, we are happy to provide further clarification.

Best regards,

Submission 35725 Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.