acceptodds
Under review as a conference paper at ICLR 2027

Let the Pixels Through: Rethinking Long Skip Connections for Scalable Pixel Diffusion

Abstract

Training pixel diffusion models remains less straightforward, owing to the ad hoc training recipes and architecture designs, along with the sensitivity to hyper-parameter choices. We show that strong denoisers can be built by refining existing pixel-native designs from the literature. Specifically, we revisit the long skip connection in U-Net and find that the optimization benefit commonly attributed to it is not necessary for its denoising benefit. We further find that the conventional linear combination of features from the skip and main backbone is inadequate for regimes of various noise levels. Motivated by these observations, we introduce two core modifications. First, we adopt asymmetric long skips, which propagate fine-grained information from input embeddings to decoder layers. Second, we introduce adaGLU, a SwiGLU-like MLP block, in decoder layers to adaptively filter information from the embeddings before being added to main-branch features. The above designs give a simple pixel diffusion transformer architecture. When trained from scratch, it scales smoothly from small-scale baselines up to 1.6B parameters, achieving 1.62 and 1.70 FID on ImageNet-256 and -512, respectively. Besides, it natively supports multimodal inputs, such as text and images, yielding competitive performance in text-to-image generation.

Primary area:
generative models
Keywords:
image generation, pixel-space diffusion model, model architecture

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

This paper introduces PiT, a pixel diffusion transformer that replaces symmetric U-Net-style skip connections with asymmetric skips and introduces adaGLU for adaptive feature fusion. The method achieves competitive results on ImageNet-256/512 and demonstrates good scalability and text-to-image generation performance.

Critical strengths
  1. Strong experimental results. PiT-G/16 achieves competitive FID scores of 1.62 and 1.70 on ImageNet-256 and ImageNet-512, respectively, with good scalability across model sizes.
  2. Simple and effective design. The proposed asymmetric skips and adaGLU improve generation quality without requiring additional pretraining or complex training procedures.
Critical weaknesses
  1. Limited novelty. The main architectural components are adaptations of existing skip connections and SwiGLU, making the methodological novelty somewhat limited.
  2. Insufficient experimental analysis. The paper lacks straight evidence of noise-adaptive gating behavior and parameter-matched comparisons to isolate the benefits of adaGLU from increased model capacity.
Questions for authors
  1. Could the authors analyze how adaGLU's gating behavior changes across different noise levels?
  2. How does adaGLU compare with standard MLP-based fusion under matched parameter counts or FLOPs?
  3. Could the authors clarify the performance comparison with JiT under equal training budgets, particularly on ImageNet-512?
2 comments

Thank you for your valuable feedback and recognition of our work. We provide our responses below and hope they could properly address your concerns.

W.1 Limited Novelty

We agree that long skip connections and SwiGLU are established primitives. Our contribution lies not in claiming either primitive as new, but in offering a novel perspective on their roles and organization in pixel diffusion.

  • First, we empirically separate the information-routing and optimization roles of long skips. Detaching skip gradients does not degrade performance (52.1 → 50.1 FID), and replacing symmetric encoder–decoder connections with asymmetric input-to-decoder pathways slightly improves FID to 47.8.

  • Second, although adaGLU is implemented using a standard SwiGLU MLP, its key distinction is that it introduces multiplicative interaction between two distinct representations (the backbone and skip features) rather than applying SwiGLU-style self-modulation to a single input.

  • Finally, these findings yield a simple architecture that scales to 1.6B parameters without AdaLN, in-context embeddings, pretraining, auxiliary losses, or scale-specific hyperparameters.

We hope these clarifications help you reassess the contributions of our work.

W.2

(Also Q.1) Direct evidence of noise-adaptive gating behavior

We visualize the relative contribution of adaGLU across noise levels and decoder layers in Appendix Figure 6 and discuss the results in Section D. For decoder layer ℓ\ell, let zℓz_{\ell} denote the backbone feature before fusion and Δzℓ=adaGLU(hℓ,zℓ)\Delta z_{\ell}=\text{adaGLU}(h_{\ell},z_{\ell}) denote the adaGLU update. We measure the relative update magnitude $r_{\ell}(\sigma)=\frac{\lVert \Delta z_{\ell}\rVert_2 }{ \lVert z_{\ell} \rVert_2} $, averaged over 1000 samples at each noise level σ\sigma. The results show that rℓ(σ)r_{\ell}(\sigma) decreases as the noise level increases across the examined decoder layers.

Since the features are fused by addition: zℓout=zℓ+Δzℓz_{\ell}^{out}=z_{\ell}+\Delta z_{\ell}, this ratio measures the relative change that the adaGLU update induces in the backbone representation. Its decrease at higher noise levels therefore shows that the relative contribution of the adaGLU update diminishes. This observation is consistent with the network relying less on skip-path information when the input is heavily corrupted.

Conversely, the larger relative updates at low noise levels support our hypothesis that adaGLU helps propagate fine-grained input information to deeper layers to facilitate pixel prediction (Section 4.2, lines 249–254).

(Also Q.2) Parameter-matched comparison with the baseline without adaGLU

To isolate the benefit of adaGLU from increased model capacity, we compare it with the standard MLP fusion baseline without adaGLU under the same training configuration. We increase the baseline’s hidden dimension to 1344 so that it has slightly more parameters than PiT-XL (709M v.s. 640M).

modelFID-10K#params
PiT-XL30.8640M
PiT-XL w/o adaGLU62.0709M

Training Settings: ImageNet-256, 200K training steps, batch size 256.

Despite having fewer parameters, PiT-XL substantially outperforms the baseline, achieving 30.8 versus 62.0 FID-10K. This improvement cannot be explained by an increase in model capacity alone and instead supports the benefit of backbone-conditioned adaptive fusion.

Q.3 Performance comparison with JiT at equal training epochs on ImageNet-512

We provide an epoch-matched comparison with JiT on ImageNet-512 in Appendix Table 10 and attach the results below.

model#paramsFIDtraining epochs
JiT-L462M2.53600
PiT-L343M2.92600
2.68800
JiT-H956M1.94600
PiT-XL646M2.31600
2.10800
JiT-G2B1.78600
PiT-G1.6B1.83600
1.72800

The FID scores above are computed using ImageNet-512 training-set statistics, following JiT, rather than the ADM statistics used in the main tables.

In the 600-epoch comparison on ImageNet-512, JiT-L and JiT-H outperform PiT-L and PiT-XL, respectively, while PiT-G nearly matches JiT-G (1.83 v.s. 1.78 FID) with approximately 20% fewer parameters. The performance gap narrows as model scale increases. With longer training and unchanged hyperparameters, all PiT models improve; PiT-G reaches 1.72 FID at 800 epochs. On ImageNet-256, PiT-G outperforms JiT-G at 600 epochs (1.61 v.s. 1.82).

These results support our claim that PiT continues to improve with longer training without scale-specific hyperparameter tuning. They do not imply uniform superiority over JiT.

Thank you for your constructive advice to help us strengthen our work. We hope these clarifications and additional experiments address your concerns, and we would be grateful if you would reconsider your assessment in light of them.

Reviewer_EYsYOpenReview ↗

Response to Authors' Rebuttal

Thanks for your detailed responses. I have carefully checked the rebuttal, and am particularly impressed by the significant improvements brought by adaGLU. All my concerns are addressed completely, and I will maintain my positive assessment.

Reviewer_k7bXOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper revisits the role of long skip connections in pixel-space diffusion transformers. Starting from experiments that detach gradients along the skip pathways, the authors investigate whether their denoising benefit requires conventional symmetric encoder–decoder connections. They propose routing input embeddings directly to decoder layers and introducing adaGLU to modulate the skip information using backbone features. Together with appropriate initialization and a Pre-Norm formulation, the resulting PiT architecture achieves strong ImageNet generation results and also supports text-to-image generation.

Critical strengths
  1. The trade-off underlying the default layer allocation could be clearer. Section 4.4 and Table 12 already specify and ablate the encoder–decoder allocation. However, allocating more layers to the decoder improves FID, while the main experiments retain equal encoder and decoder depths. Reporting the corresponding parameter counts and computational costs would help explain the practical motivation for this default.

  2. The importance of the adaGLU branch assignment is not directly isolated. The paper explains the input embedding as the information carrier and the backbone feature as the gate. This interpretation is intuitive, but Table 3 does not compare it with a variant that swaps these roles. Such a comparison would strengthen the justification for this particular design.

  3. The scope beyond pixel-space diffusion remains open. The experiments establish the effectiveness of the proposed mechanism in pixel space, but do not determine whether its benefits extend to compressed latent representations. A discussion or limited evaluation using a latent-space architecture with long skips, such as U-ViT, would help clarify the generality of the findings. I regard this as a scope question rather than a major weakness of the current contribution.

Critical weaknesses
  1. What motivates retaining equal encoder and decoder depths despite the improvement from a deeper decoder in Table 12(a)? How do parameter count and computation change across these configurations?

  2. Have the authors evaluated swapping the linear and gating inputs in adaGLU? Alternatively, could an analysis of the gated skip contribution across noise levels provide more direct evidence for the proposed adaptive-routing interpretation?

  3. Do the authors expect the proposed input-embedding pathway and adaptive fusion to benefit latent-space models with long skips, such as U-ViT? Which aspects of the observed gains are likely specific to pixel-space denoising?

Questions for authors
  1. In Table 12(a), increasing decoder depth improves FID. Could the authors elaborate on why an equal encoder–decoder depth allocation was selected as the default? Are there computational or optimization considerations that favor this configuration?

  2. In Eq. (5), the input embedding is passed through the linear branch, while the backbone feature is used for gating. Have the authors considered swapping these two inputs? I believe such an ablation would help clarify whether this particular assignment is essential to the effectiveness of adaGLU.

  3. Have the authors considered applying the proposed asymmetric skip connections and adaGLU to latent-space diffusion architectures, particularly those with long skip connections? It would be interesting to understand whether the observed improvements are specific to pixel-space denoising or reflect a more general architectural principle.

3 comments

Thank you for your valuable feedback and recognition of our work. We attach our responses below and hope they could properly address your concerns.

(Part 1/2)

W.1 and Q.1: Design rationale of equal encoder-decoder depth

Table 12(a) indeed suggests that allocating more layers to the decoder can further improve performance, and we agree in Section 4.4 that this is a promising direction. We adopt an equal encoder-decoder depth primarily for simplicity, rather than because this allocation is optimal. There is also a computational consideration. Since each decoder layer contains an additional adaGLU block, shifting layers from the encoder to the decoder increases both parameters and computation even when the total depth is fixed. As a result, the equal-depth configuration provides a simpler and balanced default, while reaching sufficiently decent performance.

We have added the exact parameter and FLOP counts for these configurations in Table 12(a). For your convenience, we attach the results below.

encdecFID-10K#ParamsFLOPs
151530.8640M165
62429.1736M190
24633.2545M141

W.2

(Also Q.2) Swapping the input order of the adaGLU

Recall that the formulation of adaGLU is defined as (in Equation 5): adaGLU(h,z)=W1(W2h⊙SiLU(W3z)).\text{adaGLU}(\bf{h},\bf{z})=W_{1}(W_{2}\bf{h}\odot\text{SiLU}(W_{3}\bf{z})). Here, h\bf{h} and z\bf{z} are skip features and backbone features, respectively. We compare the performance of the baseline adaGLU(h,z)\text{adaGLU}(\bf{h},\bf{z}) and the variant adaGLU(z,h)\text{adaGLU}(\bf{z},\bf{h}), where we swap the order of input. We show the ablation results in the following Table.

modelFID-10K
PiT-B57.5
Swap input order56.0
PiT-XL30.8
Swap input order30.0

Training Settings: ImageNet-256, 200K training steps, batch size 256

We observe both variants achieve similar performance, indicating that adaGLU’s benefit does not depend strictly on which representation enters the non-linear branch.

Both variants preserve a learned multiplicative interaction between skip and backbone features, using the backbone representation to regulate information propagated through the skip pathway. They differ primarily in which representation undergoes the additional nonlinear transformation.

We would like to emphasize that adaGLU does not designate the nonlinear branch alone as the gate. Swapping the inputs changes only which representation undergoes the additional SiLU transformation, while preserving this cross-feature multiplicative modulation.

We have clarified this interpretation in Section 4.2.

Analysis of the gated skip contribution across noise levels

We visualize the relative contribution of adaGLU across noise levels and decoder layers in Appendix Figure 6 and discuss the results in Section D. For decoder layer ℓ\ell, let zℓz_{\ell} denote the backbone feature before fusion and Δzℓ=adaGLU(hℓ,zℓ)\Delta z_{\ell}=\text{adaGLU}(h_{\ell},z_{\ell}) denote the adaGLU update. We measure the relative update magnitude $r_{\ell}(\sigma)=\frac{\lVert \Delta z_{\ell}\rVert_2 }{ \lVert z_{\ell} \rVert_2} $, averaged over 1000 samples at each noise level σ\sigma. The results show that rℓ(σ)r_{\ell}(\sigma) decreases as the noise level increases across the examined decoder layers.

Since the features are fused by addition: zℓout=zℓ+Δzℓz_{\ell}^{out}=z_{\ell}+\Delta z_{\ell}, this ratio measures the relative change that the adaGLU update induces in the backbone representation. Its decrease at higher noise levels therefore shows that the relative contribution of the adaGLU update diminishes. This observation is consistent with the network relying less on skip-path information when the input is heavily corrupted.

Conversely, the larger relative updates at low noise levels support our hypothesis that adaGLU helps propagate fine-grained input information to deeper layers to facilitate pixel prediction (Section 4.2, lines 249–254).

(Part 2/2)

W.3 and Q.3. Does our design benefit latent-space models such as U-ViT

We evaluate asymmetric skips and adaGLU on latent-space U-ViT-B.

modelFID-10K
U-ViT-B68.7
U-ViT-B + asym skip69.0
U-ViT-B + adaGLU62.0

Training Settings:ImageNet-256, 200K training steps, batch size 256. We use sd-vae-ft-ema (https://huggingface.co/stabilityai/sd-vae-ft-ema) as the VAE. All latent-space variants use the same pixel-derived hyperparameters without latent-specific tuning. These experiments therefore provide a controlled preliminary architectural comparison rather than an optimized latent-diffusion benchmark.

Replacing the U-Net-style skips with asymmetric skips leaves performance essentially unchanged, whereas adding adaGLU improves FID-10K to 62.0. Thus, the benefit of adaGLU is consistent with our pixel-space findings, while the effect of asymmetric skips differs between the two experiments. For reference, we summarize the analogous pixel-space ablations below

modelFID-10K
U-ViT-B52.1
U-ViT-B + asym skip47.8
U-ViT-B + adaGLU41.2

Training Settings: ImageNet-256, 200K training steps, batch size 1024,

These preliminary results suggest that adaGLU’s multiplicative feature fusion is not inherently pixel-specific and can also benefit latent-space diffusion. In contrast, the benefit of asymmetric input routing may depend more on the input representation. We have revised our discussion in the Appendix, section D, to reflect this distinction.

Again, we would like to thank you for providing the constructive advice to help us strengthen our work. We hope these clarifications and additional experiments address your concerns.

Reviewer_k7bXOpenReview ↗

after rebuttal

Thank you for the detailed responses and additional experiments. These have addressed my main questions and reinforced my positive assessment of the paper. I will maintain my original rating.

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

This paper re-analyze the long skip connection in pixel diffusion models, especially on diffusion transformer architectures, and propose PiT. This paper investigates and ablates many designs of long skip connections in pixel diffusion transformers and finds that detach patch embeddings also can help the convergence of pixel diffusion transformers, showing that low level information is essential for pixel diffusion transformers. The proposed PiT show competitive results on ImageNet-256 and 512, and several T2I benchmarks.

Critical strengths
  1. This paper is clear to follow and sound with regard to experiments. The motivation about hyper connections for pixel diffusion transformers is straightforward and the writting is easy to follow.

  2. The most vaulable point of this paper, is the finding that the low level token embedding, even detached one, helps a lot the performance of the convergence of pixel diffusion transformers, which the previous paper has little explored, especially on detached shortcuts.

Critical weaknesses
  1. While the improvements from long skip connections, particularly those involving xt or patch embeddings, are valuable, they are broadly consistent with current understanding and recorded by these year papers of pixel diffusion transformers without surprise, starting from PixNerd with xt in the nerf head, and then JiT using xt to transform predictions, then PixelDiT and other following-ups. Such shortcuts are widely used to help or "hack" transformers with limited hidden dimensions produce high-dimensional predictions. However, their benefits may diminish as the hidden representations gain sufficient capacity to preserve the relevant input and noise information. The paper’s claim that this design scales effectively therefore requires stronger justification, particularly comparison with plain diffusion transformers to predict plain x0 (not JiT-style x0) or velocity with large model settings.

  2. The pixel diffusion transformers are vulnerable for their patching grid effects. The figures in the paper, including Figure 4, are highly JPEG-compressed and make it hard to recognize whether it is grid effect or not. Could author provides raw png images or report grid effect relevant metrics, especially PiT/32 ones?

Questions for authors
  1. The authors mentioned no AdaLN in line 115 but comes with no arguments or explanations on no AdaLN designs in other parts of this paper. Why authors are motivated to remove AdaLN? If AdaLN presented, is it a better and stronger baseline?
5 comments

Thank you for your valuable feedback. We’ve attached our replies below and we hope they could properly address your concerns.

W.1

Relation to prior works

We agree that using input information through long-range pathways is not itself new, and that our findings are consistent with the broader understanding established by PixNerd, JiT, PixelDiT, and related work.

Our contribution is not to claim the first use of embeddings from shallow layers. Instead, we investigate a complementary question: which structural components of conventional long-skip connections are necessary for denoising, and how should skip features be fused with backbone features?

We separately examine skip topology, optimization, and feature fusion. Detaching skip gradients does not degrade performance (52.1 → 50.1 FID), suggesting that the denoising benefit of long skips does not depend on their conventional optimization role. Replacing symmetric encoder–decoder connections with asymmetric input-to-decoder pathways further improves FID to 47.8, while adaptive multiplicative fusion with adaGLU improves it to 36.2.

These results distinguish our focus on cross-layer information routing and fusion from simply introducing input information into the architecture.

Comparison with plain DiT at larger scales

We also agree that PiT’s scaling behavior alone does not establish that its advantage over plain DiT persists with increasing capacity. We therefore compare PiT with plain DiT using x0−x_0-prediction under the same training and evaluation settings.

modelFID-10K#params
PiT-L37.4339M
PiT-XL30.8640M
DiT-L51.4510M
DiT-XL47.1675M

Training settings: ImageNet-256, 200K training steps, batch size 256.

PiT substantially outperforms DiT despite using fewer parameters. These results indicate that the benefit of our design persists at the tested model scales and training budgets. It's possibile that scaling up model size of plain DiT could reduce the gap, yet the computational cost could be prohibitively expensive.

W.2 Patching grid effect of pixel diffusion transformer

We agree that the compression in the original figures makes patch-grid artifacts difficult to assess. To enable direct inspection, we have uploaded 200 randomly sampled, uncurated, lossless PNG images generated by PiT-XL/32 on ImageNet-512 using the 800-epoch checkpoint and the sampling configuration detailed in Appendix Table 9. We do not observe any discernible patch-grid artifacts in these final-checkpoint samples.

ICLR official download link to the zip file: https://openreview.net/attachment?id=UZfJEzWaom&name=supplementary_material

Q.1 Motivations for not using AdaLN

In our early scale-up experiments, AdaLN helped mitigate the optimization difficulty discussed in Section 4.3. After adopting our current residual initialization, however, we found that AdaLN was no longer necessary for stable training. To examine its remaining benefit, we compare PiT-XL with and without AdaLN under otherwise identical training settings.

modelFID-10K#params
PiT-XL15.6640M
PiT-XL w/ adaLN15.0698M

Training settings: ImageNet-256, training steps 250K, batch size 1024.

AdaLN yields a small improvement, from 15.6 to 15.0 FID-10K, while introducing an additional 58M parameters. We therefore do not claim that AdaLN is ineffective. Rather, our default initialization already enables stable optimization and strong performance without it. We omit AdaLN to retain a simpler architecture with comparable performance and fewer parameters.

We have added these ablation results to Appendix Table 12(e).

Thank you for your constructive advice to help us strengthen our work. We hope these clarifications and additional experiments address your concerns, and we would be grateful if you would reconsider your assessment in light of them.

Dear Reviewer,

Thank you again for your valuable feedback. We would greatly appreciate it if you could let us know whether our responses satisfactorily address your concerns. If any points remain unresolved or unclear, we would be grateful for the opportunity to provide further explanation during the discussion period.

Best regards,
The Authors

Dear Reviewer Luzk,

Thank you again for your initial review and valuable comments. As the discussion period is approaching its end, we would greatly appreciate any further feedback you may have on our responses. In particular, we would be grateful to know whether our responses have adequately addressed your concerns or whether any points require further clarification.

We look forward to hearing from you.

Best regards,

The Authors

Comments on the paper

4 comments
~Rajat_R_Rasal1OpenReview ↗

Potential Applications for Disentangled Representation Learning

Dear Authors,

I really enjoyed reading this piece and found it a very interesting and insightful contribution to the field!

We explored a related idea earlier this year about skip connections in our work on diffusion representation learning, which was accepted at NeurIPS 2026 (https://arxiv.org/abs/2607.05319). We also introduced a simpler (and admittedly, more rudimentary) gating mechanism, using gated additive skips rather than concatenation-based skip connections, focusing on a U-Net backbone rather than a ViT. We noticed improved stability and disentanglement when combined with our other methodological changes.

I would be very keen to test your gating setup in our setting when your code becomes available! I would be interested to see whether it could further improve stability and performance in diffusion and flow-based disentanglement, where metrics and optimisation can be very noisy.

I just wanted to express my appreciation for your insights on this problem and highlight a potential further application of your method - I would be keen to hear your thoughts!

Best of luck with the rest of the review process!

Hi, Rajat,

Thank you for sharing SteeringDRL. It would be interesting to investigate how adaGLU affects representation disentanglement in diffusion autoencoders and whether these effects translate into improved unsupervised segmentation. We would be delighted for you to try our approach in your setting, and we look forward to seeing how it performs in your setting once our code is available.

Thank you again for highlighting this potential application and for your thoughtful feedback.

Best regards, The Authors

Area_Chair_hJAxOpenReview ↗

Please engage with the authors

Dear Reviewers Luzk & k7bX,

Please read the authors' response. Evaluate whether your concerns are addressed or if anything needs more attention. Please engage with the authors while we have time.

Thanks, Your AC

AuthorsOpenReview ↗

Summary of Rebuttal

Dear Area Chair,

Thank you for facilitating the discussion. We would like to briefly summarize the reviewer concerns and how we have addressed them.

1. Summary of concerns

The concerns raised across the reviews can be grouped into five areas:

  • Novelty: Whether asymmetric skip connections and adaGLU offer sufficient methodological contributions beyond existing approaches. (Luzk, EYsY)
  • Effectiveness of adaGLU: Its contribution beyond model capacity, noise-adaptive behavior, and the importance of its input assignment. (k7bX, EYsY)
  • Architectural choices: The rationale for equal encoder–decoder depth and omitting AdaLN, including their computational implications. (k7bX, Luzk)
  • Generality and scalability: Comparisons with plain DiT and JiT, and potential benefits for latent-space models such as U-ViT. (Luzk, k7bX, EYsY)
  • Artifacts: Potential patch-grid artifacts in generated images, particularly PiT/32. (Luzk)

2. Concerns confirmed as addressed

Reviewers k7bX and EYsY both explicitly confirmed that our rebuttal and additional experiments had addressed their concerns. Both expressed satisfaction with our responses and indicated that they would maintain their original positive ratings.

3. Concerns addressed but not followed up

Reviewer Luzk has not responded to our rebuttal despite several follow-up messages. Nevertheless, we believe we have adequately addressed the concerns through additional experiments and clarifications:

  • Novelty and scalability: We clarified the distinction of our contributions and added comparisons against plain DiT at larger scales. PiT substantially outperforms DiT under matched training settings while using fewer parameters.
  • Patch-grid artifacts: We provided 200 randomly sampled, uncurated, lossless PNG images generated by PiT-XL/32 on ImageNet-512 for direct inspection. No discernible patch-grid artifacts were observed.
  • AdaLN: We clarified our motivation for omitting AdaLN and provided additional ablations showing that it offers only marginal improvement (15.6 vs. 15.0 FID) at the cost of 58M additional parameters.

Although we regret not receiving further feedback from Reviewer Luzk, we believe these additional results and explanations substantively address the concerns raised.

Finally, we would like to sincerely thank all reviewers for their constructive feedback, and thoughtful engagement throughout the review process.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.