acceptodds
Under review as a conference paper at ICLR 2027

CINDI: Coarse-to-Fine Implicit Neural Diffusion for Efficient Pixel-Space Generation

Abstract

Pixel-space diffusion and flow-based models typically perform every denoising step at the full image resolution. Although early steps primarily establish coarse structure, sampling them still incurs the full cost of fine-resolution processing. We ask whether spatial resolution can instead be scheduled along the sampling trajectory, allowing early function evaluations to operate on coarser grids while preserving final generation quality. To answer this question, we present CINDI (Coarse-to-fine Implicit Neural DIffusion), a pixel-space generative model that progressively increases spatial resolution along the sampling trajectory. CINDI features an end-to-end implicit transformer denoiser trained with the next-scale bridge loss to enable both in-scale and cross-scale prediction, and a companion coarse-to-fine sampler that operates without the need for multi-model cascading or latent-space compression. We compare CINDI against a single-scale JiT baseline that achieves state-of-the-art-level performance on ImageNet. With matched model size, training epochs, and sampling steps, CINDI achieves competitive FID scores while delivering up to a 3× speedup in inference throughput. Additionally, benefiting from its multiscale prediction capability, CINDI also natively supports generation beyond its training resolution without extra fine-tuning.

Primary area:
generative models
Keywords:
Diffusion Models, Coarse-to-Fine Image Generation, Implicit Neural Representations, Flow Matching, Efficient Pixel-Space 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_1c8EOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents CINDI, a pixel-space generative model that improves inference efficiency through coarse-to-fine sampling. A shared Transformer and implicit neural renderer predict clean images at both the active and next resolutions. The next-scale bridge loss supervises the higher-resolution predictions used at scale transitions. Experiments demonstrate comparable FID to strong JiT baselines with substantially higher throughput.

Critical strengths
  • Clear presentation and coherent design. The paper is well written and easy to follow. The motivation, two-resolution prediction interface, training objective, and sampling procedure are clearly connected. Each component serves the concrete goal of reducing expensive full-resolution computation.

  • Useful supervision for scale transitions. Jointly supervising the current and next resolutions provides a direct training signal for the predictions used during sampling. The ImageNet-64 ablations support its effectiveness. For example, adding the next-scale loss improves FID from 2.75 to 2.19 in the 32→64 setting. I consider its integration into the shared implicit denoiser a meaningful aspect of the method.

  • Encouraging efficiency gains with minimal parameter overhead. The ImageNet-256 models introduce less than 0.5% additional parameters while achieving approximately 2.0–2.2× higher BF16 throughput at comparable FID. The matched JiT comparisons provide useful controlled evidence. Figure 4 also shows an appealing quality–throughput trade-off among the evaluated operating points, with CINDI-L achieving the highest measured throughput.

Critical weaknesses
  • The relationship to prior cross-resolution objectives needs clarification. Next-scale prediction has precedents in visual generation and language modeling. More directly, Scale Space Diffusion already supervises clean-image prediction at the next resolution from a potentially lower-resolution noisy input using weighted MSE (§4.2, Eq.7). CINDI’s joint current/next-resolution supervision and implicit rendering interface appear to provide a more specific distinction, but this relationship should be discussed explicitly to establish the methodological contribution.

  • The central loss ablation is limited to ImageNet 64. Since the practical high-resolution results are an important part of the paper, removing the next-scale loss in the 128→256 setting would provide stronger evidence that this component remains effective at the main evaluation resolution.

Questions for authors

-Could the authors clarify the distinction between the proposed next-scale bridge loss and prior cross-resolution prediction objectives, particularly Scale Space Diffusion? Which benefits arise specifically from jointly supervising both resolutions through the shared implicit renderer?

  • Could the authors report an ImageNet-256 ablation without the next-scale loss, keeping the remaining training and sampling settings fixed?
3 comments

We sincerely thank the reviewer for their time, thoughtful feedback, and questions. We respond to each point below.

Q1. Could the authors clarify the distinction between the proposed next-scale bridge loss and prior cross-resolution prediction objectives, particularly Scale Space Diffusion? Which benefits arise specifically from jointly supervising both resolutions through the shared implicit renderer?

Yes, we'd like to thank the reviewer for bringing Scale Space Diffusion (SSD) [1] to our attention. We were previously unaware of this work and will cite it in the revised manuscript.

Below, we will first briefly discuss how SSD differs from CINDI, and then address the reviewer’s questions accordingly.

  • Resolution changes: SSD puts downsampling inside the forward process, so resolution is fixed by the timestep. CINDI changes resolution between flow segments, and each segment is a standard flow at its own resolution.
  • Noise at the hand-off: SSD's exact reverse step at a transition needs correlated noise (sampled with Lanczos). CINDI re-noises with plain independent noise.

What is supervised and how CINDI and SSD supervise its next-resolution, respectively:

  • SSD gives each step a single target: in-scale at most steps, and next-resolution only at the few transition steps.
  • CINDI supervises every noisy state of a non-final scale at both the current and the next resolution, at every sampled time. This is only possible thanks to our shared INR bridge design.

Benefit from jointly supervising both resolutions through the shared implicit renderer:

In addition to the benefit we mentioned earlier (CINDI requires no special correlated noise at transition), another practical benefit of joint supervision is flexibility in timestep allocation. Thanks to joint supervision, every step of CINDI's non-final scale supports a scale transition, so user can flexibly allocate coarse/fine timestep using a single trained checkpoint. Evaluation can accommodate any combination of step allocations, including an early jump midway through the coarse timesteps. On the other hand, SSD trains its resolution-increasing pathway only at the transition timesteps specified by its training schedule. Although its sampler can skip denoising steps, the resolution transitions remain tied to that schedule. Therefore, placing transitions elsewhere means training a completely new model, as shown in its schedule ablation (Fig. 5), which trains one model per schedule.

Q2:Could the authors report an ImageNet-256 ablation without the next-scale loss, keeping the remaining training and sampling settings fixed?

Yes, Please see table below for the requested ablation:

ModelDenoising resolutionNSLEpochFID ↓
CINDI L/16128→256✓2002.77
CINDL L/16128→256✗2002.93

At 200 training epochs on ImageNet-256, with all other settings fixed, adding the next-scale loss reduces CINDI L/16’s FID from 2.93 to 2.77. This result is consistent with our ImageNet-64 finding and supports the contribution of next-scale supervision to generation quality.

Finally, we sincerely thank the reviewer for taking the time to read our rebuttal. We hope the additional information and experimental results address your concerns, and we would greatly appreciate your consideration of a higher score in light of this new evidence. Thank you!

References

[1] Soumik Mukhopadhyay, Prateksha Udhayanan, and Abhinav Shrivastava. Scale Space Diffusion. CVPR, 2026, pp. 35851–35860.

Reviewer_1c8EOpenReview ↗

after rebuttal

Thank you for the clarifications and the additional ImageNet-256 ablation. The comparison with SSD helps clarify the methodological distinction, and the new ablation supports the benefit of next-scale supervision at the main evaluation resolution. These responses address my main questions. Please incorporate the additional ablation and the clarified comparison into the revised manuscript.

Thank you so much for your thoughtful feedback. We are glad that our clarifications and the additional ablation have addressed your main questions. We have added both the additional ablation and the clarified comparison with SSD into the latest revised manuscript. We sincerely appreciate your time and constructive suggestions.

Reviewer_dot6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CINDI introduces a coarse-to-fine generation framework in pixel space, which progressively increases resolution during the sampling process, thereby shifting a significant portion of the initial computation to lower resolutions. A shared Transformer, in conjunction with an implicit renderer, predicts clean images at both the current and subsequent scales, with cross-resolution transitions learned through supervision from the next scale. Experiments on ImageNet-64 and ImageNet-256 demonstrate that, when matching training and sampling budgets, this approach achieves an FID comparable to that of JiT while improving inference throughput.

Critical strengths
  1. The efficiency gains reported in the experiments are of practical value. On ImageNet-256, CINDI-L/H achieves FID scores of 2.77/2.28—comparable to the 2.79/2.29 scores of JiT—while delivering BF16 throughputs that are 2.01× and 2.20× that of the baseline, respectively. By controlling for model scale, training duration, and the number of function evaluations, the experiments provide robust support for the efficiency claims made under the tested configurations. 2.Key modules are validated through experimentation. The model learns to perform both denoising at various scales and cross-scale prediction jointly, eliminating the need to train separate generators for different stages. Removing supervision for the next scale causes the FID for the 32→64 configuration to degrade from 2.19 to 2.75; furthermore, experiments regarding multi-stage setups and step allocation effectively illustrate the trade-off between generation quality and computational cost.
Critical weaknesses
  1. The independent contribution of the implicit renderer remains insufficiently clarified. While existing ablation studies validate the role of next-scale supervision, they fail to compare implicit neural representations (INRs) against simpler learned upsampling modules under identical conditions—specifically, using the same cross-scale supervision, resolution scheduling, and sampling budget. Consequently, it is unclear whether the performance gains stem primarily from the implicit representation itself or from the general cross-scale training and computational allocation strategy. This complicates the assessment of the core module's innovation.
  2. Furthermore, the scope for validating conclusions regarding quality and efficiency is limited. Current comparisons rely on a fixed sampling budget of 50 steps, lacking quality-throughput curves that would emerge from varying the total number of sampling steps (e.g., via a "jump-in-time" or similar mechanism). Quality evaluation is also primarily based on the Fréchet Inception Distance (FID), with a lack of complementary metrics to assess diversity or the quality of fine-grained details. Additionally, formal quantitative experiments are capped at a resolution of 256×256; although Figure 6 demonstrates direct rendering at 512×512, no quantitative assessment of quality or efficiency is provided for this resolution. Thus, while the existing results support the observed gains within the tested configurations, the broader applicability of the approach remains to be established.
Questions for authors
  1. Are there any ablation studies—where the cross-scale supervision and sampling settings are kept constant, but the INR component is replaced by standard interpolation or a learned upsampling module—that could help disentangle the respective contributions of cross-scale training versus implicit neural representations (INRs)?
  2. When the total number of sampling steps is reduced via "jump-in-time" (JiT) techniques, does CINDI maintain its throughput advantage while achieving a comparable FID score?
  3. Are there any complementary quality metrics or quantitative results at higher resolutions that support the claims regarding quality preservation and generation across different training resolutions? Please clarify the specific scope of capabilities demonstrated in Figure 6.
4 comments

We sincerely appreciate the reviewer for their time, thoughtful feedback, and questions. We respond to each point accordingly. New experimental results can be found in separate Supplymentary Material pdf.

Q1. Are there any ablation studies—where the cross-scale supervision and sampling settings are kept constant, but the INR component is replaced by standard interpolation or a learned upsampling module?

Yes, please see our transition operator ablation study below:

--Transition Operator Ablation Experiment--

To isolate the contribution of our learned INR bridge, we quantitavely and qualitatively compare the learned INR brdige against nearest-neighbor, bilinear, and bicubic upsampling transitions using the same trained CINDI B/4 epoch 200 checkpoints and sampling settings.

Quantitative results are below.

Qualitative results are in the New Supplymentary Material Section A

CINDI B/4 16→64

Hand-offBest CFG swept FID-50k
Learned INR3.96
Nearest5.39
Bicubic6.33
Bilinear6.53

CINDI B/4 32→64

Hand-offBest CFG swept FID-50k
Learned INR2.19
Nearest2.62
Bicubic2.89
Bilinear3.74

Quantitatively, CINDI with a learned INR consistently outperforms the other transition operators. For the 16→64 transition, it achieves an FID of 3.96, compared with 5.39 for the next-best operator, nearest-neighbor interpolation. For 32→64, it achieves an FID of 2.19, compared with 2.62, 2.89, and 3.74 for nearest-neighbor, bicubic, and bilinear interpolation, respectively.

Qualitatively, we also provide decisive visual comparison results for these transition operators. Please refer to Supplymentary material A, Figure 1. At the jump, CINDI's learned INR bridge produces clearer bird and branch boundaries than bilinear and bicubic interpolation, while avoiding the blockiness of nearest-neighbor interpolation. After continued sampling, Learned INR bridge's final output retains a complete branch structure and natural bird pose, whereas all fixed-operator outputs show visible breaks in the branch, making the bird appear to float in the air.

Beyond fixed interpolation operators, a learned upsampler, such as a CNN, could also serve as a bridge. However, conventional upsamplers are typically designed for fixed scale factors, requiring separate modules or additional mechanisms to accommodate different resolution transitions. Crucially, one of CINDI's key features is its support for arbitrary resolution transitions, including both integer (16→64, 32→64) and non-integer (32→48, 48→64) scaling, within a single shared module. Our learned INR bridge directly satisfies this requirement without introducing scale-specific upsampling modules, making it a natural choice over conventional learned upsamplers for our flexible coarse-to-fine framework.

Q2. When the total number of sampling steps is reduced via "jump-in-time" (JiT) techniques, does CINDI maintain its throughput advantage while achieving a comparable FID score?

For reduced number of sampling steps experiment, we reduce both CINDI B/4 and JiT B/4's sampling step to 40 step, and evaluated both their ep200 checkpoint. JiT B/4 vs CINDI B/4 after CFG sweep and CINDI's optimal step allocation. we reached FID 50k 2.27 vs 2.29, and BF16 inference throughput are at JiT's 23.8 images/s vs CINDI's 66.4 images/s. Therefore, at reduced 40 step sampling budget, CINDI maintains its throuput advantage while achieving a comparable FID score.

Q3a. Are there any complementary quality metrics or quantitative results at higher resolutions that support the claims regarding quality preservation and generation across different training resolutions?

Yes, to demonstrate that CINDI’s benefits extend to training and denoising at resolutions higher than 256×256, we have added two controlled comparative experiments:

  1. unconditional FFHQ512, with matched model scale, patch size (JiT L/32 vs CINDI L/32), training epochs(100), and effective batch size (1xH200GPU, bsz=128).
  2. class-conditional Imagenet512, with matched model scale, patch size(JiT L/32 vs CINDI L/32), train epochs(60), and effective batch size (8xH200GPU, bsz=1024).

JiT denoises at 512×512 throughout, while CINDI follows a 256→512 trajectory with an explicit 512×512 denoising stage. For simplicity, both models are evaluated using matched solver, precision(BF16), and inference batch size. Unconditional FFHQ512 eval without classifier-free guidance (CFG=1.0), conditional imagenet512 follows same CFG sweep as imagenet256.

MethodDatasetNumber of classesDenoising resolutionsBatch sizeTraining epochsFID-50k ↓Throughput (images/s) ↑Speedup vs. JiT ↑
JiT-L/32FFHQ512151212810038.275.821.00×
CINDI-L/32FFHQ5121256 → 51212810027.908.641.48×
JiT-L/32Imagenet5121,0005121,024606.985.871.00×
CINDI-L/32Imagenet5121,000256 → 5121,024605.248.871.51×

To our surprise, with 100 epoch training on unconditional FFHQ512 dataset, CINDI-L/32 not only perserves the quality of JiT L/32 but actually outperforms it with FID score 27.90 vs 38.27, while still show superior throughput (CINDI 8.64 images/s vs JiT 5.82 images/s (1.48×)).

For class-conditional Imagenet512 dataset, after 60 training epochs, CINDI L/32 continues to edge out JiT L/32 in quality, but the results are now more comparable (FID 5.24 vs 6.98). In the meantime, CINDI maintains an advantage in inference throughput (CINDI 8.87 images/s vs JiT 5.87 images/s (1.51×)).

Experimental results suggest that CINDI's c2f design and Learned INR provide better breath room in higher resolution modeling(512) than JiT counterpart. As we consistently see CINDI learns data distribution faster compare to JiT with the two controlled settings.

Note both model pairs can continue be trained and reach lower FID, these experiments at least prove the quality and throughput benefits of CINDI are applicable to higher-resolution denoising at 512 × 512.

On a side note, both CINDI and JiT can be trained at even higher resolutions or for more epochs. However, training Imagenet with our setting requires 16xH200 GPUs(8 for CINDI and 8 for JiT, same for 256/512/1024 resolutions), which means long queueing time for such resources to be allocated to us. Considering we only have less than 2 weeks for the rebuttal discussion, we hope our added experimental results are sufficent for addressing the reviewer’s concerns about higher-resolution training and denoising.

Q3b. Please clarify the specific scope of capabilities demonstrated in Figure 6.

We provide Figure 6 to demonstrate a nice-to-have capability of CINDI. Although CINDI’s core contributions are its coarse-to-fine design and implicit bridge for efficient sampling, its multiscale predictive capability also allows it to produce outputs at arbitrary resolutions. We certainly would not claim that rendering at 512×512 using a checkpoint trained at 256×256 yields better quality than using a checkpoint trained at 512×512, as this is not among the contributions claimed in the introduction. However, that said, for users seeking an end-to-end diffusion model that supports arbitrary-resolution output without requiring separate VAE, CINDI could be a good option.

Finally, we sincerely thank the reviewer for taking the time to read our rebuttal. We hope the additional information and experimental results address your concerns, and we would greatly appreciate your consideration of a higher score in light of this new evidence. Thank you!

Reviewer_dot6OpenReview ↗

Response to Authors

Thank you for the clarification. The response resolves my concerns.

Thank you very much for your kind reply. We sincerely appreciate the time and effort you have devoted to providing a thoughtful review and constructive suggestions to improve our paper. We are pleased that our clarifications have addressed your concerns. Thank you again for your valuable feedback and support!

Reviewer_RxK4OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper introduces CINDI, a coarse-to-fine pixel-space diffusion model that reduces sampling cost by performing early denoising steps at lower resolutions. A shared Transformer and coordinate-based implicit neural renderer support both in-scale denoising and next-scale clean-image prediction. The method combines next-scale bridge supervision with a sampler that re-noises the predicted image at scale transitions. Experiments on class-conditional ImageNet at 64×64 and 256×256 show comparable FID to JiT with higher inference throughput under matched model scales, training epochs, and sampling budgets.

Critical strengths
  1. The motivation is clear and intuitive: early denoising mainly establishes coarse structure, so allocating less spatial computation to these steps is a reasonable way to improve sampling efficiency.

  2. The shared implicit denoiser provides a coherent mechanism for cross-scale prediction without requiring a pretrained autoencoder or separately trained cascaded models.

Critical weaknesses
  1. The related-work discussion needs a more systematic organization and sharper positioning. Section 2.4 omits relevant approaches such as Pyramid Flow (arXiv:2410.05954) and Pyramidal Patchification Flow (arXiv:2506.23543), which also allocate computation along the denoising trajectory. Combining diffusion with INRs is not itself new, as illustrated by Kim and Kim (2024, arXiv:2403.10255), which the paper already cites. Although CINDI distinguishes its in-trajectory renderer from prior latent-space decoders, a more explicit comparison of the mechanisms and their benefits is needed. The system-level PixelFlow comparison in Table 3 does not isolate the advantage of learned implicit bridging under matched conditions.
  2. The method presentation is harder to follow than necessary. Sections 3–4 introduce substantial notation for an otherwise intuitive workflow; symbols used only locally could be reduced. Figure 2 would benefit from a clearer visual separation of in-scale denoising, next-scale prediction, and re-noising.
  3. The experimental scope remains limited to class-conditional ImageNet at 64×64 and 256×256. The 512×512 example in Appendix D is a qualitative rendering comparison against nearest-neighbor upsampling, rather than a quantitative high-resolution generation benchmark. Performance and efficiency at higher resolutions and on text-to-image generation therefore remain unestablished.
Questions for authors
  1. Could the authors systematically compare CINDI with Pyramid Flow, Pyramidal Patchification Flow, and prior diffusion–INR approaches, clarifying the respective roles of resolution scheduling, patch-size scheduling, and learned implicit bridging? Where settings permit, could matched comparisons or controlled transition-operator ablations isolate the benefit of the proposed bridge?
  2. Could the authors simplify the notation and revise Figure 2 to make the three operations—within-scale denoising, cross-scale clean prediction, and re-noising—easier to follow?
  3. Could the authors provide quantitative quality and efficiency evaluations at 512×512 or higher, distinguishing direct rendering from denoising at the higher resolution? Is there evidence that the benefits extend to text-to-image generation?
7 comments
AuthorsOpenReview ↗

Part1: Regarding related work positioning and isolation of benefit of Learned INR bridge

We sincerely thank the reviewer for providing quality review and questions. We will address each concern accordingly. New experiment results can be found in separate Supplymentary Material pdf, paper content and figure2 are revised based on reviewer's suggestion.

1a. Regarding related work positioning concern:

Q:Could the authors systematically compare CINDI with Pyramid Flow, Pyramidal Patchification Flow, and prior diffusion–INR approaches, clarifying the respective roles of resolution scheduling, patch-size scheduling, and learned implicit bridging?

Yes, to better address the related-work positioning concern, we systematically compare CINDI with the relevant methods the reviewer mentioned and categorize them in the table below by where generation runs and how computation is reduced along the trajectory; (all cited in the revised manuscript)

MethodSpaceTaskSampling strategyTransition operator / INR roleC2F computation-saving capable
Pyramid Flow [1]Latent (2 stage)* Video * generationDynamic state resolution; fixed patch sizeNearest-neighbor upsampling + corrective re-noisingYes
Pyramidal Patchification Flow (PPFlow) [2]Latent (2 stage)Image generationDynamic patch size; fixed state resolutionN/AYes
Kim & Kim [3]Latent (2 stage)Image generationNo scheduling on resolution or patchsizeN/A; INR for post hoc arbitrary output resolutionNo
PixelFlow [4]Pixel (end-to-end)Image generationDynamic state resolution; fixed patch size (fine-grained, degrade on coarse-grained)Nearest-neighbor upsampling + corrective re-noisingYes
CINDIPixel (end-to-end)Image generationDynamic state resolution; fixed patch size (coarse-grained)Learned INR clean prediction + re-noisingYes

All methods in the table except Kim & Kim save computation by processing fewer tokens in early sampling stages and full tokens in later stages. Pyramid Flow, PixelFlow, and CINDI achieve this by progressively increasing state resolution at fixed patch size, whereas PPFlow decreases patch size while keeping the latent-state resolution fixed. Kim & Kim instead use an implicit decoder for arbitrary-scale output, but exihbit no coarse-to-fine efficiency capability. Pixel Flow is the closest c2f method compared to CINDI, however Pixel Flow only works well with fine grained patchification(4x4), with significant degradation at coarse grained patch(8x8) reported in Table 2 of their paper, imagenet256 FID degrade from 2.81 to 4.65. CINDI instead still perform well at even coarser grained patchification(16x16) and with fewer training epoch, imagenet256 FID at 2.28.

One of CINDI’s core distinction is its next-scale-supervised INR bridge, which directly predicts the finer-resolution clean image used to initialize the next sampling stage. And we agree with the reviewer, that our transition operator's benefit need to be isolated and ablated as well.

1b. Isolation of Benefit of Learned INR bridge:

Q:Where settings permit, could matched comparisons or controlled transition-operator ablations isolate the benefit of the proposed bridge?

Yes, please see requested ablation below:

Transition Operator Ablation Experiment

To isolate the contribution of our learned INR bridge, we quantitavely and qualitatively compare the learned INR brdige against nearest-neighbor, bilinear, and bicubic upsampling transitions using the same trained CINDI B/4 checkpoints and sampling settings (so only transition operator is swapped).

Quantitative results are below.

Qualitative Result in New Supplymentary Material, Section A

CINDI B/4 16→64

Hand-offBest CFG swept FID-50k
Learned INR3.96
Nearest5.39
Bicubic6.33
Bilinear6.53

CINDI B/4 32→64

Hand-offBest CFG swept FID-50k
Learned INR2.19
Nearest2.62
Bicubic2.89
Bilinear3.74

Quantitatively, CINDI with a learned INR consistently outperforms the other transition operators. For the 16→64 transition, it achieves an FID of 3.96, compared with 5.39 for the next-best operator, nearest-neighbor interpolation. For 32→64, it achieves an FID of 2.19, compared with 2.62, 2.89, and 3.74 for nearest-neighbor, bicubic, and bilinear interpolation, respectively.

Qualitatively, we also provide decisive visual comparison results for these transition operators. Please refer to Supplymentary material Section A, Figure 1. At the jump, CINDI's learned INR bridge produces clearer bird and branch boundaries than bilinear and bicubic interpolation, while avoiding the blockiness of nearest-neighbor interpolation. After continued sampling, Learned INR bridge's final output retains a complete branch structure and natural bird pose, whereas all fixed-operator outputs show visible breaks in the branch, making the bird appear to float in the air.

AuthorsOpenReview ↗

Part2: Regarding limited dataset scope and ability for higher-resolution training and generation.

2.Regarding limited dataset scope and ability for higher-resolution training and generation.

Q: Could the authors provide quantitative quality and efficiency evaluations at 512×512 or higher, distinguishing direct rendering from denoising at the higher resolution?

Yes, to provide quantitative evaluations of CINDI’s generation quality and efficiency when denoising at 512×512 (rather than denoising at 256×256, then rendering at 512×512, as in Appendix D), we have added two controlled comparison experiments:

  1. unconditional FFHQ512, with matched model scale, patch size (JiT L/32 vs CINDI L/32), training epochs(100), and effective batch size (1xH200GPU, bsz=128).
  2. class-conditional Imagenet512, with matched model scale, patch size(JiT L/32 vs CINDI L/32), train epochs(60), and effective batch size (8xH200GPU, bsz=1024).

JiT denoises at 512×512 throughout, while CINDI follows a 256→512 trajectory with an explicit 512×512 denoising stage. For simplicity, both models are evaluated using matched solver, precision(BF16), and inference batch size. unconditional FFHQ512 eval without classifier-free guidance (CFG=1.0), conditional imagenet512 follows same CFG sweep as imagenet256.

MethodDatasetNumber of classesDenoising resolutionsBatch sizeTraining epochsFID-50k ↓Throughput (images/s) ↑Speedup vs. JiT ↑
JiT-L/32FFHQ512151212810038.275.821.00×
CINDI-L/32FFHQ5121256 → 51212810027.908.641.48×
JiT-L/32Imagenet5121,0005121,024606.985.871.00×
CINDI-L/32Imagenet5121,000256 → 5121,024605.248.871.51×

To our surprise, with 100 epoch training on unconditional FFHQ512 dataset, CINDI-L/32 not only perserves the quality of JiT L/32 but actually outperforms it with FID score 27.90 vs 38.27, while still show superior throughput (CINDI 8.64 images/s vs JiT 5.82 images/s (1.48×)).

For class-conditional Imagenet512 dataset, after 60 training epochs, CINDI L/32 continues to edge out JiT L/32 in quality, but the results are now more comparable (FID 5.24 vs 6.98). In the meantime, CINDI maintains an advantage in inference throughput (CINDI 8.87 images/s vs JiT 5.87 images/s (1.51×)).

Experimental results suggest that CINDI's c2f design and Learned INR provide better breath room in higher resolution modeling(512) than JiT counterpart. As we consistently see CINDI learns data distribution faster compare to JiT with the two controlled settings.

Note both model pairs can continue be trained and reach lower FID, these experiments at least prove the quality and throughput benefits of CINDI extend beyond class-conditional ImageNet and also apply to higher-resolution denoising at 512 × 512.

On a side note, both CINDI and JiT can be trained at even higher resolutions or for more epochs. However, training Imagenet with our setting requires 16xH200 GPUs(8 for CINDI and 8 for JiT, same for 256/512/1024 resolution), which means long queueing time for such resources to be allocated to us. Considering we only have less than 2 weeks for rebuttal discussion, we hope our added experimental results are sufficent to address the reviewer’s concerns about higher-resolution denoising and dataset diversity.

Q: Is there evidence that the benefits extend to text-to-image generation?

Regarding text-to-image generation, we appreciate the reviewer’s interest in this extension; CINDI’s cross-scale interface and next-scale supervision is conditioning irrelavant, class embedding conditioning can be replaced with text embedding. However,that said, our paper never claimed any contribution related to text-to-image generation, and we focused on class-conditional image generation (same as our baseline JiT paper), validating this extension is beyond the scope of the present study. Thank you for your understanding.

Finally, we sincerely thank the reviewer for taking the time to read our rebuttal. We hope the additional information and experimental results address your concerns, and we would greatly appreciate your consideration of a higher score in light of this new evidence. Thank you!

References

[1] Jin et al. Pyramidal Flow Matching for Efficient Video Generative Modeling. ICLR 2025.

[2] Li et al. Pyramidal Patchification Flow for Visual Generation. ICLR 2026.

[3] Kim and Kim. Arbitrary-Scale Image Generation and Upsampling using Latent Diffusion Model and Implicit Neural Decoder. CVPR 2024.

[4] Chen et al. PixelFlow: Pixel-Space Generative Models with Flow. arXiv:2504.07963, 2025.

Dear Reviewer RxK4,

Thank you again for your constructive questions and suggestions, as well as your positive remarks on our work.

We would like to follow up on the responses and revisions we have provided to address your concerns:

  1. We have added a systematic comparison with the related methods you mentioned to clarify CINDI’s distinct positioning and differences from prior work.
  2. We have added transition-operator ablations to isolate the contribution of CINDI’s learned INR bridge.
  3. We have added quantitative evaluations of two models trained at 512×512, providing evidence that CINDI’s quality and throughput benefits extend to modeling at resolutions above 256×256.
  4. We have also further simplified the mathematical notation in the implicit-denoiser writing and revised Figure 2 to provide clearer visual separation between the different operations in the latest manuscript revision(Nov 16). And we will continue to refine the writing and would welcome any further suggestions for improving clarity.

As the discussion period progresses, we would greatly appreciate it if you could let us know whether our responses have adequately addressed your concerns. We remain available to answer any remaining questions you may have.

Thank you for your time and thoughtful consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.