acceptodds
Desk rejected at ICLR 2027

DC-Gen 2.0: Efficient and High-Fidelity Image Generation in Deeply Compressed Latent Spaces

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing diffusion models excel at generating high-quality images, but scaling them to higher resolutions remains computationally expensive. Deep-compression autoencoders reduce latent token counts, yet maintaining reconstruction fidelity and synthesizing fine details in the compressed latent space remain challenging. To address these challenges, we present DC-Gen 2.0, built on our DC-AE 2.0 family. For DC-AE 2.0-Structured, we propose Pyramid-Aligned Decoder Training to facilitate downstream generation while maintaining high reconstruction quality. Building on the structured latent space introduced by DC-AE 1.5, it aligns channel budgets with hierarchical decoder paths and Gaussian-pyramid targets, encouraging coarse-to-fine reconstruction across decoder stages. To improve generation quality under high compression, we introduce Deep-Compression Detail Enhancement, which combines targeted data curation with generation post-training through diffusion adaptation and reinforcement learning. Curated text and face data support fine-detail preservation during autoencoder training and synthesis during diffusion post-training. Experiments show that DC-Gen 2.0 improves reconstruction quality and enables efficient, high-quality image generation under high compression. For FLUX.2-dev, DC-Gen 2.0 f32 achieves 3.39x speedup at 1024 and DC-Gen 2.0 f64 achieves 53.91x speedup at 4096.

Primary area:
generative models
Keywords:
efficient high-resolution diffusion model, post-training model acceleration, deep compression autoencoder

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

The paper proposes DC-GEN 2.0. The paper trains DC-AE 2.0 as a family of deep-compression autoencoders on a targeted dataset containing images with small text and small faces, and proposes pyramid-aligned decoder training (PADT) to separate latent channels for low-to-high-frequency reconstruction. The paper adapts FLUX1 and FLUX2 to the DC-AE 2.0 latent space to compare quality and speedup.

Critical strengths

Deep-compression AEs are an important and practical problem, and the paper proposes new improvements on top of a prior series of works.

Critical weaknesses

Novelty is limited.

  1. The structured latent space was previously proposed in DC-AE 1.5, which masks a random prefix of latent channels to reconstruct full-resolution images. The only essential difference in PADT is that it instead reconstructs images at different pyramid resolutions.
  2. Another difference from DC-AE 1.0 and 1.5 is the inclusion of targeted data, namely image datasets containing small text and small faces. This is very effective as shown in the paper but constitutes limited novelty.

The experimental results do not support the claims.

  1. In Table 2, DC-AE 2.0 seems to outperform DC-AE 2.0-structured in reconstruction.
  2. In Table 4 (the main results), DC-AE 2.0-structured does not clearly outperform DC-AE 2.0 (non-structured) for FLUX2-dev. DC-AE 2.0 (non-structured) even clearly outperforms DC-AE 2.0-structured for FLUX1-Krea. It is also highly questionable that the paper does not include DC-AE 2.0 (non-structured) as a comparison for FLUX2 at 4096px.
  3. The results do not support the claim that PADT contributes to either reconstruction or downstream diffusion, which undermines the main contribution of the paper.

The measurement of speedup is not convincing.

  1. In Table 4, speedup and throughput are measured only for the DiT, without including the VAE component. This is concerning because DC-AE can be more computationally expensive than the original FLUX-VAE and should therefore be included in the measurement. Alternatively, a separate comparison of the speed of DC-AE and FLUX-VAE should be provided.

The training comparison is not fair.

  1. In Table 4, the compared FLUX model with the original AE is not fine-tuned at all, whereas the FLUX models with DC-AE 2.0 are fine-tuned on a custom dataset. This is no longer a fair comparison because the gain may result from the dataset rather than the change in latent space. The authors should at least fine-tune the FLUX model using the matched dataset for the same number of epochs.
Questions for authors

Please address the concerns listed in critical weaknesses section.

5 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/2)

Thank you for recognizing the practical importance of deep-compression autoencoders and our efforts to improve them. Below, we first present reconstruction and generation evidence for PADT, then clarify our contributions. We also report end-to-end generation speed and compare against FLUX.2-dev fine-tuned with its original autoencoder under matched training settings.

1 Experimental evidence for PADT

Structured latent spaces help diffusion models learn high-dimensional latents under deep compression. However, the original DC-AE 1.5 structured-training recipe leads to a loss of reconstruction quality, which can limit the fidelity of downstream generation. We propose PADT to address this issue, aiming to achieve reconstruction quality comparable to standard DC-AE 2.0 while retaining structured latents. Table 2 evaluates this objective through comparisons with both the original DC-AE 1.5 recipe and standard DC-AE 2.0. To further assess the downstream benefits, we have expanded the standard-versus-Structured comparisons in Table 4 and added class-conditional ImageNet generation experiments.

Reconstruction quality

As intended, DC-AE 2.0-Structured achieves reconstruction quality comparable to standard DC-AE 2.0 at f32c128 while retaining structured latents, consistent with our statement in Sec. 4.2. Across the three benchmarks, the absolute PSNR differences are below 0.022 dB. Compared with the original DC-AE 1.5 structured-training recipe under identical initialization, training data, and training settings, PADT improves all reported reconstruction metrics by a large margin (Table 2). On ImageNet, FFHQ, and OmniDoc, respectively, PSNR increases by 0.6271/0.6863/0.3679 dB.

ImageNet

AutoencoderPSNR ↑SSIM ↑LPIPS ↓rFID ↓
DC-AE 2.030.12920.84780.07060.2433
DC-AE 2.0-Structured30.13230.84510.07350.2390
DC-AE 1.529.50520.82340.07960.2502

FFHQ

AutoencoderPSNR ↑SSIM ↑LPIPS ↓rFID ↓
DC-AE 2.036.12770.91770.06560.2276
DC-AE 2.0-Structured36.14900.91610.06760.2127
DC-AE 1.535.46270.90980.07110.2396

OmniDoc

AutoencoderPSNR ↑SSIM ↑LPIPS ↓rFID ↓NED ↑
DC-AE 2.026.66150.92860.03001.11700.9269
DC-AE 2.0-Structured26.64990.92800.03041.19260.9287
DC-AE 1.526.28200.91310.03231.20760.9252

Generation quality

Our focus is fine-grained detail generation under deep compression. GenEval and DPG-Bench primarily evaluate prompt alignment, whereas LongText-Bench directly evaluates text rendering. LongText-Bench is therefore a primary benchmark for assessing the benefits of Structured latents for fine-grained detail generation.

To provide more complete standard-versus-Structured comparisons, we expand Table 4 with standard DC-AE 2.0 f32c128 results for FLUX.1-Krea at 1024 × 1024 and standard DC-AE 2.0 f64c128 results for FLUX.2-dev at 4096 × 4096. The paired results below show consistent LongText-Bench gains across the evaluated backbones, AE settings, and resolutions, supporting PADT's effectiveness for fine-grained generation under deep compression.

FLUX.2-dev, 1024 × 1024

AutoencoderAE settingLongText EN ↑LongText ZH ↑
DC-AE 2.0f32c1280.96750.9338
DC-AE 2.0-Structuredf32c1280.96920.9467
DC-AE 2.0f64c1280.92220.9187
DC-AE 2.0-Structuredf64c1280.92490.9233

FLUX.2-dev, 4096 × 4096, f64c128

AutoencoderAE settingLongText EN ↑LongText ZH ↑
DC-AE 2.0f64c1280.95280.9321
DC-AE 2.0-Structuredf64c1280.96010.9440

FLUX.1-Krea, 1024 × 1024

AutoencoderAE settingLongText EN ↑
DC-AE 2.0f32c320.8867
DC-AE 2.0f32c1280.8607
DC-AE 2.0-Structuredf32c1280.8658

The original FLUX.1-Krea comparison used standard f32c32 and Structured f32c128, which differ in channel count. The higher score at f32c32 is consistent with lower-dimensional latents being easier for the smaller FLUX.1 backbone to learn. As shown in the table above, we have added standard f32c128 to compare the two variants at the same channel count. Structured still achieves a higher LongText-Bench score (0.8658 vs. 0.8607), consistent with the improvements observed in the other matched comparisons.

Beyond text-to-image generation, we additionally evaluate class-conditional generation on ImageNet. At the same f32c128 configuration, DC-AE 2.0-Structured reduces FID from 39.92 to 25.33 compared with standard DC-AE 2.0, providing complementary evidence of its generation benefits.

AutoencoderAE settingFID ↓
DC-AE 2.0f32c12839.92
DC-AE 2.0-Structuredf32c12825.33
AuthorsOpenReview ↗

Official Comment by Authors (2/2)

2 Novelty and contributions

PADT design

PADT addresses a practical limitation of the DC-AE 1.5 structured-training recipe: supervising short channel prefixes against full-resolution images degrades reconstruction quality. We propose a new training design that jointly aligns channel budgets, decoder stages, and target scales, matching reconstruction demands to the available channel capacity and decoder depth. The controlled comparison above demonstrates its benefit for recovering reconstruction quality while retaining structured latents for better generation performance. This improved reconstruction fidelity raises the potential quality ceiling for downstream generation.

Engineering and community contributions

Targeted data curation and post-training are engineering contributions of this work. Applying these existing techniques effectively in deeply compressed latent spaces requires nontrivial adaptation and tuning of the training stages, data composition, and balance between training targets. Tables 3 and 5 quantify the benefits of targeted data for reconstruction and generation: OmniDoc PSNR increases from 24.1228 to 26.6615, and Chinese LongText-Bench increases from 0.0745 to 0.9290. Table 9 further shows the benefit of our post-training pipeline for text and face generation. We will release our pipeline and the resulting models, providing the community with reusable resources for efficient high-resolution generation and further research.

3 End-to-end generation speed

Thank you for pointing this out. We agree that DiT-only measurements do not fully capture generation efficiency. We have replaced them with end-to-end measurements, including text encoding and AE decoding, and updated Table 4, Section 4.3, the abstract, and the speedup charts in Figure 1 accordingly. All measurements use a single GB300 GPU, BF16 precision, and 20 sampling steps. Including these costs, FLUX.2-dev with DC-AE 2.0 achieves speedups of 3.33× and 6.60× at 1024 × 1024 with f32 and f64, respectively, and 48.5× at 4096 × 4096 with f64.

4 Training comparison with native FLUX.2-dev

Thank you for highlighting this confounding factor. Table 4 compares our complete system with the original FLUX.2-dev, but this comparison alone does not isolate the effects of targeted-data fine-tuning from those of autoencoder adaptation. To address this concern, we additionally fine-tuned FLUX.2-dev with its original autoencoder on the same targeted dataset using the same training settings and number of iterations. Targeted-data fine-tuning substantially improves the original model’s text-rendering performance, particularly in Chinese. Our adapted model nevertheless achieves higher scores on all four metrics.

AutoencoderAE settingTrainingLongText EN ↑LongText ZH ↑TextPecker EN ↑TextPecker ZH ↑
Original AEf8c32Original0.95060.752697.2373.94
Original AEf8c32Fine-tuned0.95780.923797.6785.78
DC-AE 2.0-Structuredf32c128Adapted0.96920.946798.0387.81

I appreciate the authors' response. However, my overall assessment of the paper remains borderline. My primary concerns are as follows:

  1. PADT yields only marginal improvements in generation quality. This raises the question of whether the significance of the contribution justifies the additional engineering complexity. For example, training PADT VAE also incurs additional computational costs due to the pyramidal alignment loss, a factor that is not considered at all in the paper. If the method demonstrated substantial improvements, it would provide a stronger case for acceptance. Unfortunately, the reported gains are not sufficiently convincing.

  2. Dataset curation appears to play a much larger role than PADT in improving generation quality. I appreciate the authors providing additional results from fine-tuning FLUX2 on the same curated dataset. However, these results further suggest that dataset curation is a much more significant factor in improving model quality than the proposed PADT technique.

I will reserve my final rating for now and remain open to further discussion with the other reviewers.

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

This paper proposes DC-Gen 2.0 with two major improvements to the Deep Compression Autoencoder (DC-AE). First, on the VAE (or DC-AE) design side, it proposes the DC-AE 2.0 series, with Pyramid-Aligned Decoder Training (PADT) as the main improvement. Specifically, intermediate decoder features at different resolutions are aligned, using different numbers of channels, with the corresponding levels of the Gaussian pyramid decomposition of the original image. Second, it proposes Deep-Compression Detail Enhancement (DCDE), a comprehensive data and post-training pipeline. The training data are expanded to include faces as well as Chinese and English text, while the post-training pipeline incorporates Diffusion Adaptation and RL. Experimental results demonstrate that DC-Gen 2.0 achieves a better trade-off between generation speed and quality. DC-AE 2.0 itself also achieves better reconstruction quality at the same compression ratio. Overall, this work presents effective improvements to both DC-AE and diffusion training.

Critical strengths
  1. PADT imposes a structured constraint on the distribution of information across latent channels. Tables 1 and 2 show that this constraint does not significantly improve the compression and reconstruction efficiency of DC-AE 2.0 itself. However, Table 4 indicates that it provides some benefits for text generation in downstream diffusion models.
  2. DCDE develops a comprehensive data and post-training pipeline. Multiple experiments demonstrate that this pipeline substantially improves the capabilities of both DC-AE 2.0 itself and downstream diffusion models. It would be great if this pipeline could be open-sourced.
Critical weaknesses

W1. I consider PADT to be the core contribution of this work. It attempts to introduce a better training strategy for learning structured latents in DC-AE. However, the associated effectiveness analysis, ablation studies, and main experimental results are not yet sufficiently comprehensive.

  • Regarding the analysis of its effectiveness: Table 4 demonstrates the benefits of PADT for text generation, and Fig. 8 in the appendix illustrates its impact on the frequency distribution. However, why is this particular distribution of information across channels beneficial for diffusion training? The paper currently reuses the discussion of the Structured Latent Space from DC-AE 1.5 without providing a deeper analysis or explanation of why PADT works. As a result, PADT appears more like an empirical exploration than a generalizable finding, which limits the overall contribution of this component.

  • Regarding ablation studies: Several aspects of PADT worth ablation studies but are not investigated. These include the blur operation in the Gaussian pyramid, the contribution of LPIPS in the loss function after blurring, the role of aligning features at each pyramid level, and the number of channels aligned at each level. Without these ablations, the overall contribution of PADT remains unclear.

  • Regarding experiments: First, the comparison between DC-AE 2.0 and DC-AE 2.0-Structured is insufficient, particularly in Tables 1 and 4. Some settings report results only for DC-AE 2.0, while others report results only for DC-AE 2.0-Structured. Second, both DC-AE and DC-AE 1.5 report their performance on general domain benchmark like ImageNet, whereas neither DC-AE 2.0 nor DC-Gen 2.0 provides corresponding results. This makes it difficult to fully assess their actual performance levels.

W2. DCDE does not appear to be a sufficiently strong contribution on its own. It seems more like an engineering pipeline that combines existing techniques, including different datasets, post-training strategies, DMD, RL, and reward models. Given the aforementioned concerns regarding PADT, the overall contribution of the work requires further evaluation.

Minors

  • The caption of Table 2 remains unclear. A clearer explanation is needed for how DC-AE 1.5, DC-AE 2.0, and DC-AE 2.0-Structured are trained.
  • At first glance, the bar charts in Fig. 1 appear quite different. It would be helpful to emphasize that these charts are generated by DC-Gen 2.0 rather than manually designed. If possible, the style and visual details of the bar charts should be made more consistent. In addition, it would be better if the speed comparison at 4096 × 4096 resolution also include the f32 model.
  • In Table 4, the best results for Prompt Alignment are not highlighted in bold, making the comparisons less clear.
Questions for authors

I slightly lean toward a positive initial evaluation because PADT presents an interesting direction for improving DC-AE and demonstrates some degree of effectiveness. DCDE, as a complete engineering pipeline, also represents a substantial amount of work. However, I sincerely hope to address the following issues before further evaluation.

Q1. Could you provide a deeper analysis of why PADT benefits diffusions? More specifically, why is the channel-wise frequency distribution in Fig. 8(a) preferable to that in Fig. 8(b)? Why does PADT produce a better structured latent than DC-AE 1.5?

Q2. Could you provide ablation analysis on the implementation of PADT? For instance, the Gaussian pyramid involves both blurring and downsampling. What are the respective roles of these two operations in extracting low-frequency components? Furthermore, does LPIPS in Lrec\mathcal{L}_{\mathrm{rec}}, as a perceptual loss, provide sufficient supervision for low-frequency signals after repeated blurring and downsampling? How do the supervision applied at each resolution level and the number of supervised channels at each level affect the final results?

Given the limited response time, the following questions are relatively lower priority.

Q3. What is the compression limit of DC-AE 2.0, as earlier versions of DC-AE explored f128. How does DC-AE 2.0 perform at f128?

Q4. Is it possible to complete the comparisons between DC-AE 2.0 and DC-AE 2.0-Structured across different settings (f32c32, f32c128, f32c256, f64c128) in Tables 1 and 4? In addition, the generation results of DC-Gen 2.0 on general benchmarks, such as ImageNet (256x256 or 512x512).

Q5. Since DCDE primarily adds training data covering faces and Chinese and English text, how well does it generalize to other out-of-domain details, such as text in other languages?

6 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/4)

Thank you for recognizing PADT as an interesting direction for improving DC-AE and for acknowledging the substantial effort behind DCDE. Following your suggestion, we will release our models and post-training pipeline to support reproducibility and further research. Below, we address your concerns through a more detailed analysis of PADT, new ablations on implementation choices, and additional evaluations.

W1 and Q1: Why PADT can benefit diffusion

Advantage over DC-AE 1.5: Higher reconstruction fidelity while retaining structured latents

PADT builds on the structured latent space of DC-AE 1.5 and improves reconstruction fidelity while preserving its structured latent space. Under identical initialization, data, and training settings, Table 2 shows that PADT improves all reported reconstruction metrics over the DC-AE 1.5 recipe and achieves quality comparable to standard DC-AE 2.0. Figure 8(a) shows the retained coarse-to-fine channel organization. Better reconstruction raises the potential fidelity ceiling for downstream generation.

A frequency perspective on channel-masked diffusion training

DC-AE 1.5 [1] uses channel-masked diffusion training to emphasize learning object structure. We further explain why this can help from a frequency perspective. We have added this analysis to Appendix A.4 of the revised paper.

Prior work supports the benefit of low-frequency-biased latents for diffusion learning. Skorokhodov et al. [2] show that scale-equivariance regularization reshapes latent frequency profiles and improves DiT convergence and generation quality. SSVAE [3] further analyzes the link between local correlations and low-frequency energy using the Wiener–Khinchin theorem, and empirically demonstrates improved video generation through local correlation regularization.

In DC-AE 2.0-Structured (Fig. 8(a)), early channels have larger relative DCT magnitudes at low frequencies, while later channels have larger relative magnitudes at high frequencies. In standard DC-AE 2.0 (Fig. 8(b)), the channel spectra overlap without clear frequency ordering.

This channel organization provides a frequency-based interpretation of our channel-masked diffusion training. Since prefix lengths are sampled throughout training, short-prefix objectives emphasize low-frequency structure, while longer-prefix objectives additionally require modeling fine details. We hypothesize that jointly learning these tasks helps the shared denoiser acquire structural features from the simpler short-prefix objectives and reuse them when learning the richer long-prefix objectives.

[1] Junyu Chen et al. DC-AE 1.5: Accelerating Diffusion Model Convergence with Structured Latent Space. ICCV, 2025.

[2] Ivan Skorokhodov et al. Improving the Diffusability of Autoencoders. ICML, 2025.

[3] Shizhan Liu et al. Delving into Latent Spectral Biasing of Video VAEs for Superior Diffusability. arXiv:2512.05394, 2025.

AuthorsOpenReview ↗

Official Comment by Authors (2/4)

W1 and Q2: PADT implementation ablations

Constructing low-frequency supervision

To examine the role of Gaussian blur, we replace Gaussian-pyramid target construction with bilinear downsampling without explicit Gaussian smoothing (antialias=False), keeping the target resolutions and all other PADT settings unchanged. We conduct this ablation on DC-AE 2.0-Structured f32c128 using the Table 2 baseline and report the results in Appendix E.

ImageNet 256

Target constructionPSNR ↑SSIM ↑LPIPS ↓rFID ↓
Gaussian pyramid30.13230.84510.07350.2390
Bilinear only29.62450.84230.07710.2622

FFHQ 1024

Target constructionPSNR ↑SSIM ↑LPIPS ↓rFID ↓
Gaussian pyramid36.14900.91610.06760.2127
Bilinear only36.01320.91010.07000.2291

OmniDoc 256

Target constructionPSNR ↑SSIM ↑LPIPS ↓rFID ↓NED ↑
Gaussian pyramid26.64990.92800.03041.19260.9287
Bilinear only26.04810.91910.03261.26750.9129

Removing Gaussian blur reduces PSNR by 0.5078, 0.1358, and 0.6018 dB on ImageNet, FFHQ, and OmniDoc, respectively; SSIM, LPIPS, and rFID also worsen on all three benchmarks. On OmniDoc, NED decreases from 0.9287 to 0.9129, a drop of 0.0158.

These consistent degradations support explicit low-pass filtering before downsampling. Blurring suppresses high-frequency content and reduces aliasing, while downsampling sets the target spatial resolution. The results are consistent with our design motivation: coarse-scale supervision should match the capacity of shorter channel prefixes and earlier decoder stages. PADT supports different target-construction methods, including Gaussian pyramids and antialiased resizing; the Gaussian pyramid provides more effective supervision than the bilinear-only alternative tested here.

LPIPS at coarse scales

To ablate LPIPS at coarse scales, we remove the auxiliary LPIPS loss from the intermediate decoder heads and retain L1 supervision, keeping the Gaussian pyramid and original full-resolution objective unchanged. We conduct this ablation on DC-AE 2.0-Structured f32c128 using the Table 2 baseline and report the results in Appendix E. The differences are small across all benchmarks, indicating that auxiliary LPIPS provides a modest complementary benefit.

ImageNet 256

Auxiliary lossPSNR ↑SSIM ↑LPIPS ↓rFID ↓
L1 + LPIPS30.13230.84510.07350.2390
L1 only30.11390.84510.07430.2503

FFHQ 1024

Auxiliary lossPSNR ↑SSIM ↑LPIPS ↓rFID ↓
L1 + LPIPS36.14900.91610.06760.2127
L1 only36.11460.91590.06820.2203

OmniDoc 256

Auxiliary lossPSNR ↑SSIM ↑LPIPS ↓rFID ↓NED ↑
L1 + LPIPS26.64990.92800.03041.19260.9287
L1 only26.59800.92760.03071.23760.9272

Supervision levels and channel budgets

Our downstream channel-masked diffusion training follows the DC-AE 1.5 recipe, which uses channel budgets of [32, 64, 96, 128] for c=128. We therefore use the same prefixes for PADT supervision, pairing them with decoder outputs and targets at H/8, H/4, H/2, and H, respectively.

To examine whether supervising additional prefixes improves reconstruction, we expand the channel budgets from [32, 64, 96, 128] to [16, 32, 48, 64, 96, 128]. Each of the six channel prefixes is paired with a distinct target resolution, maintaining a one-to-one correspondence between prefixes and supervision resolutions. We compare DC-AE 2.0-Structured f64c128 models trained with PADT in Phase 1, evaluating full 128-channel reconstruction on MJHQ at 1024 resolution. Increasing the number of supervised prefixes from four to six does not improve full-channel reconstruction in this setting, supporting our choice of four supervision levels.

Channel budgetsPSNR ↑SSIM ↑LPIPS ↓
[32, 64, 96, 128]29.78840.85190.0891
[16, 32, 48, 64, 96, 128]29.49310.84620.0917
AuthorsOpenReview ↗

Official Comment by Authors (3/4)

W1 Q3 and Q4: Experiments coverage

Thank you for suggesting broader comparisons and evaluations on general-domain benchmarks. We have expanded the paired comparisons between standard DC-AE 2.0 and DC-AE 2.0-Structured, and added class-conditional ImageNet generation results. Due to time and resource constraints, we are not able to present the f128c512 results, and we will add this comparison in the final revision.

Paired comparisons between standard and Structured variants. Tables 1 and 4 include paired results at f32c128 and f64c128 for reconstruction and FLUX.2-dev generation at 1024 × 1024. We now add DC-AE 2.0-Structured f32c256 reconstruction results to Table 1. As shown below, it achieves reconstruction quality comparable to the standard variant across ImageNet, FFHQ, and OmniDoc.

ImageNet 256

AutoencoderPSNR ↑SSIM ↑LPIPS ↓rFID ↓
DC-AE 2.033.230.91150.04140.18
DC-AE 2.0-Structured33.280.91160.04130.18

FFHQ 1024

AutoencoderPSNR ↑SSIM ↑LPIPS ↓rFID ↓
DC-AE 2.039.500.95610.03590.13
DC-AE 2.0-Structured39.650.95640.03550.11

OmniDoc 256

AutoencoderPSNR ↑SSIM ↑LPIPS ↓rFID ↓NED ↑
DC-AE 2.029.780.96020.01750.730.9650
DC-AE 2.0-Structured29.750.96000.01780.780.9650

We also expand Table 4 with standard DC-AE 2.0 f32c128 results for FLUX.1-Krea at 1024 × 1024 and standard DC-AE 2.0 f64c128 results for FLUX.2-dev at 4096 × 4096. For FLUX.1-Krea, the Structured variant improves English LongText-Bench from 0.8607 to 0.8658. For FLUX.2-dev at 4096 × 4096, it improves English/Chinese LongText-Bench from 0.9528/0.9321 to 0.9601/0.9440. These results show consistent improvements in text rendering across the two added comparisons.

Class-conditional ImageNet generation. While our main experiments focus on text-to-image generation under deep compression, we agree that ImageNet generation results provide a complementary evaluation. We therefore add class-conditional ImageNet generation experiments at 512 × 512, using DiT-XL trained for 400k steps. We use 30 sampling steps without classifier-free guidance (CFG) and compute FID on 50,000 generated images. At f32c32, DC-AE 2.0 achieves generation FID comparable to DC-AE 1.0 (13.90 vs. 14.25). Its gains in fine-detail reconstruction may not be fully reflected in generation FID, which evaluates overall distributional similarity. At f32c128, DC-AE 2.0-Structured reduces FID from 39.92 to 25.33 compared with standard DC-AE 2.0, demonstrating the benefit of structured latents.

AutoencoderAE settingFID ↓
DC-AE 1.0f32c3214.25
DC-AE 2.0f32c3213.90
DC-AE 2.0f32c12839.92
DC-AE 2.0-Structuredf32c12825.33
Reviewer_JoMAOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces DC-Gen 2.0 to improve the efficiency of high-resolution image generation through deeply compressed latent spaces while mitigating the resulting degradation in reconstruction fidelity and fine-detail synthesis. Specifically, it presents DC-AE 2.0 for improved reconstruction under high compression, introduces Pyramid-Aligned Decoder Training to organize latent channels from global semantics to local details, and proposes Deep-Compression Detail Enhancement, which combines targeted text and face data with diffusion adaptation, distribution matching distillation, and reinforcement-learning-based post-training to improve generation quality on challenging tasks.

Critical strengths

Rather than focusing on a single architectural component such as DC-AE 2.0, the paper develops a comprehensive training pipeline encompassing multiple stages, including reinforcement learning, and outperforms existing autoencoders on text-rich, general-image, and facial reconstruction benchmarks.

Critical weaknesses

As a research paper, I primarily assess whether the work offers sufficient novelty and new insights. Although DC-Gen 2.0 is comprehensive and effective, its architecture and optimization pipeline rely heavily on existing methods. For example:

  • DC-AE 2.0 largely builds on the DC-AE family, while components such as global skip connections have already appeared in Qwen-Image-VAE-2.0.
  • Gaussian pyramids and multi-scale supervision are established techniques.
  • The post-training pipeline draws substantially from existing methods, including AnyFlow, DMD, and HPSv3++.

Therefore, despite the strong empirical results, the paper’s core methodological novelty appears limited.

  • Ablation Studies: It is unclear whether some improvements arise from the proposed architecture or the training data. For example, adding targeted text data improves Chinese LongText-Bench from 0.0745 to 0.9290, suggesting that much of the gain may come from large-scale specialized data rather than DC-AE 2.0 itself.
Questions for authors

see Weaknesses.

2 comments

Thank you for recognizing the comprehensiveness of DC-Gen 2.0 and its strong empirical results. We appreciate your constructive feedback. Below, we first clarify our methodological and engineering contributions, then present ablation studies to explain the sources of improvement in reconstruction and generation.

1 Methodological novelty and contribution

Deep-compression tokenizers address an important practical challenge in diffusion generation: reducing the cost of high-resolution generation while preserving quality. Fewer latent tokens can substantially improve efficiency, but aggressive spatial compression can degrade reconstruction and lower the potential fidelity ceiling for downstream generation. Increasing the latent channel count helps preserve information, yet makes the latent distribution harder for diffusion models to learn. DC-AE 1.5 alleviates this difficulty through structured latents and channel-masked diffusion training, but its original structured-training recipe introduces a reconstruction drop. To address this remaining bottleneck, we propose PADT, a new decoder-training strategy that recovers reconstruction quality while retaining structured latents. PADT is the core methodological contribution of our work.

Our targeted data curation and post-training are complementary engineering contributions that require nontrivial adaptation and tuning of existing techniques for deeply compressed latent spaces. These efforts aim to deliver a practical, efficient generation model that preserves fine details. We will release our models and post-training pipeline, providing reusable resources and practical guidance for the community.

Relationship to DC-AE and Qwen Image VAE

DC-AE 2.0 builds on the DC-AE family and adopts global skip connections from Qwen-Image-VAE-2.0, as acknowledged in Appendix A.1. However, we do not present these components as new technical contributions.

Joint alignment of channel budgets decoder outputs and target scales

We agree that Gaussian pyramids and multi-scale supervision are established techniques. However, PADT is motivated by the reconstruction drop of the original DC-AE 1.5 structured-training recipe. To address this practical problem, we propose a new training design that jointly aligns channel budgets, decoder stages, and target scales, recovering reconstruction quality while retaining structured latents. To the best of our knowledge, this design has not been proposed in prior deep-compression tokenizers.

Role of the post-training pipeline

Targeted data curation and post-training are complementary engineering contributions of our work to improve fine-grained detail generation in deeply compressed latent spaces. Our post-training pipeline builds on AnyFlow, DMD, and HPSv3++, as acknowledged in Sec. 3.4 and Appendix B. Table 9 demonstrates the effectiveness of our post-training pipeline for text and face generation, while the targeted-data ablations are discussed below.

2 Ablation studies and attribution of improvements

We agree that targeted training data are important, and data curation is one of our engineering contributions. Our controlled ablations demonstrate its benefits for both reconstruction and generation.

For autoencoder training, targeted data improve all five OmniDoc reconstruction metrics while keeping the architecture and training settings unchanged (Table 3). Separately, under identical downstream adaptation and post-training data and settings (Appendix B, Table 8), DC-AE 2.0 also improves fine-detail generation over DC-AE 1.0 at the same f32c32 configuration using FLUX.2-dev at 1024 × 1024 (Table 4), improving English TextPecker score from 95.30 to 97.14, Chinese TextPecker score from 77.44 to 82.40, FaceScore mean from 4.050 to 4.459, and P10 from 0.129 to 2.149.

For diffusion training, targeted data increase the Chinese LongText-Bench score from 0.0745 to 0.9290, with the autoencoder and all other training and evaluation settings held fixed (Table 5).

Beyond data curation, PADT introduces a new decoder-training design that preserves structured latents while recovering the reconstruction quality lost under the DC-AE 1.5 recipe. Table 2 isolates this algorithmic contribution by comparing the two training approaches under identical initialization, data, and training settings. PADT improves all reported reconstruction metrics across ImageNet, FFHQ, and OmniDoc, achieving reconstruction quality comparable to standard DC-AE 2.0 while retaining a structured latent space.

Table 4 further evaluates the benefits of PADT for fine-grained text generation. On FLUX.2-dev, DC-AE 2.0-Structured consistently improves both English and Chinese LongText-Bench scores across the evaluated resolutions and compression ratios. At 1024 × 1024, it improves the Chinese score from 0.9338 to 0.9467 at f32c128 and from 0.9187 to 0.9233 at f64c128. At 4096 × 4096, it improves the Chinese score from 0.9321 to 0.9440 at f64c128.

Dear Reviewer JoMA,

With less than a day remaining in the discussion period, we wanted to briefly follow up on our rebuttal. We would appreciate hearing whether our response addresses your concerns about methodological novelty and the attribution of improvements, and whether any points remain unclear.

We would be happy to discuss any remaining questions before the deadline. Thank you again for your time and constructive feedback.

Best regards, The Authors

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

Related papers

Discussion (0)

Sign in to comment.