pith. sign in

arxiv: 2605.23902 · v1 · pith:Z2OHCR7Cnew · submitted 2026-05-22 · 💻 cs.CV

PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion

Yifan Lu , Qi Wu , Jay Zhangjie Wu , Zian Wang , Huan Ling , Sanja Fidler , Xuanchi Ren This is my paper

Pith reviewed 2026-05-25 04:16 UTC · model grok-4.3

classification 💻 cs.CV
keywords pixel diffusion decoderlatent decodinghigh-resolution image synthesisdiffusion modelsimage upsamplingtext-to-image generationVAE latentssemantic latents
0
0 comments X

The pith

PiD reformulates latent decoding as conditional pixel diffusion to synthesize high-resolution images directly from compact latents.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces PiD to address the bottleneck of reconstruction-oriented decoders in latent diffusion and autoregressive image systems, which grow costly at megapixel scales. It casts decoding as a generative pixel-space diffusion process conditioned on latents, so that upsampling to 4x or 8x resolution occurs inside the same denoising loop. A lightweight sigma-aware adapter feeds noise-corrupted latents into the backbone, allowing the latent diffusion stage to stop early. Distillation with DMD2 reduces the process to four steps. The result is practical high-resolution output from 512x512 latents in under a second on consumer GPUs, with lower memory than cascaded super-resolution pipelines.

Core claim

PiD is a pixel diffusion decoder that unifies latent decoding and upsampling by denoising directly in high-resolution pixel space; a sigma-aware adapter injects noise-corrupted latents into the diffusion backbone so that partially denoised latents can be decoded and the latent diffusion process terminated early, with further distillation to four inference steps, yielding 2048x2048 outputs from 512x512 latents in under one second on an RTX 5090 while supporting both VAE and semantic latents.

What carries the argument

The sigma-aware adapter that injects noise-corrupted latents into the pixel diffusion backbone to enable conditional generation and early termination of latent diffusion.

If this is right

  • Decoding and upsampling become a single generative module instead of sequential reconstruction and super-resolution stages.
  • Latent diffusion pipelines can stop at intermediate denoising steps without major quality degradation.
  • The distilled four-step model runs at 210 ms on a GB200 GPU while using 13 GB peak memory on an RTX 5090.
  • The same architecture works for both conventional VAE latents and semantic latents such as SigLIP or DINOv2.
  • Inference is approximately six times faster than cascaded diffusion-based super-resolution pipelines at equal or higher visual fidelity.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The approach could support interactive high-resolution editing tools where latency must stay below one second.
  • Similar conditioning adapters might accelerate other latent-based generative tasks such as video or 3D synthesis.
  • Further reduction in step count or memory footprint could make megapixel generation feasible on edge devices.
  • The unification of decoding and upsampling raises the question of whether end-to-end training from text to pixels can bypass separate latent stages altogether.

Load-bearing premise

The lightweight sigma-aware adapter can successfully inject noise-corrupted latents into the pixel diffusion backbone so that the model can decode partially denoised latents and terminate the latent diffusion process early without major quality loss.

What would settle it

A controlled comparison in which early termination of latent diffusion with the adapter produces visibly lower fidelity or more artifacts than full latent diffusion plus a standard decoder would falsify the efficiency-without-quality-loss claim.

Figures

Figures reproduced from arXiv: 2605.23902 by Huan Ling, Jay Zhangjie Wu, Qi Wu, Sanja Fidler, Xuanchi Ren, Yifan Lu, Zian Wang.

Figure 1
Figure 1. Figure 1: PiD directly decodes latents from VAE or vision encoders into higher-resolution images, replacing the decode–then–upsample cascade while achieving lower latency and higher visual quality. © 2026 NVIDIA. All rights reserved. arXiv:2605.23902v1 [cs.CV] 22 May 2026 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: 4K decoding results. PiD synthesize more details at 4k resolution. 2 [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of PiD. PiD unifies latent decoding and upsampling as a single latent-conditioned pixel diffusion model that predicts the target-resolution pixel-space velocity field. Noise-corrupted latent training and sigma-aware gating make the decoder robust to partially denoised latents, enabling early exit from the base LDM while preserving high-resolution output quality. rotary positional encoding (RoPE) [… view at source ↗
Figure 4
Figure 4. Figure 4: Pairwise image quality preference judged by closed-source MLLMs. Three MLLMs compare PiD-decoded images against cascaded baselines (original VAE decoder followed by SR). All three MLLMs consistently prefer PiD, with high 2-round consistency under image order swap [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Image reconstruction comparison. Given a latent encoded from a clean image, PiD reconstructs the image at higher resolution with sharper details than the original VAE / RAE decoder. VAE decoding PiD decoding 16/28 step 20/28 step 24/28 step 26/28 step full (28) step [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: VAE decoding and PiD decoding at different LDM termination steps. Top: VAE decoding. Bottom: PiD decoding. With the full LDM denoising steps, PiD is faithful to the latent’s VAE decoding results; at intermediate steps, because the base latent diffusion model has not denoised all the subtle details, it allows PiD to imagine additional details. 4.4. Qualitative Evaluation Reconstruction on real-world image’s… view at source ↗
Figure 7
Figure 7. Figure 7: PiD vs. cascaded super-resolution. From a FLUX.1 [dev] latent of a 5122 image, baselines apply a super-resolution model on VAE decoding output, while PiD decodes directly to 20482 . PiD produces sharper detail at lower latency. Latency is measured on a single GB200 GPU with torch.compile. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Image quality of PiD at different LDM termination step for FLUX.1 [dev] (28 denoising steps in total) [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison with native 2048 × 2048 px generation. Coupling a low-resolution LDM with PiD substantially reduces inference time while maintaining image quality competitive to native high-resolution generation, and in some cases surpassing it in fine-grained details. Latency (in lower left corner) is measured on a single GB200 GPU without torch.compile. compare native 2K generation of FLUX.2, PixelDiT, and FL… view at source ↗
Figure 10
Figure 10. Figure 10: MLLM evaluation sample 1. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: MLLM evaluation sample 2. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
read the original abstract

Most practical high-resolution text-to-image systems, including latent diffusion and autoregressive models, perform generation in a compact latent space, and a decoder maps the generated latents back to pixels. Yet the latent-to-pixel decoder is reconstruction-oriented, optimized to invert the encoder rather than synthesize more details, and becomes increasingly costly at megapixel scale. This drawback calls for a more expressive and efficient decoding paradigm. Motivated by recent progress in scalable pixel-space diffusion, we introduce PiD, a Pixel diffusion Decoder that reformulates latent decoding as conditional pixel diffusion, unifying decoding and upsampling into one generative module. By denoising directly in high-resolution pixel space, PiD synthesizes $4\times$ and even $8\times$ upscaled images with low latency. For latent conditioning, a lightweight sigma-aware adapter injects noise-corrupted latents into the pixel diffusion backbone, enabling PiD to decode partially denoised latents and terminate the latent diffusion process early. To further improve efficiency, we distill the model using DMD2, reducing inference to just 4 steps. PiD applies to both conventional VAE latents and semantic latents (e.g., SigLIP, DINOv2) used in recent RAE-based models. PiD decodes latents of $512 \times 512$ images into $2048 \times 2048$ pixels in under 1 second with 13 GB peak memory on a consumer RTX 5090, and as fast as 210 ms on a GB200 GPU, about $6\times$ faster than cascaded diffusion-based super-resolution pipelines with better visual fidelity.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper introduces PiD, a Pixel diffusion Decoder that reformulates latent decoding as conditional pixel diffusion to unify decoding and upsampling. It uses a lightweight sigma-aware adapter to inject noise-corrupted latents into a pixel diffusion backbone, enabling early termination of latent diffusion, and applies DMD2 distillation to reduce inference to 4 steps. The method is claimed to apply to both VAE and semantic latents, delivering 2048x2048 outputs from 512x512 latents in under 1s (13GB on RTX 5090) or 210ms on GB200, with 6x speedup and better fidelity versus cascaded diffusion super-resolution.

Significance. If the reported speed, memory, and quality advantages hold under rigorous evaluation, PiD could meaningfully simplify high-resolution latent generative pipelines by replacing separate reconstruction decoders and cascaded upsamplers with a single generative module. The unification of decoding with pixel-space diffusion and support for semantic latents are potentially impactful for both conventional LDMs and recent RAE-style models.

major comments (2)
  1. [Abstract] Abstract: performance claims (under-1s decoding, 6x speedup, better visual fidelity) are stated without any reference to experimental protocol, datasets, baselines, ablations, or quantitative metrics (e.g., FID, LPIPS, user studies). This absence prevents verification of the central efficiency and quality assertions.
  2. [Abstract] Abstract: the effectiveness of the sigma-aware adapter for injecting partially denoised latents and safely terminating latent diffusion early is presented as a core enabling mechanism, yet no supporting derivation, training objective, or ablation is referenced, leaving the weakest assumption unexamined.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed feedback. We address the two major comments on the abstract below. The full manuscript contains the requested experimental details, but we agree the abstract can be strengthened for clarity.

read point-by-point responses

  1. Referee: [Abstract] Abstract: performance claims (under-1s decoding, 6x speedup, better visual fidelity) are stated without any reference to experimental protocol, datasets, baselines, ablations, or quantitative metrics (e.g., FID, LPIPS, user studies). This absence prevents verification of the central efficiency and quality assertions.

    Authors: The abstract is a concise summary and therefore omits explicit citations to the evaluation protocol. The manuscript reports these details in the Experiments section, including datasets (ImageNet, COCO, LAION subsets), baselines (cascaded diffusion SR pipelines), quantitative metrics (FID, LPIPS), user studies, and ablations. We will revise the abstract to add a short clause referencing the evaluation protocol and key metrics used to support the claims. revision: partial

  2. Referee: [Abstract] Abstract: the effectiveness of the sigma-aware adapter for injecting partially denoised latents and safely terminating latent diffusion early is presented as a core enabling mechanism, yet no supporting derivation, training objective, or ablation is referenced, leaving the weakest assumption unexamined.

    Authors: The abstract highlights the adapter's role at a high level. The manuscript provides the adapter architecture, sigma-aware conditioning derivation, training objective, and dedicated ablations in Sections 3 and 4. We will revise the abstract to include a brief reference to these supporting analyses in the body of the paper. revision: partial

Circularity Check

0 steps flagged

No significant circularity

full rationale

The provided abstract and description present PiD as a new construction that reformulates latent decoding as conditional pixel diffusion, motivated by external pixel-space diffusion progress. No equations, fitted parameters, or predictions are shown that reduce to the authors' own prior results by definition. The sigma-aware adapter, DMD2 distillation, and early termination are described as engineering choices without self-definitional or self-citation load-bearing reductions. The reader's assessment of score 1.0 is consistent with an independent construction; absent any quoted derivation chain that collapses to inputs, the default non-circular finding applies.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract only; no explicit free parameters, axioms, or invented entities are identifiable from the provided text.

pith-pipeline@v0.9.0 · 5842 in / 1264 out tokens · 34397 ms · 2026-05-25T04:16:51.863903+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

65 extracted references · 23 canonical work pages · 13 internal anchors

  1. [1]

    Arora, Z

    A. Arora, Z. Tu, Y. Wang, R. Bai, J. Wang, and S. Ma. Guidesr: Rethinking guidance for one-step high-fidelity diffusion-based super-resolution. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 3914–3922, 2025. 5

    2025

  2. [2]

    S. Bai, Y. Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Ge, W. Ge, Z. Guo, Q. Huang, J. Huang, F. Huang, B. Hui, S. Jiang, Z. Li, M. Li, M. Li, K. Li, Z. Lin, J. Lin, X. Liu, J. Liu, C. Liu, Y. Liu, D. Liu, S. Liu, D. Lu, R. Luo, C. Lv, R. Men, L. Meng, X. Ren, X. Ren, S. Song, Y. Sun, J. Tang, J. Tu, J. Wan, P. Wang, P. Wang, Q....

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Betker, G

    J. Betker, G. Goh, L. Jing, T. Brooks, J. Wang, L. Li, L. Ouyang, J. Zhuang, J. Lee, Y. Guo, et al. Improving image generation with better captions.Computer Science, 2(3):8, 2023. 4

    2023

  4. [4]

    T. Bi, X. Zhang, Y. Lu, and N. Zheng. Vision foundation models can be good tokenizers for latent diffusion models. arXiv preprint arXiv:2510.18457, 2025. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    Ntk-aware scaled rope.https://www.reddit.com/r/LocalLLaMA/comments/14lz7j5/ntkaware_scaled_ rope_allows_llama_models_to_have/

    bloc97. Ntk-aware scaled rope.https://www.reddit.com/r/LocalLLaMA/comments/14lz7j5/ntkaware_scaled_ rope_allows_llama_models_to_have/. 6

  6. [6]

    H. Cai, S. Cao, R. Du, P. Gao, S. Hoi, Z. Hou, S. Huang, D. Jiang, X. Jin, L. Li, et al. Z-image: An efficient image generation foundation model with single-stream diffusion transformer.arXiv preprint arXiv:2511.22699, 2025. 9

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  7. [7]

    S. Cao, J. Li, X. Li, Y. Pu, K. Zhu, Y. Gao, S. Luo, Y. Xin, Q. Qin, Y. Zhou, et al. Unipercept: Towards unified perceptual- level image understanding across aesthetics, quality, structure, and texture.arXiv preprint arXiv:2512.21675, 2025. 9, 10

    work page arXiv 2025

  8. [8]

    J. Chen, H. Cai, J. Chen, E. Xie, S. Yang, H. Tang, M. Li, Y. Lu, and S. Han. Deep compression autoencoder for efficient high-resolution diffusion models.arXiv preprint arXiv:2410.10733, 2024. 4

    work page arXiv 2024

  9. [9]

    J. Chen, C. Ge, E. Xie, Y. Wu, L. Yao, X. Ren, Z. Wang, P. Luo, H. Lu, and Z. Li. PixArt-Σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. InEuropean Conference on Computer Vision (ECCV), 2024. 5

    2024

  10. [10]

    L. Dong, Q. Fan, Y. Guo, Z. Wang, Q. Zhang, J. Chen, Y. Luo, and C. Zou. Tsd-sr: One-step diffusion with target score distillation for real-world image super-resolution. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 23174–23184, 2025. 5, 9, 10

    2025

  11. [11]

    Esser, S

    P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Müller, H. Saini, Y. Levi, D. Lorenz, A. Sauer, F. Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. InForty-first international conference on machine learning, 2024. 5, 9

    2024

  12. [12]

    J. Heek, E. Hoogeboom, T. Mensink, and T. Salimans. Unified latents (ul): How to train your latents.arXiv preprint arXiv:2602.17270, 2026. 4

    work page arXiv 2026

  13. [13]

    J. Ho, C. Saharia, W. Chan, D. J. Fleet, M. Norouzi, and T. Salimans. Cascaded diffusion models for high fidelity image generation.Journal of Machine Learning Research, 23(47):1–33, 2022. 5

    2022

  14. [14]

    X. Hu, R. Wang, Y. Fang, B. Fu, P. Cheng, and G. Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment.arXiv preprint arXiv:2403.05135, 2024. 9

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  15. [15]

    Jeong, S

    J. Jeong, S. Han, J. Kim, and S. J. Kim. Latent space super-resolution for higher-resolution image generation with diffusion models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 2355–2365, 2025. 5

    2025

  16. [16]

    Kawai, T

    K. Kawai, T. Oba, K. Tokoro, K. Akita, and N. Ukita. Efficient burst super-resolution with one-step diffusion. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 864–873, 2025. 5

    2025

  17. [17]

    J. Ke, Q. Wang, Y. Wang, P. Milanfar, and F. Yang. Musiq: Multi-scale image quality transformer. InProceedings of the IEEE/CVF international conference on computer vision, pages 5148–5157, 2021. 9 19 PiD : Fast and High-Resolution Latent Decoding with Pixel Diffusion

    2021

  18. [18]

    D. P. Kingma and M. Welling. Auto-encoding variational Bayes. InInternational Conference on Learning Representations (ICLR), 2014. 3, 4

    2014

  19. [19]

    B. F. Labs. Flux.https://github.com/black-forest-labs/flux, 2024. 9

    2024

  20. [20]

    B. F. Labs. FLUX.2: Frontier Visual Intelligence.https://bfl.ai/blog/flux-2, 2025. 9

    2025

  21. [21]

    Back to Basics: Let Denoising Generative Models Denoise

    T. Li and K. He. Back to basics: Let denoising generative models denoise.arXiv preprint arXiv:2511.13720, 2025. 3, 5

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  22. [22]

    T. Li, Y. Tian, H. Li, M. Deng, and K. He. Autoregressive image generation without vector quantization. InAdvances in Neural Information Processing Systems (NeurIPS), 2024. 3, 4

    2024

  23. [23]

    X. Lin, J. He, Z. Chen, Z. Lyu, B. Dai, F. Yu, Y. Qiao, W. Ouyang, and C. Dong. DiffBIR: Toward blind image restoration with generative diffusion prior. InEuropean Conference on Computer Vision, pages 430–448. Springer, 2024. 5

    2024

  24. [24]

    Flow Matching for Generative Modeling

    Y. Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022. 7

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  25. [25]

    X. Liu, C. Gong, and Q. Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022. 7

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  26. [26]

    completely blind

    A. Mittal, R. Soundararajan, and A. C. Bovik. Making a “completely blind” image quality analyzer.IEEE Signal processing letters, 20(3):209–212, 2012. 9

    2012

  27. [27]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V. Vo, M. Szafraniec, V. Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El- Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y. Huang, S.-W. Li, I. Misra, M. Rabbat, V. Sharma, G. Synnaeve, H. Xu, H. Jégou, J. Mairal, P. Labatut, A. Joulin, and P. Bojanowski. DINOv2: Learning robust visual features without super...

    2024

  28. [28]

    Razin, D

    A. Razin, D. Kazantsev, and I. Makarov. One small step in latent, one giant leap for pixels: Fast latent upscale adapter for your diffusion models.arXiv preprint arXiv:2511.10629, 2025. 5, 9, 10

    work page arXiv 2025

  29. [29]

    Stochasticbackpropagationandapproximateinferenceindeepgenerative models

    D.J.Rezende, S.Mohamed, andD.Wierstra. Stochasticbackpropagationandapproximateinferenceindeepgenerative models. InInternational Conference on Machine Learning (ICML), 2014. 3, 4

    2014

  30. [30]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695,

  31. [31]

    Saharia, J

    C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi. Image super-resolution via iterative refinement. arXiv preprint arXiv:2104.07636, 2021. 5

    work page arXiv 2021

  32. [32]

    J. Shi, C. Wu, J. Liang, X. Liu, and N. Duan. DiVAE: Photorealistic images synthesis with denoising diffusion decoder,

  33. [33]

    J. Su, M. Ahmed, Y. Lu, S. Pan, W. Bo, and Y. Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 6

    2024

  34. [34]

    P. Sun, Y. Jiang, S. Chen, S. Zhang, B. Peng, P. Luo, and Z. Yuan. Autoregressive model beats diffusion: Llama for scalable image generation.arXiv preprint arXiv:2406.06525, 2024. 3, 4

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  35. [35]

    Y. Sun, L. Sun, S. Liu, R. Wu, Z. Zhang, and L. Zhang. One-step diffusion for detail-rich and temporally consistent video super-resolution. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. 5

    2025

  36. [36]

    G. Team, M. Riviere, S. Pathak, P. G. Sessa, C. Hardin, S. Bhupatiraju, L. Hussenot, T. Mesnard, B. Shahriari, A. Ramé, et al. Gemma 2: Improving open language models at a practical size.arXiv preprint arXiv:2408.00118, 2024. 8

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  37. [37]

    K. Tian, Y. Jiang, Z. Yuan, B. Peng, and L. Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. InAdvances in Neural Information Processing Systems (NeurIPS), 2024. 3, 4 20 PiD : Fast and High-Resolution Latent Decoding with Pixel Diffusion

    2024

  38. [38]

    S. Tong, B. Zheng, Z. Wang, B. Tang, N. Ma, E. Brown, J. Yang, R. Fergus, Y. LeCun, and S. Xie. Scaling text-to-image diffusion transformers with representation autoencoders, 2026. 4, 9

    2026

  39. [39]

    Vallaeys, J

    T. Vallaeys, J. Verbeek, and M. Cord. SSDD: Single-step diffusion decoder for efficient image tokenization.arXiv preprint arXiv:2510.04961, 2025. 3, 4, 9

    work page arXiv 2025

  40. [40]

    Van Den Oord, O

    A. Van Den Oord, O. Vinyals, et al. Neural discrete representation learning.Advances in neural information processing systems, 30, 2017. 4

    2017

  41. [41]

    J. Wang, S. Lin, Z. Lin, Y. Ren, M. Wei, Z. Yue, S. Zhou, H. Chen, Y. Zhao, C. Yang, et al. Seedvr2: One-step video restoration via diffusion adversarial post-training.arXiv preprint arXiv:2506.05301, 2025. 9, 10

    work page arXiv 2025

  42. [42]

    J. Wang, Z. Yue, S. Zhou, K. C. K. Chan, and C. C. Loy. Exploiting diffusion prior for real-world image super-resolution,

  43. [43]

    X. Wang, L. Xie, C. Dong, and Y. Shan. Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data. InIEEE/CVF International Conference on Computer Vision Workshops (ICCVW), 2021. 5, 9, 10

    2021

  44. [44]

    X. Wang, K. Yu, S. Wu, J. Gu, Y. Liu, C. Dong, C. C. Loy, X. Tang, and Y. Qiao. ESRGAN: Enhanced super-resolution generative adversarial networks. InEuropean Conference on Computer Vision Workshops (ECCVW), 2018. 5

    2018

  45. [45]

    H. Wu, Z. Zhang, W. Zhang, C. Chen, C. Li, L. Liao, A. Wang, E. Zhang, W. Sun, Q. Yan, X. Min, G. Zhai, and W. Lin. Q-align: Teaching lmms for visual scoring via discrete text-defined levels.arXiv preprint arXiv:2312.17090, 2023. Equal Contribution by Wu, Haoning and Zhang, Zicheng. Project Lead by Wu, Haoning. Corresponding Authors: Zhai, Guangtai and Li...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  46. [46]

    R. Wu, T. Yang, L. Sun, Z. Zhang, S. Li, and L. Zhang. SeeSR: Towards semantics-aware real-world image super- resolution. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25456–25467, 2024. 5

    2024

  47. [47]

    T. Wu, J. Zou, J. Liang, L. Zhang, and K. Ma. Visualquality-r1: Reasoning-induced image quality assessment via reinforcement learning to rank.arXiv preprint arXiv:2505.14460, 2025. 9, 10, 14

    work page arXiv 2025

  48. [48]

    X. Wu, J. Xin, J. Hao, H. Gao, J. Li, N. Wang, and X. Gao. One-step diffusion-based real-world image super-resolution with visual perception distillation.Neurocomputing, page 133066, 2026. 5

    2026

  49. [49]

    E. Xie, J. Chen, J. Chen, H. Cai, H. Tang, Y. Lin, Z. Zhang, M. Li, L. Zhu, Y. Lu, and S. Han. SANA: Efficient high-resolution image synthesis with linear diffusion transformer, 2024. 5

    2024

  50. [50]

    S. Yang, T. Wu, S. Shi, S. Lao, Y. Gong, M. Cao, J. Wang, and Y. Yang. Maniqa: Multi-dimension attention network for no-reference image quality assessment. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1191–1200, 2022. 9

    2022

  51. [51]

    T. Ye, S. Fei, and L. Zhu. UltraFlux: Data-model co-design for high-quality native 4k text-to-image generation across diverse aspect ratios.arXiv preprint arXiv:2511.18050, 2025. 5, 8

    work page arXiv 2025

  52. [52]

    T. Yin, M. Gharbi, T. Park, R. Zhang, E. Shechtman, F. Durand, and W. T. Freeman. Improved distribution matching distillation for fast image synthesis.Advances in neural information processing systems, 37:47455–47487, 2024. 4, 7, 8

    2024

  53. [53]

    W. You, M. Zhang, L. Zhang, X. Zhou, K. Shi, and S. Gu. Consistency trajectory matching for one-step generative super-resolution. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 12747–12756,

  54. [54]

    Z. You, X. Cai, J. Gu, T. Xue, and C. Dong. Teaching large language models to regress accurate image quality scores using score distribution. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 14483–14494,

  55. [55]

    F. Yu, J. Gu, Z. Li, J. Hu, X. Kong, X. Wang, J. He, Y. Qiao, and C. Dong. Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25669–25680, 2024. 5 21 PiD : Fast and High-Resolution Latent Decoding with Pixel ...

    2024

  56. [56]

    J. Yu, Y. Xu, J. Y. Koh, T. Luong, G. Baid, Z. Wang, V. Vasudevan, A. Ku, Y. Yang, B. K. Ayan, B. Hutchinson, W. Han, Z. Parekh, X. Li, H. Zhang, J. Baldridge, and Y. Wu. Scaling autoregressive models for content-rich text-to-image generation.Transactions on Machine Learning Research, 2022. 3, 4

    2022

  57. [57]

    L. Yu, J. Lezama, N. B. Gundavarapu, L. Versari, K. Sohn, D. Minnen, Y. Cheng, V. Birodkar, A. Gupta, X. Gu, et al. Language model beats diffusion–tokenizer is key to visual generation.arXiv preprint arXiv:2310.05737, 2023. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  58. [58]

    Q. Yu, M. Weber, X. Deng, X. Shen, D. Cremers, and L.-C. Chen. An image is worth 32 tokens for reconstruction and generation.Advances in Neural Information Processing Systems, 37:128940–128966, 2024. 4

    2024

  59. [59]

    Y. Yu, W. Xiong, W. Nie, Y. Sheng, S. Liu, and J. Luo. Pixeldit: Pixel diffusion transformers for image generation.arXiv preprint arXiv:2511.20645, 2025. 3, 5

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  60. [60]

    Z. Yue, K. Liao, and C. C. Loy. Arbitrary-steps image super-resolution via diffusion inversion. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 23153–23163, 2025. 5, 9, 10

    2025

  61. [61]

    Z. Yue, J. Wang, and C. C. Loy. ResShift: Efficient diffusion model for image super-resolution by residual shifting. arXiv preprint arXiv:2307.12348, 2023. 5

    work page arXiv 2023

  62. [62]

    X. Zhai, B. Mustafa, A. Kolesnikov, and L. Beyer. Sigmoid loss for language image pre-training. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975–11986, 2023. 4, 9

    2023

  63. [63]

    Zhang, A

    L. Zhang, A. Rao, and M. Agrawala. Adding conditional control to text-to-image diffusion models. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3836–3847, 2023. 6

    2023

  64. [64]

    L. Zhao, S. Woo, Z. Wan, Y. Li, H. Zhang, B. Gong, H. Adam, X. Jia, and T. Liu. Epsilon-VAE: Denoising as visual decoding. InProceedings of the 42nd International Conference on Machine Learning (ICML), 2025. 3, 4

    2025

  65. [65]

    Diffusion Transformers with Representation Autoencoders

    B. Zheng, N. Ma, S. Tong, and S. Xie. Diffusion transformers with representation autoencoders.arXiv preprint arXiv:2510.11690, 2025. 3, 4, 9 22

    work page internal anchor Pith review Pith/arXiv arXiv 2025