Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

A single-step, GAN-free diffusion decoder can beat KL-VAE on reconstruction quality and speed.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 11:21 UTC pith:57IXJSIC

load-bearing objection A genuine step forward for single-step diffusion decoding, with a clear metric-framing caveat and a few mechanical fixes needed before the central claim is fully trustworthy. the 4 major comments →

arxiv 2510.04961 v2 pith:57IXJSIC submitted 2025-10-06 cs.CV

SSDD: Single-Step Diffusion Decoder for Efficient Image Tokenization

Th\'eophane Vallaeys , Jakob Verbeek , Matthieu Cord This is my paper
classification cs.CV
keywords single-step diffusion decoderimage tokenizationflow matchingGAN-free trainingdistillationKL-VAEperception-distortion trade-offlatent diffusion
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Standard image tokenizers rely on a deterministic KL-VAE decoder trained with reconstruction, perceptual, and adversarial losses. This paper argues that a diffusion decoder—trained without any GAN loss and then distilled to a single step—can replace that decoder and win on the metrics that matter for generation: perceptual quality and distribution shift. On ImageNet 256×256 at 8×4 compression, the smallest SSDD model achieves rFID 0.50 versus 0.87 for KL-VAE while decoding 1.4× faster; larger models push rFID to 0.38. When plugged into a DiT-XL/2 generator, SSDD preserves or improves generation FID while cutting total sampling time by up to 3.8× (at f16c4 compression with SSDD-L). The paper explicitly accepts lower PSNR/SSIM in exchange, framing the result as a deliberate move along the perception–distortion trade-off.

Core claim

Central claim: the standard recipe for image tokenization—a deterministic decoder plus adversarial losses—can be replaced by a generative diffusion decoder that runs in a single step and requires no GAN. The authors build a U-Net/transformer hybrid pixel-space decoder (U-ViT style), train it with a flow-matching objective plus LPIPS and REPA feature-alignment losses, then distill the multi-step sampler into a one-step student. The result, SSDD, reconstructs images with lower rFID, LPIPS, and DreamSim than KL-VAE across encoder configurations (f8c4, f16c16, f16c4), with higher decoding throughput, and it improves DiT generation FID while being up to 3.8× faster end-to-end at higher latent com

What carries the argument

The load-bearing mechanism is the single-step distilled decoder produced from a teacher trained to navigate the perception–distortion trade-off. The teacher is a pixel-space flow-matching decoder built as a U-Net with a transformer middle block (U-ViT style); it is trained without any GAN loss using L_fm + 0.5*LPIPS + 0.25*REPA, where REPA aligns transformer features to DINOv2 representations. The number of sampling steps N (with scheduler t_i = ((N-i+1)/N)^rho, rho=2, N=8) selects where the model sits on the perception–distortion curve: more steps improve rFID/LPIPS but hurt PSNR. Distillation then fine-tunes the student with the same losses computed against teacher outputs, freezing the se

Load-bearing premise

The 'higher reconstruction quality' claim stands or falls on using perceptual and distribution-shift metrics (rFID, LPIPS, DreamSim) as the definition of quality; SSDD is consistently worse on PSNR and SSIM, so any application that treats pixel-level distortion as operative inverts the conclusion.

What would settle it

A reader could settle the claim by computing rFID and PSNR on a held-out set of images where KL-VAE and SSDD are matched in throughput and parameter count: if SSDD's rFID advantage disappears or if the PSNR gap exceeds the perceptual gains on a diverse test set (e.g., faces or text, not just ImageNet), the 'higher reconstruction quality' claim would not generalize as stated. More directly, reproducing Table S2 at f16c4 tests the distillation assumption: the paper already shows rFID degrades from 2.25 to 2.49 for SSDD-S after distillation, so a broader range of encoders/compression ratios where

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Existing KL-VAE-based latent diffusion pipelines can swap in a distilled SSDD decoder without retraining the encoder or generator, since SSDD trains successfully on frozen encoders (the paper demonstrates this with SD-VAE and DiTo encoders).
  • The 3.8× end-to-end speedup at f16c4 shows that high-compression latents are no longer penalized by expensive decoding, making high-compression settings more attractive for generative modeling.
  • The optimal-step-count behavior tied to LPIPS regularization means sampling-step choice can serve as a dial on the perception–distortion trade-off, and distillation can freeze that dial into a one-step decoder.
  • A GAN-free training recipe with LPIPS+REPA losses suggests that adversarial training is not necessary to match or beat KL-VAE tokenization quality, which could simplify and stabilize tokenizer training at scale.

Where Pith is reading between the lines

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

  • An implication the author leaves implicit: the paper's own Table S2 shows the distilled student degrades more at high compression (f16c4: rFID rises 2.25→2.49 for SSDD-S), so the 'single-step without quality loss' claim is best read as configuration-dependent; very high compression ratios may still need multi-step decoding or stronger distillation.
  • Because SSDD wins on perceptual/distribution-shift metrics and loses on PSNR/SSIM, downstream tasks that require pixel-level fidelity—image compression with distortion guarantees, scientific imaging, or editing that expects preservation—would likely prefer the KL-VAE; the paper's stated goal is generative realism, not exact recovery.
  • A natural testable extension is to distill teachers trained with different sampling schedules (different rho or N) to produce a family of single-step decoders spanning the perception–distortion curve, letting applications pick fidelity vs. realism in one forward pass instead of eight.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper introduces SSDD, a diffusion-decoder tokenizer designed to replace KL-VAE decoders in latent image-generation pipelines. The decoder uses a U-ViT-inspired hybrid U-Net/transformer architecture, is trained with flow matching plus LPIPS and REPA losses (no adversarial loss), and is distilled to a single sampling step. On ImageNet 256×256, the authors report that SSDD improves reconstruction FID relative to KL-VAE and prior diffusion decoders (e.g., 0.87 → 0.50 for f8c4), improves LPIPS and DreamSim, increases decoder throughput by about 1.4× at f8c4, and improves downstream DiT generation FID while speeding up the full generation pipeline by up to 3.8× with an f16c4 latent. The paper includes extensive ablations (Table 4), scaling studies (Tables 2 and S4), distillation analysis (Table S2, Figures S1–S3), and frozen-encoder experiments (Table S3).

Significance. If the results hold, SSDD is a meaningful step: it is the first diffusion-decoder tokenizer that is both GAN-free and single-step, and it appears to Pareto-dominate KL-VAE on perceptual/distribution-level reconstruction metrics while also improving generation speed and quality. The paper is unusually transparent: it reports the configurations where the method degrades (e.g., f16c4 distillation in Table S2), discloses that PSNR/SSIM are worse, and provides code and detailed hyperparameters. The empirical basis is solid enough that the central technical claim — that a distilled, GAN-free diffusion decoder can replace a KL-VAE in common latent-diffusion pipelines — is credible. The main weaknesses are in the framing of the headline claims, which are stated more broadly than the evidence supports.

major comments (4)
  1. [Abstract and Section 1] The abstract states 'SSDD improves reconstruction FID from 0.87 to 0.46', but Section 1 and Table 1 report 0.87 → 0.50 for the f8c4 configuration. The same inconsistency appears in the abstract block and the full text. Since this is the headline quantitative claim, the discrepancy must be resolved before the paper can be relied upon.
  2. [Abstract, Section 1, and Table 1] The unqualified claims 'higher reconstruction quality' and 'faster sampling than KL-VAE' are metric- and configuration-dependent. Table 1 shows SSDD is consistently worse on PSNR and SSIM (e.g., f8c4: PSNR 25.12 vs 25.68; SSIM 0.77 vs 0.82 for SSDD-H vs KL-VAE), and Section 4.2 itself says these distortion metrics 'should be de-emphasized.' The abstract should say 'perceptual and distribution-level reconstruction quality' rather than imply a universal quality improvement. Similarly, 'faster sampling than KL-VAE' is not true at f16c16: Table 1 lists KL-VAE at 1196 images/s versus SSDD-S at 985 images/s. The speed advantage is specific to f8c4 (1.4×) or to the full DiT pipeline with f16c4 (3.8×), not to the decoder alone in all configurations.
  3. [Section 3.3 vs Section B] The sampling and distillation protocol is inconsistent. Section 3.3 states 'We use a shift parameter of ρ=2 with N=8 sampling steps for all our SSDD models', and Table S2 reports results for '8-steps sampling'. Section B, however, says 'For distillation, we use a 7-steps teacher model'. This discrepancy affects the reproducibility of the distillation results and the interpretation of Table S2. Please specify which teacher step count was actually used and align the text, tables, and supplementary material.
  4. [Table 1 and Section 3.4] The comparison in Table 1 is described as 'All methods use the same encoder architecture', but SSDD uses a shared encoder trained jointly with an SSDD-M decoder (Section 3.4), whereas the KL-VAE and DiTo baselines use their own encoders. Because the encoder changes the latent distribution, the reconstruction comparison conflates decoder improvements with encoder differences. The frozen-encoder results in Table S3 are helpful, but they are not the main comparison table. Please clarify whether the f8c4 encoders in Table 1 are the same weights or simply the same architecture, and report the main f8c4 comparison with a frozen KL-VAE encoder if possible.
minor comments (4)
  1. [Table 1] The Images/s entries for SSDD-S (1027) and KL-VAE f16c16 (1196) are formatted as '11027' and '11196', which is ambiguous. Please add a space or separator.
  2. [Throughout] The paper alternates between 'DiT' and 'DiTo' when referring to the diffusion decoder of Chen et al. (2025d). Please use a consistent notation to avoid confusion with the DiT generation model.
  3. [Section 4.1] The text says 'the 50k images of the ImageNet evaluation set'; ImageNet-1k has a 50k validation set. Please use the standard terminology.
  4. [Section A.2, Eq. (4)] The notation ∇L = (∇_ν̂ L)(x̂0, x0) is terse. Defining the arguments of the LPIPS gradient explicitly would improve readability, especially since this equation is central to the argument about non-straight velocity fields.

Circularity Check

0 steps flagged

No meaningful circularity: the central claims are benchmarked against external metrics and the design choices are ablated, not derived from their conclusions.

full rationale

The paper's central claims do not reduce to their inputs by construction. Reconstruction quality is measured with rFID, which is an external distribution-level benchmark (Fréchet distance on Inception features), and generation quality with gFID from a separately trained DiT; neither is a term in the training objective. The training loss L = L_FM + 0.5 L_LPIPS + 0.25 L_REPA (Section 3.2) is not equivalent to the reported rFID or gFID results. Distillation is a teacher-student alignment whose success is evaluated externally, not asserted by definition. The paper does not rely on any load-bearing self-citation: the cited prior diffusion-decoder works (Chen et al., Zhao et al., Sargent et al.) are third-party, and no uniqueness theorem from the authors' own prior work is invoked to force a choice. LPIPS does appear both in the loss and in Table 1, which can favor the method, but this is an evaluation-overlap concern rather than an equation-level reduction, and the KL-VAE baselines are also trained with LPIPS; rFID and DreamSim provide independent axes. The paper's own admitted limitation—lower PSNR/SSIM in Section 4.2 ('it yields somewhat worse results in terms of low-level distortion metrics (SSIM, PSNR)')—and the abstract/body rFID discrepancy (0.46 vs 0.50) are framing and correctness issues, not circularity. Therefore no circular step is present.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The design rests on empirically chosen loss weights (lambda_LPIPS=0.5, lambda_REPA=0.25, lambda_KL=1e-6), a chosen operating point (N=8, rho=2, 50k distillation iterations), and a set of domain assumptions—most notably that perceptual/feature losses can replace GANs and that rFID/LPIPS/DreamSim, not PSNR/SSIM, define decode quality. No invented entities. The only mild circularity is LPIPS being present in both the training loss and the evaluation table.

free parameters (6)
  • lambda_LPIPS = 0.5
    Perceptual loss weight (Section 3.2, B). Sets how aggressively the decoder trades pixel distortion for perceptual alignment; tuned by hand.
  • lambda_REPA = 0.25
    Feature-alignment (DINOv2) weight (Section 3.2). Stabilizes transformer representations; ablated as a major rFID contributor (Table 4, -0.44).
  • lambda_KL = 1e-6
    Latent KL weight on the shared encoder (Section B), keeping the latent space close to the standard KL-VAE setup.
  • teacher sampling steps N = 8 (7 in Section B)
    Operating point on the perception-distortion curve (Section 3.3, Fig. S1). Chosen because rFID is lowest at 8 steps; distillation inherits this operating point.
  • scheduler shift rho = 2
    t-spacing shift from Sargent et al. (2025); balances realism vs fidelity at N steps (Section 3.3).
  • distillation budget = 50k iterations
    Fine-tune length for the single-step student (Section B); 5k is used for Fig. S3. Empirically set.
axioms (5)
  • standard math Flow-matching L2 objective on straight paths x_t = (1-t)x + t*eps is a valid generative training objective
    Optimal-transport flow matching from Lipman et al. 2023; used as the main L_FM (Section 3.2).
  • domain assumption Perceptual alignment (LPIPS) and representation alignment (REPA/DINOv2) substitute for adversarial distribution matching
    Core design bet of the paper: GAN-free training needs these auxiliary losses. Ablation (Table 4) supports it empirically, but it is an empirical premise, not a theorem.
  • domain assumption The perception-distortion framework (Blau and Michaeli 2018) licenses de-emphasizing PSNR/SSIM when claiming 'higher reconstruction quality'
    Section 4.2: 'these metrics are not representative... should be de-emphasized.' If a reader requires distortion parity, SSDD is worse than KL-VAE (PSNR 25.12 vs 25.68 at H size).
  • domain assumption Hybrid U-Net + latent transformer (U-ViT variant) is a stronger pixel-space decoder backbone than conv U-Net
    Section 3.1; empirical claim validated vs DiTo baseline in Table 4 (-1.16 rFID) but not proven generally.
  • standard math The LPIPS-regularized flow-matching field is non-straight with a step-count-dependent optimum (derived in Section A.2)
    A.2 derives that L_FM + lambda*L_LPIPS is equivalent to fitting a shifted velocity nu - (lambda*t/2)*grad L; supports the N=8 operating point. The derivation is the paper's own but internally consistent.

pith-pipeline@v1.3.0-alltime-deepseek · 24587 in / 19245 out tokens · 126211 ms · 2026-08-04T11:21:35.319234+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of SSDD: Single-Step Diffusion Decoder for Efficient Image Tokenization." pith.science (2026). https://pith.science/paper/57IXJSIC

@misc{pith2026251004961,
  author       = {Pith},
  title        = {Pith review of: SSDD: Single-Step Diffusion Decoder for Efficient Image Tokenization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/57IXJSIC}},
  note         = {Machine review of arXiv:2510.04961}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Tokenizers are a key component of state-of-the-art generative image models, extracting the most important features from the signal while reducing data dimension and redundancy. Most current tokenizers are based on KL-regularized variational autoencoders (KL-VAE), trained with reconstruction, perceptual and adversarial losses. Diffusion decoders have been proposed as a more principled alternative to model the distribution over images conditioned on the latent. However, matching the performance of KL-VAE still requires adversarial losses, as well as a higher decoding time due to iterative sampling. To address these limitations, we introduce a new pixel diffusion decoder architecture for improved scaling and training stability, benefiting from transformer components and GAN-free training. We use distillation to replicate the performance of the diffusion decoder in an efficient single-step decoder. This makes SSDD the first diffusion decoder optimized for single-step reconstruction trained without adversarial losses, reaching higher reconstruction quality and faster sampling than KL-VAE. In particular, SSDD improves reconstruction FID from $0.87$ to $0.46$ with $1.4\times$ higher throughput and preserve generation quality of DiTs with $3.8\times$ faster sampling. As such, SSDD can be used as a drop-in replacement for KL-VAE, and for building higher-quality and faster generative models.

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

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

    cs.CV 2026-05 unverdicted novelty 6.0

    PiD is a pixel diffusion decoder that performs latent-to-pixel conversion and 4-8x upsampling in one generative step, enabling early stopping of latent diffusion and achieving sub-second 2048x2048 decoding with claime...

Reference graph

Works this paper leans on

24 extracted references · 12 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Cosmos world foundation model platform for physical AI

    Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, Daniel Dworakowski, Jiaojiao Fan, Michele Fenzi, Francesco Ferroni, Sanja Fidler, Dieter Fox, Songwei Ge, Yunhao Ge, Jinwei Gu, Siddharth Gururani, Ethan He, Jiahui Huang, Jacob Huffman, Pooya Jannaty, Jingyi Jin, S...

    Pith/arXiv arXiv

  2. [2]

    Models directly trained at128×128

    8 steps Figure S4 GAN sampling.Evolution of the metrics depending on the number of stepsN. Models directly trained at128×128. Evaluated on ImageNet128×128. Impact of GAN loss on sampling.According to Zhao et al. (2025a), the fact that a low number of sampling steps is optimal forε-V AEresults from thedenoising trajectory matching, referring to their adapt...

    2022

  3. [4]

    Masked autoencoders are effective tokenizers for diffusion models

    Hao Chen, Yujin Han, Fangyi Chen, Xiang Li, Yidong Wang, Jindong Wang, Ze Wang, Zicheng Liu, Difan Zou, and Bhiksha Raj. Masked autoencoders are effective tokenizers for diffusion models. InICML, 2025a. Hao Chen, Ze Wang, Xiang Li, Ximeng Sun, Fangyi Chen, Jiang Liu, Jindong Wang, Bhiksha Raj, Zicheng Liu, and Emad Barsoum. SoftVQ-VAE: Efficient 1-dimensi...

    Pith/arXiv arXiv

  4. [8]

    Mixture-of-transformers: A sparse and scalable architecture for multi-modal foundation models

    Weixin Liang, Lili Yu, Liang Luo, Srini Iyer, Ning Dong, Chunting Zhou, Gargi Ghosh, Mike Lewis, Wen tau Yih, Luke Zettlemoyer, and Xi Victoria Lin. Mixture-of-transformers: A sparse and scalable architecture for multi-modal foundation models. InICLR 2025 Workshop on World Models: Understanding, Modelling and Scaling,

    2025

  5. [9]

    DGAE: Diffusion-guided autoencoder for efficient latent representation learning.arXiv preprint, 2506.09644,

    Dongxu Liu, Yuang Peng, Haomiao Tang, Yuwei Chen, Chunrui Han, Zheng Ge, Daxin Jiang, and Mingxue Liao. DGAE: Diffusion-guided autoencoder for efficient latent representation learning.arXiv preprint, 2506.09644,

    arXiv

  6. [10]

    Knowledge distillation in iterative generative models for improved sampling speed.arXiv preprint, 2101.02388,

    Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed.arXiv preprint, 2101.02388,

    Pith/arXiv arXiv

  7. [12]

    Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du

    Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Sing...

    Pith/arXiv arXiv

  8. [15]

    DiVAE: Photorealistic images synthesis with denoising diffusion decoder.arXiv preprint, 2206.00386,

    Jie Shi, Chenfei Wu, Jian Liang, Xiang Liu, and Nan Duan. DiVAE: Photorealistic images synthesis with denoising diffusion decoder.arXiv preprint, 2206.00386,

    Pith/arXiv arXiv

  9. [16]

    Autoregressive model beats diffusion: Llama for scalable image generation.arXiv preprint, 2406.06525,

    Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation.arXiv preprint, 2406.06525,

    Pith/arXiv arXiv

  10. [17]

    Chameleon: Mixed-modal early-fusion foundation models.arXiv preprint, 2405.09818,

    Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models.arXiv preprint, 2405.09818,

    Pith/arXiv arXiv

  11. [18]

    Latent denoising makes good visual tokenizers

    Jiawei Yang, Tianhong Li, Lijie Fan, Yonglong Tian, and Yue Wang. Latent denoising makes good visual tokenizers. arXiv preprint, 2507.15856,

    arXiv

  12. [19]

    Language model beats diffusion - tokenizer is key to visual generation

    Lijun Yu, Jose Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A Ross, and Lu Jiang. Language model beats diffusion - tokenizer is key to visual generation. InICLR, 2024a. Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, an...

    Pith/arXiv arXiv

  13. [20]

    Z x∈E−1(z) ∆(x,ˆy) # P(D g(z) =y)dydx(2) = min ˆy

    14 A Theoretical motivations A.1 Distortion-distribution shift trade-off As discussed in Section 1, autoencoders have to navigate along the fundamental trade-off between image- to-imagedistortion∆( x,ˆx)that can be minimized by a deterministic decoder, anddistribution shift d(Px, Pˆx), best optimized using a generative decoder. We show here an illustratio...

    2018

  14. [22]

    Impact of sampling steps on reconstruction.Quality of samples from diffusion models usually improves with a higher number of sampling steps

    8 steps Figure S1 Evolution of reconstruction metrics depending on the number of sampling stepsN.Evaluated on ImageNet 256×256. Impact of sampling steps on reconstruction.Quality of samples from diffusion models usually improves with a higher number of sampling steps. But as noted by Zhao et al. (2025a), diffusion decoders have an optimalnumber of samplin...

    2025

  15. [23]

    We show in Table S3 that SSDDoutperforms the original decoders on reconstruction performance, despite being conditioned on features optimized for a different architecture

    SD-V AE 0.69 0.061 0.040 25.22 0.77 SD-V AE-XL 0.64 0.0590.03525.37 0.79 SSDD(8)-M 0.47 0.060 0.046 24.44 0.75 SSDD(8)-L 0.45 0.058 0.044 24.52 0.76 SSDD(8)-H 0.41 0.0560.043 24.67 0.76 DiTo-XL-LPIPS (Chen et al., 2025d) DiTo-XL-LPIPS 0.78 0.102 0.058 24.10 0.71 SSDD(8)-M 0.52 0.062 0.049 24.64 0.76 Reconstruction from existing encoders.To evaluate the ca...

    arXiv

  16. [2006]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. InNeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications,

    2021

  17. [2009]

    We use the following loss coefficients: λLPIPS = 0 .5, λREPA = 0 .25, λKL = 10 −6

    using the RAdamW schedule-free optimizer (Defazio et al., 2024), a weight decay of0.001and no scheduler. We use the following loss coefficients: λLPIPS = 0 .5, λREPA = 0 .25, λKL = 10 −6. We maintain an exponential moving average of the weights with decay rate0.999, starting from 50k iterations. During the first shared pre-training stage, we train on128×1...

    2024

  18. [2018]

    GLU variants improve transformer.arXiv preprint, 2002.05202,

    Noam Shazeer. GLU variants improve transformer.arXiv preprint, 2002.05202,

    Pith/arXiv arXiv 2002

  19. [2020]

    High- fidelity image compression with score-based generative models.CoRR, abs/2305.18231, 2023a

    Emiel Hoogeboom, Eirikur Agustsson, Fabian Mentzer, Luca Versari, George Toderici, and Lucas Theis. High- fidelity image compression with score-based generative models.CoRR, abs/2305.18231, 2023a. Emiel Hoogeboom, Jonathan Heek, and Tim Salimans. Simple diffusion: End-to-end diffusion for high resolution images. InICML, 2023b. Emiel Hoogeboom, Thomas Mens...

    Pith/arXiv arXiv

  20. [2021]

    Latent consistency models: Synthesizing high-resolution images with few-step inference.arXiv preprint, 2310.04378,

    Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference.arXiv preprint, 2310.04378,

    Pith/arXiv arXiv

  21. [2022]

    Flow to the mode: Mode-seeking diffusion autoencoders for state-of-the-art image tokenization.arXiv preprint, 2503.11056,

    Kyle Sargent, Kyle Hsu, Justin Johnson, Li Fei-Fei, and Jiajun Wu. Flow to the mode: Mode-seeking diffusion autoencoders for state-of-the-art image tokenization.arXiv preprint, 2503.11056,

    arXiv

  22. [2023]

    Diffusion autoencoders are scalable image tokenizers.arXiv preprint, 2501.18593, 2025d

    Yinbo Chen, Rohit Girdhar, Xiaolong Wang, Sai Saketh Rambhatla, and Ishan Misra. Diffusion autoencoders are scalable image tokenizers.arXiv preprint, 2501.18593, 2025d. Katherine Crowson, Stefan Andreas Baumann, Alex Birch, Tanishq Mathew Abraham, Daniel Z Kaplan, and Enrico Shippole. Scalable high-resolution pixel-space image synthesis with hourglass dif...

    Pith/arXiv arXiv

  23. [2024]

    Vighnesh Birodkar, Gabriel Barcik, James Lyon, Sergey Ioffe, David Minnen, and Joshua V. Dillon. Sample what you can’t compress.arXiv preprint, 2409.02529,

    arXiv

  24. [2025]

    Imagen 3.arXiv preprint, 2408.07009,

    Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Yichang Chen, SanderDieleman, YuqingDu, ZachEaton-Rosen, HongliangFei, NandodeFreitas, YilinGao, EvgenyGladchenko, Sergio Gómez Colmenarejo, Mandy Guo, Alex Haig, Will Hawkins, Hexiang Hu, Huilian Huang, Tobenna Peter Igwe, Christos Kaplanis, Siavash Khodadadeh, Yel...

    arXiv