Pith. sign in

REVIEW 4 major objections 5 minor 61 references

A post-training pass using Fréchet distance over decoded images as the sole objective cuts FID by 41.4% on average across eight autoregressive image generators, and lifts the best checkpoint from 2.42 to 1.43 FID at no extra inference cost.

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

2026-08-05 00:38 UTC pith:IEAK5LNA

load-bearing objection FD-loss post-training for discrete AR generators is a plausible, well-engineered improvement, but the headline metrics partly optimize the evaluation objective, so the true distributional generalization remains unproven. the 4 major comments →

arxiv 2608.00562 v1 pith:IEAK5LNA submitted 2026-08-01 cs.CV

Beyond Token-Level Cross-Entropy: Fr\'echet Distributional Post-Training for Autoregressive Image Generation

Jinhua Zhang , Yisong Lin , Wei Long , Shuhang Gu This is my paper
classification cs.CV
keywords autoregressive image generationFréchet distance losspost-trainingstraight-through estimatorteacher forcingdistribution matchingclass-conditional ImageNetdiscrete tokenization
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.

Autoregressive image generators are trained token-by-token under ground-truth contexts, yet judged on the distribution of whole decoded images. The paper argues this creates two fixable mismatches: token-level cross-entropy treats every codeword error as equally costly, and teacher forcing never exposes the model to its own inference-time outputs. The proposed remedy is FD-loss post-training, which uses representation-space Fréchet distance as the sole objective and evaluates it on images decoded from model-generated, detached rollout contexts via a probability-level straight-through estimator. Across eight ImageNet 256×256 configurations spanning four generator families, FID falls by 41.4% and FD_r6 by 52.0% on average, with no added parameters or inference steps. If correct, the paper establishes a cheap, general post-training upgrade for pretrained discrete generators that does not touch the tokenizer or the decoding policy.

Core claim

Token-level cross-entropy under teacher forcing treats every codeword error alike and trains on ground-truth histories while inference conditions on the model's own outputs. The paper claims both mismatches are fixed by post-training a pretrained discrete generator with representation-space Fréchet distance as the sole objective. A gradient-free rollout under the native inference policy supplies detached contexts; a probability-level straight-through estimator keeps hard argmax decoding in the forward pass while carrying image-level FD gradients through soft probabilities. Only the generator is updated. Across eight ImageNet 256×256 configurations from four generator families, FID drops 41.4

What carries the argument

The load-bearing object is the probability-level straight-through estimator p_st = p_soft + stopgrad(p_hard − p_soft): the forward pass decodes hard argmax tokens, so the surrogate image is a genuine discrete decoding, while the backward pass differentiates temperature-scaled softmax probabilities, letting image-level Fréchet gradients reach the generator. It is paired with detached rollout-context replay — no-gradient rollouts under the native inference policy are replayed to recompute logits, so optimization happens on the model's own free-running contexts. The Fréchet objective uses EMA-estimated means and covariances in three frozen feature spaces (Inception-v3, MAE, SigLIP) for stable s

Load-bearing premise

The load-bearing premise is that a biased gradient — computed through argmax tokens on self-generated contexts, against only three feature spaces — genuinely improves the model's true sampling distribution, rather than merely overfitting Inception-v3, MAE, and SigLIP statistics that the evaluation metrics partly share.

What would settle it

Measure distributional quality of post-trained checkpoints in a feature space disjoint from Inception-v3, MAE, and SigLIP (e.g., a ViT trained only on non-ImageNet data, or a hand-crafted texture descriptor). If the relative FID-type improvement over the pretrained baseline vanishes there, the reported gains reflect overfitting to the three trained statistics rather than improved sampling. A second check: rerun post-training under several seeds at the short budget and verify the FID/FD_r6 deltas are stable, since all ablations are single runs.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Any pretrained discrete AR generator can be upgraded without changing its architecture, tokenizer, or sampling policy; the gains appear across token-wise (LlamaGen, GigaTok), scale-wise (VAR), and iterative masked (TiTok) families.
  • Teacher-forced replay is not merely weaker but harmful — FID worsens from 5.30 to 14.31 on LlamaGen and from 3.32 to 43.07 on VAR-d16 — so self-generated contexts are the essential ingredient, not the FD loss alone.
  • Feature-space choice trades single-metric performance for cross-representation alignment: Inception-only training reaches FID 3.06 but leaves FD_r6 near baseline, while the three-space default reaches FD_r6 12.30 with FID 4.09.
  • The method is framed as a complement to likelihood pretraining: it starts from cross-entropy-trained checkpoints and improves their decoded-image distribution without replacing token-level learning.

Where Pith is reading between the lines

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

  • The recipe generalizes: any differentiable image-level objective (adversarial discriminators, perceptual losses, human-preference scoring) could be ported to discrete AR generators using the same detached-rollout plus probability-level-STE pattern, targeting whatever quality axis one wants.
  • The jump on VAR-d24 (FID 2.15 to 1.51) suggests part of the historical FID gap between autoregressive and diffusion generators traces to objective/context mismatch rather than capacity — plausible from the cross-family consistency, but not directly tested here.
  • Because rollout trajectories are outside the computation graph and the STE gradient is biased, the approach likely has a ceiling; differentiating through the sampling process itself, or blending in a small likelihood term to guard against drift, are testable extensions the paper leaves open.

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 / 5 minor

Summary. The paper proposes FD-loss post-training for discrete autoregressive image generators. The method freezes the tokenizer, decoder, and feature extractors, and updates only the generator parameters to minimize an EMA-based Fréchet distance in three frozen feature spaces (Inception-v3, MAE, SigLIP). A dual-pass scheme first performs a gradient-free rollout under the model's native inference policy to obtain detached contexts, then recomputes replay logits and uses a probability-level straight-through estimator to backpropagate image-level gradients through hard argmax tokens. Experiments on class-conditional ImageNet at 256×256 cover eight pretrained configurations from LlamaGen, TiTok, VAR, and GigaTok, reporting average FID and FDr6 reductions of 41.4% and 52.0%, respectively, with the best FID improving from 2.42 to 1.43. Ablations examine replay-context source, replay temperature, feature-space composition, and EMA decay. The paper is transparent about several limitations, including the aggregate nature of FDr6 and the bias of the STE.

Significance. If the reported improvements generalize beyond the optimized feature spaces, this is a meaningful contribution: it offers a simple, inference-cost-free post-training complement to token-level cross-entropy, demonstrates applicability across four discrete AR families, and provides code. The detached-rollout-plus-STE mechanism is a practical and clearly described way to bridge the discrete non-differentiability barrier. The paired experimental design and the authors' explicit statement of the objective-overlap limitation are strengths. However, the headline FID is exactly the Inception-v3 Fréchet term in the training objective, and FDr6 only partially overlaps the training objective; the manuscript does not report a held-out-only FD decomposition, so the central claim of improved distributional quality beyond the training objective is not yet established.

major comments (4)
  1. [Experiments / Eq. (12), Table 1] The headline FID is computed in Inception-v3 feature space, which is one of the three FD terms minimized in Eq. (12). FDr6 is an aggregate of six normalized FD ratios, of which three (Inception-v3, MAE, SigLIP) are optimized and three (ConvNeXt-v2, DINOv2, CLIP) are held out. Only the aggregate FDr6 is reported in Table 1. Table 4 provides direct evidence that the aggregate can move mainly through the optimized spaces: Inception-only training gives the best FID (3.06) but leaves FDr6 essentially unchanged (16.97 vs. pretrained 16.76). The paper itself concedes this in the 'Feature-space composition' paragraph. To support the claim that FD-loss post-training improves true distributional quality, the authors should report per-space FD ratios, especially the held-out-only component, for at least the main configurations in Table 1. Without that, the central empirical claim is not established
  2. [Method, Eqs. (5)–(8), Algorithm 1] The gradient is computed on a hard replay surrogate that 'need not equal the token sampled at the corresponding rollout position' (text after Eq. (7)), and no gradient flows through the sampled rollout trajectory. The paper acknowledges the bias but provides no quantitative measure of the mismatch between the sampled rollout and the replay argmax tokens. If the two diverge substantially, optimizing the surrogate may not optimize the actual sampling distribution that is evaluated at test time. Please report, for example, the token-level agreement/overlap between sampled rollout tokens and replay argmax tokens during post-training, or otherwise bound the surrogate bias. This is load-bearing for the method's validity.
  3. [Experiments / Post-training details vs. Table S7] The main text states: 'For each generator, we retain the optimizer, initial learning rate, learning-rate schedule, and associated optimization hyperparameters of the corresponding baseline implementation without modification.' However, Table S7 reports a single common post-training setting for all models (AdamW with betas (.9,.95), initial LR 1e-6, 1k warm-up + cosine), which conflicts with the baseline settings in Table S6 (e.g., LlamaGen constant LR 1e-4, VAR linear schedules with different base LRs). Please clarify which setting was actually used. This inconsistency is important for reproducibility.
  4. [Table 1 and Ablation protocol] All main results appear to be single-run evaluations without seeds or confidence intervals; the paper only characterizes the ablations as single-run short-budget runs. Given the stochastic post-training procedure and the many free hyperparameters (τ, β, feature coefficients, LR, batch, iterations), the reported average reductions need at least a small number of seeds or an explicit statement of computational constraints. This is especially relevant because the metric being optimized is also the headline evaluation metric.
minor comments (5)
  1. [Eq. (8) and surrounding text] The typesetting of the codebook embedding is corrupted: 'eet = (pst_t)^ op C, ex=D( ee1, . . . ,eeT )' should likely be \tilde e_t and \tilde x. Please fix.
  2. [FDr6 definition] The paper says FDr6 'averages normalized FD ratios' but does not state the normalization procedure in the main text. Since the 52.0% average reduction is a headline number, the normalization (e.g., per-space division by a reference FD) should be specified.
  3. [Table 1 header] The header reads 'Bold indicates the betterFDr6, FID, and IS values within each AR pair.' A missing space and comma make it awkward; also clarify whether Prec./Rec. are computed on the same 50k samples.
  4. [Eq. (1) and iterative generators] Equation (1) presents a strict factorization p_θ(y|c) = ∏_s p_θ(y^(s)|y^(<s),c). For TiTok and GigaTok, described as iterative masked generators, this factorization is not the native generation model. Explain how the method and equations extend to those cases, since Algorithm 1 is written generically.
  5. [Algorithm 1, step 2] The initialization step says 'Generate an initialization set with p_θ0 under P_base.' The size of this initialization set and how it is used to set EMA moments are not specified. Please add the value.

Circularity Check

2 steps flagged

Headline FID and FD_r6 gains are partly the training objective itself; held-out evidence is only partially separated.

specific steps
  1. fitted input called prediction [Abstract; Eq. (4); Eq. (12); Experimental Setup (Datasets and metrics)]
    "Across eight completed configurations ... FD-loss post-training reduces FID and FD_r6 by 41.4% and 52.0% on average. ... For the Inception-v3 representation used by the standard evaluation protocol, Eq. (4) corresponds to FID (Heusel et al. 2017). ... LFD = 1/3 sum_{m in M} FD_{phi_m}(r,g)."

    The headline FID is the Inception-v3 instance of Eq. (4), and Eq. (12) minimizes exactly that term (together with MAE and SigLIP). Thus the average 41.4% FID reduction is a direct result of optimizing the reported metric, not an independent test of distributional quality. The same overlap affects three of six FD_r6 components. Because the paper reports only the aggregate FD_r6, the reader cannot separate the forced component from any held-out generalization.

  2. self definitional [Experimental Setup (Datasets and metrics); Feature-space composition ablation (Table 4)]
    "Inception-v3, MAE, and SigLIP define the training objective; ConvNeXt-v2, DINOv2, and CLIP are held out from optimization. Because FD_r6 mixes optimized and held-out spaces, we interpret it as an aggregate alignment metric rather than a fully held-out measure. ... Inception-only training achieves the lowest FID (3.06) ... but leaves FD_r6 nearly unchanged relative to the pretrained baseline (16.97 versus 16.76)."

    FD_r6 is defined as an average over six normalized FD ratios, three of which are the spaces in the training objective. Therefore an FD_r6 improvement is partly guaranteed by the definition of the metric. Table 4 confirms this: Inception-only training barely changes FD_r6, while adding MAE and SigLIP lowers it to 12.30, showing the aggregate gain is driven by the optimized spaces rather than by the held-out spaces, whose separate values are never reported.

full rationale

The paper's central method is to minimize representation-space Fréchet distances in Inception-v3, MAE, and SigLIP. The headline FID metric is exactly the Inception-v3 FD term of that objective, and FD_r6 includes all three training spaces in its definition; hence part of the claimed improvement is the value of the training objective itself, not an independent discovery. However, the paper is unusually transparent: it states that FD_r6 'mixes optimized and held-out spaces' and concludes that the evidence is bounded by 'an aggregate FD_r6 metric that partly overlaps the training objective.' It also reports held-out metrics (IS, precision, recall) that improve, and the FD_r6 aggregate does include three non-optimized spaces, so the central claim has independent content and is not wholly forced. No load-bearing self-citation was found: the only overlapping self-citations (Zhang et al. 2026 MVAR; possibly Lin et al. 2026b GEAR) are related-work mentions, not used to justify the method or forbid alternatives. The main circularity is partial and metric-level rather than a derivation that reduces to its inputs. Score 6 reflects that one or more headline 'predictions' (FID and partly FD_r6) reduce to the optimized objective by construction, while the held-out evidence mitigates, but is not separately reported for the held-out feature spaces.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The method introduces no new entities, only a training procedure. The central claim relies on the feature-space FD proxy, the STE approximation, the rollout policy match, and the public checkpoints, all explicitly stated. The main free parameters are tau, beta, and feature-space weights, each selected via small ablations.

free parameters (5)
  • replay temperature tau = 1.0
    Chosen by hand and justified via ablation in Table 3; controls the softness of the backward surrogate but not forward tokens.
  • EMA decay beta = 0.999
    Chosen by hand and justified via ablation in Table S5; balances responsiveness and temporal smoothing of feature moments.
  • feature-space coefficients = equal weights 1/3 each
    Chosen by hand; Table 4 shows the multi-space objective improves FD_r6 but slightly worsens FID compared to Inception-only.
  • post-training learning rate = 1e-6
    Optimization hyperparameter from the FD-loss protocol; not physically fitted.
  • post-training batch size and iterations = 16 / 100,000 (main); 32 / 3,000 (ablations)
    Chosen for compute; ablations use a shorter budget and are explicitly local sensitivity runs.
axioms (5)
  • domain assumption FD distance with EMA-estimated moments is a valid, differentiable proxy for distributional quality.
    Used in Eq. (11)-(13); the entire method depends on this proxy being a useful training signal.
  • domain assumption Frozen tokenizer, decoder, and feature extractors are differentiable with respect to their inputs, letting gradients reach the generator.
    Stated in the post-training protocol; if any frozen module has non-differentiable operations, the gradient path breaks.
  • domain assumption The probability-level STE (Eq. 8) provides a useful, sufficiently unbiased gradient direction for discrete token selection.
    The paper acknowledges the bias of the STE and relies on it for all updates.
  • domain assumption The fixed base inference policy P_base used for rollout matches the evaluation policy.
    Eq. (5) and the experimental setup; if rollout and evaluation policies diverge, post-training would not align with inference.
  • domain assumption The public pretrained checkpoints are used as-is, and baselines are evaluated under the same unified pipeline.
    Experimental setup; any discrepancy in checkpoint handling or inference settings could bias the paired comparison.

pith-pipeline@v1.3.0-alltime-deepseek · 16368 in / 10043 out tokens · 112093 ms · 2026-08-05T00:38:14.221224+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Beyond Token-Level Cross-Entropy: Fr\'echet Distributional Post-Training for Autoregressive Image Generation." pith.science (2026). https://pith.science/paper/IEAK5LNA

@misc{pith2026260800562,
  author       = {Pith},
  title        = {Pith review of: Beyond Token-Level Cross-Entropy: Fr\'echet Distributional Post-Training for Autoregressive Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IEAK5LNA}},
  note         = {Machine review of arXiv:2608.00562}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Autoregressive image generators are commonly pretrained with token-level cross-entropy under teacher forcing, yet evaluated by the distributional quality of decoded images. This creates an objective mismatch, because categorical errors have unequal image-level consequences, and a context mismatch, because inference conditions on model-generated histories. We introduce FD-loss post-training, which adapts a pretrained discrete generator using representation-space Fr\'echet distance as the sole objective. A dual-pass scheme first constructs detached rollout contexts through gradient-free generation under the model's native inference configuration, then performs differentiable replay with a probability-level straight-through estimator (STE) that preserves hard argmax decoding in the forward pass while propagating image-level gradients through temperature-scaled probabilities. Only the generator is updated, while the tokenizer and feature extractors remain frozen. Across eight completed configurations from four generator families on class-conditional ImageNet at $256\times256$, FD-loss post-training reduces FID and $\mathrm{FD}_{r6}$ by 41.4% and 52.0% on average. The strongest FID result improves from 2.42 to 1.43 without adding parameters or inference steps.

Figures

Figures reproduced from arXiv: 2608.00562 by Jinhua Zhang, Shuhang Gu, Wei Long, Yisong Lin.

Figure 1
Figure 1. Figure 1: Teacher-forced token pretraining versus our dual [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our FD-loss post-training framework. The dual-pass scheme first performs a gradient-free rollout under the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of pretrained generators and their FD-loss post-trained counterparts on class-conditional [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

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

61 extracted references · 19 canonical work pages · 3 internal anchors

  1. [1]

    Advances in Neural Information Processing Systems , volume=

    Neural discrete representation learning , author=. Advances in Neural Information Processing Systems , volume=

  2. [2]

    Advances in neural information processing systems , volume=

    Gans trained by a two time-scale update rule converge to a local nash equilibrium , author=. Advances in neural information processing systems , volume=

  3. [3]

    arXiv preprint arXiv:2010.11929 , year=

    An image is worth 16x16 words: Transformers for image recognition at scale , author=. arXiv preprint arXiv:2010.11929 , year=

    Pith/arXiv arXiv 2010

  4. [4]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Taming transformers for high-resolution image synthesis , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  5. [5]

    arXiv preprint arXiv:2110.04627 , year=

    Vector-quantized image modeling with improved vqgan , author=. arXiv preprint arXiv:2110.04627 , year=

    Pith/arXiv arXiv

  6. [6]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Autoregressive image generation using residual quantization , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  7. [7]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Maskgit: Masked generative image transformer , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  8. [8]

    arXiv preprint arXiv:2207.12598 , year=

    Classifier-free diffusion guidance , author=. arXiv preprint arXiv:2207.12598 , year=

    Pith/arXiv arXiv

  9. [9]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Magvit: Masked generative video transformer , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  10. [10]

    Autoregressive Model Beats Diffusion:

    Sun, Peize and Jiang, Yi and Chen, Shoufa and Zhang, Shilong and Peng, Bingyue and Luo, Ping and Yuan, Zehuan , journal=. Autoregressive Model Beats Diffusion:

    work page

  11. [11]

    arXiv preprint arXiv:2406.11838 , year=

    Autoregressive Image Generation without Vector Quantization , author=. arXiv preprint arXiv:2406.11838 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2409.16211 , year=

    Maskbit: Embedding-free image generation via bit tokens , author=. arXiv preprint arXiv:2409.16211 , year=

    Pith/arXiv arXiv

  13. [13]

    Visual autoregressive modeling:

    Tian, Keyu and Jiang, Yi and Yuan, Zehuan and Peng, Bingyue and Wang, Liwei , journal=. Visual autoregressive modeling:

    work page

  14. [14]

    arXiv preprint arXiv:2412.15205 , year=

    Flowar: Scale-wise autoregressive image generation meets flow matching , author=. arXiv preprint arXiv:2412.15205 , year=

    Pith/arXiv arXiv

  15. [15]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Beyond next-token: Next-x prediction for autoregressive visual generation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    work page

  16. [16]

    International Conference on Learning Representations , volume=

    Fluid: Scaling autoregressive text-to-image generative models with continuous tokens , author=. International Conference on Learning Representations , volume=

    work page

  17. [17]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Randomized autoregressive visual generation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  18. [18]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Randar: Decoder-only autoregressive visual generation in random orders , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    work page

  19. [19]

    arXiv preprint arXiv:2503.07076 , year=

    NFIG: Autoregressive Image Generation with Next-Frequency Prediction , author=. arXiv preprint arXiv:2503.07076 , year=

    arXiv

  20. [20]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Next patch prediction for autoregressive visual generation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    work page

  21. [21]

    arXiv preprint arXiv:2604.28190 , year=

    Representation Fréchet Loss for Visual Generation , author=. arXiv preprint arXiv:2604.28190 , year=

    Pith/arXiv arXiv

  22. [22]

    Advances in Neural Information Processing Systems , volume=

    Self forcing: Bridging the train-test gap in autoregressive video diffusion , author=. Advances in Neural Information Processing Systems , volume=

    work page

  23. [23]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Scalable image tokenization with index backpropagation quantization , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    work page

  24. [24]

    arXiv preprint arXiv:1903.05662 , year=

    Understanding straight-through estimator in training activation quantized neural nets , author=. arXiv preprint arXiv:1903.05662 , year=

    Pith/arXiv arXiv 1903

  25. [25]

    arXiv preprint arXiv:2411.18704 , year=

    Exponential moving average of weights in deep learning: Dynamics and benefits , author=. arXiv preprint arXiv:2411.18704 , year=

    Pith/arXiv arXiv

  26. [26]

    Advances in Neural Information Processing Systems , volume=

    An image is worth 32 tokens for reconstruction and generation , author=. Advances in Neural Information Processing Systems , volume=

  27. [27]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Gigatok: Scaling visual tokenizers to 3 billion parameters for autoregressive image generation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    work page

  28. [28]

    arXiv preprint arXiv:2304.07193 , year=

    Dinov2: Learning robust visual features without supervision , author=. arXiv preprint arXiv:2304.07193 , year=

    Pith/arXiv arXiv

  29. [29]

    Advances in neural information processing systems , volume=

    Scheduled sampling for sequence prediction with recurrent neural networks , author=. Advances in neural information processing systems , volume=

  30. [30]

    Advances in neural information processing systems , volume=

    Professor forcing: A new algorithm for training recurrent networks , author=. Advances in neural information processing systems , volume=

  31. [31]

    European conference on computer vision , pages=

    Incorporating reinforced adversarial learning in autoregressive image generation , author=. European conference on computer vision , pages=. 2020 , organization=

    work page 2020

  32. [32]

    Liao, Xinyao and He, Qiyuan and Xu, Kai and Qu, Xiaoye and Li, Yicong and Wei, Wei and Yao, Angela , journal=

    work page

  33. [33]

    International conference on learning representations , year=

    Categorical reparameterization with gumbel-softmax , author=. International conference on learning representations , year=

    work page

  34. [34]

    arXiv preprint arXiv:2605.00503 , year=

    End-to-end autoregressive image generation with 1d semantic tokenizer , author=. arXiv preprint arXiv:2605.00503 , year=

    Pith/arXiv arXiv

  35. [35]

    GEAR: Guided End-to-End AutoRegression for Image Synthesis

    GEAR: Guided End-to-End AutoRegression for Image Synthesis , author=. arXiv preprint arXiv:2606.32039 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  36. [36]

    RankE: End-to-End Post-Training for Discrete Text-to-Image Generation with Decoder Co-Evolution

    RankE: End-to-End Post-Training for Discrete Text-to-Image Generation with Decoder Co-Evolution , author=. arXiv preprint arXiv:2605.21195 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  37. [37]

    arXiv preprint arXiv:2003.11774 , year=

    Image generation via minimizing fr 'echet distance in discriminator feature space , author=. arXiv preprint arXiv:2003.11774 , year=

    Pith/arXiv arXiv 2003

  38. [38]

    arXiv preprint arXiv:2009.14075 , year=

    Backpropagating through Fr 'echet Inception Distance , author=. arXiv preprint arXiv:2009.14075 , year=

    Pith/arXiv arXiv 2009

  39. [39]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Rethinking the inception architecture for computer vision , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  40. [40]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Masked autoencoders are scalable vision learners , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  41. [41]

    arXiv preprint arXiv:2502.14786 , year=

    Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features , author=. arXiv preprint arXiv:2502.14786 , year=

    Pith/arXiv arXiv

  42. [42]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Convnext v2: Co-designing and scaling convnets with masked autoencoders , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  43. [43]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  44. [44]

    arXiv preprint arXiv:2310.05737 , year=

    Language Model Beats Diffusion--Tokenizer is Key to Visual Generation , author=. arXiv preprint arXiv:2310.05737 , year=

    Pith/arXiv arXiv

  45. [45]

    2026 , url=

    Jinhua Zhang and Wei Long and Minghao Han and Weiyi You and Shuhang Gu , booktitle=. 2026 , url=

    work page 2026

  46. [46]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Spectralar: Spectral autoregressive visual generation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    work page

  47. [47]

    Generative Refinement Networks for Visual Synthesis

    Generative Refinement Networks for Visual Synthesis , author=. arXiv preprint arXiv:2604.13030 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  48. [48]

    arXiv preprint arXiv:2602.09024 , year=

    Autoregressive image generation with masked bit modeling , author=. arXiv preprint arXiv:2602.09024 , year=

    arXiv

  49. [49]

    arXiv preprint arXiv:2601.17124 , year=

    iFSQ: Improving FSQ for Image Generation with 1 Line of Code , author=. arXiv preprint arXiv:2601.17124 , year=

    arXiv

  50. [50]

    arXiv preprint arXiv:2508.12811 , year=

    Next Visual Granularity Generation , author=. arXiv preprint arXiv:2508.12811 , year=

    work page arXiv

  51. [51]

    arXiv preprint arXiv:2506.05289 , year=

    Towards Sequence Modeling Alignment between Tokenizer and Autoregressive Model , author=. arXiv preprint arXiv:2506.05289 , year=

    arXiv

  52. [52]

    arXiv preprint arXiv:2503.10568 , year=

    Autoregressive image generation with randomized parallel decoding , author=. arXiv preprint arXiv:2503.10568 , year=

    arXiv

  53. [53]

    arXiv preprint arXiv:2601.01535 , year=

    Improving Flexible Image Tokenizers for Autoregressive Image Generation , author=. arXiv preprint arXiv:2601.01535 , year=

    arXiv

  54. [54]

    arXiv preprint arXiv:2603.10744 , year=

    Just-in-time: Training-free spatial acceleration for diffusion Transformers , author=. arXiv preprint arXiv:2603.10744 , year=

    arXiv

  55. [55]

    arXiv preprint arXiv:2602.04770 , year=

    Generative Modeling via Drifting , author=. arXiv preprint arXiv:2602.04770 , year=

    Pith/arXiv arXiv

  56. [56]

    arXiv preprint arXiv:2512.02012 , year=

    Improved mean flows: On the challenges of fastforward generative models , author=. arXiv preprint arXiv:2512.02012 , year=

    Pith/arXiv arXiv

  57. [57]

    arXiv preprint arXiv:2601.22158 , year=

    One-step Latent-free Image Generation with Pixel Mean Flows , author=. arXiv preprint arXiv:2601.22158 , year=

    Pith/arXiv arXiv

  58. [58]

    arXiv preprint arXiv:2507.23268 , year=

    Pixnerd: Pixel neural field diffusion , author=. arXiv preprint arXiv:2507.23268 , year=

    Pith/arXiv arXiv

  59. [59]

    European Conference on Computer Vision , pages=

    Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  60. [60]

    Advances in Neural Information Processing Systems , volume=

    Representation entanglement for generation: Training diffusion transformers is much easier than you think , author=. Advances in Neural Information Processing Systems , volume=

    work page

  61. [61]

    arXiv preprint arXiv:2507.15856 , year=

    Latent denoising makes good visual tokenizers , author=. arXiv preprint arXiv:2507.15856 , year=

    arXiv