Pith. sign in

REVIEW 3 major objections 6 minor 18 references

Efficient Generative Modeling with Residual Vector Quantization-Based Tokens

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read ResGen predicts the cumulative embedding of masked RVQ tokens rather than individual tokens, so inference steps no longer scale with quantization depth, reaching FID 1.93 on ImageNet and beating autoregressive TTS baselines.

desk verdict ResGen's core idea—predicting cumulative RVQ embeddings to decouple sampling from depth—is genuinely useful and backed by controlled ablations, but the ELBO derivation in Section 3.2 overreaches. read the letter →

arxiv 2412.10208 v3 pith:7HHJ2SZT submitted 2024-12-13 cs.LG

Jaehyeon Kim , Taehong Moon , Keon Lee , Jaewoong Cho This is my paper · ORCID
classification cs.LG
keywords residualvectorquantizationmaskedtokenpredictiondiscretediffusioncumulativeembeddinggenerativemodelingimagegenerationtext-to-speechsynthesismixtureofGaussians
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

ResGen is a generative model for residual vector quantization (RVQ) tokens that sets out to make sampling cost independent of both sequence length and quantization depth. RVQ packs fidelity into a token grid by quantizing residuals at several depths, but autoregressive models pay a sampling step for every depth at every position. ResGen instead trains the model to predict the cumulative vector embedding — the sum of the masked embeddings across depths — and then quantizes that vector back into tokens. On ImageNet 256x256 it reports FID 1.93 with classifier-free guidance in 63 steps, and in zero-shot text-to-speech it beats autoregressive baselines on word and character error rates with far fewer inference steps. The point is that adding RVQ depth for fidelity no longer has to cost sampling speed.

What carries the argument

The central object is the cumulative embedding prediction: instead of classifying each masked RVQ token separately, the model outputs a mixture of Gaussians over the sum of the masked depth embeddings at each position, $z_i = \sum_j e(x_{i,j}; j) \odot (1 - m_{i,j})$, with a modified training objective that forces every mixture component to be used. At sampling time the predicted vector is quantized back into tokens with the RVQ codebook, and the most confident tokens are unmasked first, following a schedule in which deeper (finer) tokens are masked first and filled in last. This machinery carries the argument by collapsing the depth axis into a single regression target, so the number of generation steps depends only on the masking schedule and not on sequence length or RVQ depth.

What would settle it

Compute the dropped ELBO terms on held-out latents: estimate $-\log p(x^{(0)}|z,x^{(t)})$ for the trained RVQ quantizer and the KL divergence between the quantization posterior and the dequantization distribution, and check whether their sum is small relative to $-\log p_\theta(z|x^{(t)})$; if it is large, the Section 3.2 simplification fails. A cheaper decisive test is to train the same architecture with the cumulative-embedding loss but without the hypergeometric diffusion masking, using a fixed uniform mask ratio instead; identical FID would show the probabilistic framing is not what drives the result.

Watch

Extended reading notes

Core claim

The paper's central claim is that predicting the summed vector embedding of the masked tokens at a position, $z_i = \sum_j e(x_{i,j}; j) \odot (1 - m_{i,j})$, is both sufficient and better than predicting individual tokens, because the RVQ decoder consumes vector embeddings, not token indices. This design removes the dependence of inference steps on RVQ depth: one forward pass predicts all depths at once, and re-quantizing the predicted embedding into tokens adds negligible overhead. The authors cast masking and unmasking as a discrete diffusion process with a multivariate hypergeometric forward step, deriving the training objective $-\log p_\theta(z|x^{(t)})$ as the surviving term of a variational bound after assuming the reconstruction and entropy terms cancel. Empirically the claim is supported by a controlled ablation in which the cumulative-embedding variant beats direct token prediction at equal capacity (FID 2.43 versus 2.91 with CFG on ImageNet under the same RVQ tokens), by an autoregressive-depth ablation that isolates the depth-prediction gain (FID 5.22 in 3.05 seconds versus 5.50 in 5.35 seconds for RQ-Transformer), and by scaling results showing that deeper RVQ improves generation quality at the same step count.

Load-bearing premise

The load-bearing premise is that the terms dropped from the evidence lower bound in Section 3.2 — the reconstruction likelihood of the original tokens and the entropy of the quantization-versus-dequantization distributions — are negligible or cancel; if they do not, the training loss is not the claimed variational bound, even though it may still work empirically.

Editorial extensions

If this is right

  • RVQ depth can be scaled for higher fidelity without increasing sampling steps: 16-depth beats 8-depth at the same step count, and the audio model runs 25 steps with 72 quantization levels.
  • In the controlled comparison under identical RVQ tokens, ResGen reaches FID 2.43 with CFG in 63 steps where RQ-Transformer needs 1024 steps for FID 5.50; the depth-only ablation is 1.8x faster at better FID.
  • For zero-shot text-to-speech, ResGen posts the lowest word and character error rates on the cross-sentence task among the compared models (WER 1.70, CER 0.46) with 25 inference steps.
  • Standard diffusion tooling transfers to the framework: classifier-free guidance, temperature, and top-p each tune the quality-diversity balance of the unmasking process, as the paper demonstrates in its sampling ablations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 'predict the cumulative embedding, then re-quantize' loop should transfer to any hierarchical codec whose decoder consumes additive vectors — multi-group bit tokens in LFQ-style models, audio codec language models, or video tokenizers — none of which the paper tests.
  • Because step count and representation depth are now independent, an adaptive-compute regime becomes possible that the paper does not pursue: draft with few steps and refine with more, choosing the budget per application.
  • A non-variational version of the same loss — cumulative-embedding regression without the diffusion-masking story — is a testable extension that would reveal whether the probabilistic framing is load-bearing; the paper itself notes there is no theory for why so few steps suffice.
  • Combining progressive unmasking with key-value caching, which the paper lists as future work, would make per-step cost shrink as generation proceeds — an efficiency gain orthogonal to the depth decoupling.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes ResGen, a masked generative model over Residual Vector Quantization (RVQ) tokens. Instead of predicting individual token indices, the model predicts the sum of the embeddings of the masked tokens at each sequence position, modeled with a mixture of Gaussians. Masking is applied from the highest RVQ depth downward, and sampling iteratively unmasks tokens using confidence scores. The paper claims that this design decouples the number of inference steps from both RVQ depth and sequence length. It reports a controlled ablation on ImageNet 256x256 with matched RVQ tokens, parameter counts, and training budget against RQ-Transformer and a MaskGIT variant, alongside scaling ablations and benchmark comparisons for image generation and zero-shot text-to-speech synthesis.

Significance. If the empirical results hold, the cumulative-embedding prediction idea is a simple and practically useful contribution: it avoids autoregressive decoding over RVQ depth and shortens the sampling trajectory. The controlled ablation in Table 1 with matched tokens, parameters, and training budget is a strong piece of evidence that the design, rather than model scale or tokenizer choice, drives the gains. The paper also provides useful ablations of sampling hyperparameters and an honest limitations section. However, the theoretical framing in Section 3.2 currently overclaims, and the selective reporting in the AR-ResGen ablation needs to be addressed.

major comments (3)
  1. [Section 3.2] The variational derivation drops the reconstruction term -log p(x^(0)|z,x^(t)) and the normalization term -log q(z|x^(0),x^(t)) by asserting that p is RVQ quantization and q is RVQ dequantization. This is only valid if the deterministic mapping z_i = sum_j e(x_{i,j};j)(1-m_{i,j}) is inverted exactly by the RVQ quantization used in Algorithm 2. For partial masks, z_i contains only the masked (higher-depth) embeddings, so greedy residual quantization of z_i cannot recover unmasked lower-depth tokens from the sum alone; even for fully masked sequences, residual quantization is not guaranteed to be the inverse of embedding summation. If RVQ(z_i) disagrees with the training target x_{i,·}, the dropped term is not a constant and Eq. (2) is not an ELBO for p_theta(x^(0)|x^(t)). Please either prove the inverse property, verify it empirically on the actual RVQ tokenizer, and specify how unmasked depths are handled, or explicitly present Eq. (1) as a surrogate latent-regression objective rather than as a variational bound.
  2. [Section 3.2, Eq. (2)] The reduction of the ELBO to L_simple(x^(0);theta) = -log p_theta(x^(0)|x^(t)) is not a derivation: the ELBO contains a sum over diffusion steps and a reconstruction term, and replacing them with a single cross-entropy at an unspecified time t under 'equal emphasis' is a heuristic simplification. The connection from the ELBO to Eq. (1) therefore involves two unstated approximations: ELBO to L_simple, and L_simple to Lmask. Please state explicitly which parts are exact bounds and which are design choices, and give the distribution of the training time t used in Eq. (2).
  3. [Table 1 (right panel)] The right-hand side of Table 1 shows that without classifier-free guidance, AR-ResGen is substantially worse than RQ-Transformer at every step budget (e.g., FID 23.48 versus 15.71 at 8 steps), while with CFG it is comparable or better. The text's claim that the depth-prediction strategy 'delivers substantial improvements in both the generation speed and the sample quality' is therefore only supported in the CFG evaluation. Please report and discuss the w/o-CFG numbers in the main text and scope the 'comparable quality' claim accordingly.
minor comments (6)
  1. [Section 3.2] The formulas for the marginal q(x^(t)|x^(0)) and the conditional q(x^(t)|x^(t+1),x^(0)) are hard to parse because the binomial coefficients appear to contain the random variables k_i^(tau) on both sides; please rewrite them with standard multivariate hypergeometric notation and explicitly define n^(tau).
  2. [Section 3.1 and Algorithm 1] The text says the counts k_i are 'sampled without replacement from a multinomial distribution'; this should be a multivariate hypergeometric distribution, and the sampling procedure in Algorithm 2 should be specified more precisely.
  3. [Appendix B.1] The confidence score formula log p(x_{i,j}|z_i) proportional to a Gaussian log-density omits the proportionality constant and is a heuristic posterior over tokens; please state this explicitly so readers do not confuse it with the model's actual predictive distribution.
  4. [Table 1] In the right panel, 'FID of 5.33s' should read 'FID of 5.33' (the unit 's' appears to be a typo).
  5. [Tables 2 and 3] The benchmark comparisons rely largely on numbers reported in the original papers, which may use different training data, tokenizers, and evaluation protocols; the paper should state more prominently that the controlled ablation in Table 1 is the only directly comparable evidence for the claims of superiority.
  6. [Section 5.2.3] The sentence 'deeper RVQ depth enhances leads to improved reconstruction quality' contains a grammatical error and should be reworded.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central cumulative-embedding prediction is validated by controlled ablations and external benchmarks; only minor self-citations and an admitted heuristic variational derivation appear.

full rationale

The central claim—predicting cumulative RVQ embeddings so inference steps are independent of sequence length and depth—is not extracted from the evaluation metric. The training loss in Eq. 1 is defined directly on the cumulative embedding z_i = sum_j e(x_i,j;j)*(1-m_i,j), and the controlled ablation in Table 1 (ResGen vs. direct token prediction, RQ-Transformer, MaskGIT) plus external ImageNet FID and TTS WER/CER benchmarks provide independent evidence. Section 3.2's probabilistic derivation is heuristic rather than circular: it drops -log p(x(0)|z,x(t)) and the q-denominator under the assumption that p and q are RVQ quantization/dequantization, but this simplification is not guaranteed to recover the masked tokens; that is a derivation weakness, not a reduction of the loss to its own input. Self-citations (CLaM-TTS for mixture-of-Gaussians vector prediction and low-rank projection; DiTTo-TTS as a baseline) are building blocks or baselines, not load-bearing support for the central claim. The Limitations section (Appendix D) explicitly concedes that no theoretical justification is offered for the low number of inference steps, and the variational simplification in Section 3.2 is likewise an unproven assumption rather than a circular step. No fitted parameter is renamed as a prediction, and no uniqueness theorem from the authors is invoked. The paper is therefore not circular in any load-bearing sense.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The ledger's dominant items are tuned sampling hyperparameters (temperature, top-p, CFG schedule, step count) and architectural choices (K=1024 mixture components). These are the main degrees of freedom behind the reported numbers. The variational derivation rests on an unproven simplification, flagged as ad hoc.

free parameters (6)
  • masking schedule gamma(r) = gamma(r) = sqrt(1 - r^2)
    Chosen by hand and used for all training and sampling; Appendix C.3 shows FID depends on the schedule.
  • choice temperature tau = 28.0
    Selected via configuration search (Figure 4c) to optimize FID with CFG.
  • top-p nucleus threshold = 0.94, 0.96, or 0.98 depending on sampling steps
    Tuned per configuration in Table 4; Figure 4b shows FID varies with top-p.
  • CFG schedule = linear from 0.02 to 2.4 (or 2.2)
    Tuned per configuration in Table 4.
  • number of sampling steps = 28, 48, 64, or 100
    Ablated in Figure 4a; more steps improve FID, so the 'fast' claim is tied to a step budget.
  • number of Gaussian mixture components K = 1024
    Chosen so output projection cost is comparable to a ~64K softmax (Appendix A.2); not swept.
assumptions (4)
  • domain assumption Higher RVQ depths encode finer details, so masking from the highest depth down is a valid coarse-to-fine corruption.
    Invoked in Section 3.1 to define the masking strategy; assumed to hold for both ImageNet and speech.
  • domain assumption The VAE decoder depends only on the sum z_i = sum_j e(x_i,j;j), so predicting z_i is sufficient for generation.
    True by RVQ construction, but the claim that modeling z suffices for high-quality generation is an assumption discussed in Section 3.1.
  • domain assumption A mixture of Gaussians with identity covariance and low-rank projections can represent the distribution of cumulative embeddings accurately enough for RVQ quantization to recover good tokens.
    The training objective uses this model; no theoretical guarantee is given in Appendix A.2.
  • ad hoc to paper The variational bound simplifies to Lmask by dropping log p(x^(0)|z,x^(t)) and normalization terms.
    Section 3.2 states 'Assuming... we focus on the remaining term' without proof; if false, the probabilistic grounding of the loss is incomplete.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Generative Modeling with Residual Vector Quantization-Based Tokens." pith.science (2026). https://pith.science/paper/7HHJ2SZT

@misc{pith2026241210208,
  author       = {Pith},
  title        = {Pith review of: Efficient Generative Modeling with Residual Vector Quantization-Based Tokens},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7HHJ2SZT}},
  note         = {Machine review of arXiv:2412.10208}
}
read the original abstract

We introduce ResGen, an efficient Residual Vector Quantization (RVQ)-based generative model for high-fidelity generation with fast sampling. RVQ improves data fidelity by increasing the number of quantization steps, referred to as depth, but deeper quantization typically increases inference steps in generative models. To address this, ResGen directly predicts the vector embedding of collective tokens rather than individual ones, ensuring that inference steps remain independent of RVQ depth. Additionally, we formulate token masking and multi-token prediction within a probabilistic framework using discrete diffusion and variational inference. We validate the efficacy and generalizability of the proposed method on two challenging tasks across different modalities: conditional image generation on ImageNet 256x256 and zero-shot text-to-speech synthesis. Experimental results demonstrate that ResGen outperforms autoregressive counterparts in both tasks, delivering superior performance without compromising sampling speed. Furthermore, as we scale the depth of RVQ, our generative models exhibit enhanced generation fidelity or faster sampling speeds compared to similarly sized baseline models.

Figures

Figures reproduced from arXiv: 2412.10208 by the authors.

Figure 1
Figure 1. An overview of the forward masking and reverse unmasking processes is shown at the top, with a detailed depiction of the reverse unmasking process below. In the top figure, forward masking proceeds from right to left, incrementally masking more tokens, while reverse unmasking progresses from left to right, iteratively revealing the masked tokens. White boxes denote masked tokens and colored boxes represent tokens th… view at source ↗
Figure 2
Figure 2. The left figure shows the trade-off between sampling speed and generation quality across various generative models. For ResGen, dotted lines indicate performance across different sampling steps, highlighting step-dependent performance improvements. For other models, solid lines connect results corresponding to variations in parameter size. Note that in ResGen-rvq8 and ResGen-rvq16, the number specifies the depth of … view at source ↗
Figure 3
Figure 3. Effect of RVQ depth on both the autoencoder’s reconstruction quality and our method’s generation quality. We compare two configurations, rvq8 and rvq16, corresponding to RVQ depths of 8 and 16, respectively. As the RVQ depth increases, the autoencoder achieves better reconstruction quality (lower rFID). Despite the increased number of tokens at deeper depth, our generative models show better generation quality (lowe… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: a, increasing the number of sampling steps improves generation quality in both scenarios: with classifier-free guidance (CFG) and without it. This demonstrates that additional steps enable the model to refine its outputs more effectively, resulting in higher-quality ge…
Figure 5
Figure 5. Figure 5: Model comparison on ImageNet 256×256 benchmark. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_5.png]
Figure 6
Figure 6. Figure 6: Randomly generated 256×256 samples by ResGen trained on ImageNet. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 9 canonical work pages

  1. [1]

    L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    arXiv

  2. [2]

    To increase the depth of RVQ, we warm-start from the 4-depth RQ-V AE checkpoint (Lee et al., 2022), excluding the attention layers, and reduce the latent dimension from 256 to

    1 2 and applied throughout all training. To increase the depth of RVQ, we warm-start from the 4-depth RQ-V AE checkpoint (Lee et al., 2022), excluding the attention layers, and reduce the latent dimension from 256 to

    work page 2022

  3. [4]

    B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D

    Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361,

    arXiv 2001

  4. [5]

    doi: 10.1162/tacl a 00618

    ISSN 2307-387X. doi: 10.1162/tacl a 00618. Kim, J., Kim, S., Kong, J., and Yoon, S. Glow-tts: A gen- erative flow for text-to-speech via monotonic alignment search.Advances in Neural Information Processing Sys- tems, 33:8067–8077,

    doi:10.1162/tacl

  5. [6]

    Gemini 1.5: Unlocking multi- modal understanding across millions of tokens of context

    Reid, M., Savinov, N., Teplyashin, D., Lepikhin, D., Lilli- crap, T., Alayrac, J.-b., Soricut, R., Lazaridou, A., Firat, O., Schrittwieser, J., et al. Gemini 1.5: Unlocking multi- modal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530,

    arXiv

  6. [9]

    Llama: Open and efficient foundation lan- guage models.arXiv preprint arXiv:2302.13971,

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozi`ere, B., Goyal, N., Hambro, E., Azhar, F., et al. Llama: Open and efficient foundation lan- guage models.arXiv preprint arXiv:2302.13971,

    arXiv

  7. [11]

    Training details A.1

    12 Efficient Generative Modeling with Residual Vector Quantization-Based Tokens A. Training details A.1. Configurations for Training Vision modelsWe train our method using an architecture similar to DiT (Peebles & Xie, 2023), adopting the XLarge version while modifying the adaptive layer normalization layers for conditioning by replacing their linear laye...

    work page 2023

  8. [16]

    This technique, previously used in CLaM-TTS(Kim et al., 2024), significantly mitigates overhead. 13 Efficient Generative Modeling with Residual Vector Quantization-Based Tokens Training Objective Modification.From Equation 1, the log-likelihood of the target embedding zi is formulated as logp θ(zi|x⊙m) =−loga i + logP ν π(ν) i N( ˜zi;µµµ(ν) i ,I) , where ...

    work page 2024

Show all 18 references
  1. [17]

    These log probabilities are derived from the squared distance between token embeddings and the sampled latent zi at each position i

    and GIVT (Tschannen et al., 2024), we unmask tokens based on the log probabilities computed for all masked tokens. These log probabilities are derived from the squared distance between token embeddings and the sampled latent zi at each position i. The log probability logp(x i,...

    2024

  2. [18]

    As shown in Table 8, these models achieve FID scores of 1.78 and 1.62, respectively

    and MaskBit (Weber et al.). As shown in Table 8, these models achieve FID scores of 1.78 and 1.62, respectively. Our ResGen-rvq16 (with CFG) achieves an FID of 1.93 with a comparable number of sampling steps. While their FID scores are lower than 1.93 of our ResGen-rvq16, ResG...

    2024

  3. [32]

    The real and imaginary parts are concatenated and then encoded through an encoder composed of three blocks, each consisting of three 1D ConvNeXt layers (Siuzdak, 2024; Liu et al.,

    2024

  4. [64]

    For the RVQ quantizer, we employ the probabilistic RVQ method from Kim et al

    Each V AE is further trained for an additional 1M steps, both with and without adversarial training, following the same configuration as prior work. For the RVQ quantizer, we employ the probabilistic RVQ method from Kim et al. (2024), which updates the RVQ codebook embeddings ...

    2024

  5. [1024]

    The resulting token embeddings obtained from the RVQ quantizer are then fed into ResGen, where they are projected to match the hidden size via a linear layer. Audio modelsFor the Text-to-Speech task, our model, based on the DiT XLarge architecture as in the vision task, is tra...

    2024

  6. [2017]

    Neural codec language models are zero-shot text to speech synthesizers

    Wang, C., Chen, S., Wu, Y ., Zhang, Z., Zhou, L., Liu, S., Chen, Z., Liu, Y ., Wang, H., Li, J., et al. Neural codec language models are zero-shot text to speech synthesizers. arXiv preprint arXiv:2301.02111,

    arXiv

  7. [2021]

    Lu- miere: A space-time diffusion model for video generation

    Bar-Tal, O., Chefer, H., Tov, O., Herrmann, C., Paiss, R., Zada, S., Ephrat, A., Hur, J., Liu, G., Raj, A., et al. Lu- miere: A space-time diffusion model for video generation. InSIGGRAPH Asia 2024 Conference Papers, pp. 1–11,

    2024

  8. [2022]

    K., Asawaroengchai, C., Nguyen, D

    Rubenstein, P. K., Asawaroengchai, C., Nguyen, D. D., Bapna, A., Borsos, Z., Quitry, F. d. C., Chen, P., Badawy, D. E., Han, W., Kharitonov, E., et al. Audiopalm: A large language model that can speak and listen.arXiv preprint arXiv:2306.12925,

    arXiv

  9. [2023]

    Soundstorm: Efficient par- allel audio generation.arXiv preprint arXiv:2305.09636,

    Borsos, Z., Sharifi, M., Vincent, D., Kharitonov, E., Zeghi- dour, N., and Tagliasacchi, M. Soundstorm: Efficient par- allel audio generation.arXiv preprint arXiv:2305.09636,

    arXiv

  10. [2024]

    Tang, H., Wu, Y ., Yang, S., Xie, E., Chen, J., Chen, J., Zhang, Z., Cai, H., Lu, Y ., and Han, S

    URL https: //openreview.net/forum?id=vY9nzQmQBw. Tang, H., Wu, Y ., Yang, S., Xie, E., Chen, J., Chen, J., Zhang, Z., Cai, H., Lu, Y ., and Han, S. Hart: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812,

    arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.