Pith. sign in

REVIEW 3 major objections 4 minor 6 cited by

STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Normalizing flows hit diffusion-level image quality at scale.

desk verdict STARFlow is a genuine scaling demonstration for latent autoregressive flows, with a mostly sound empirical core, but the FID attribution to the flow versus the finetuned decoder and a few overclaims need tightening before it should appear. read the letter →

arxiv 2506.06276 v1 pith:XJYHRPXO submitted 2025-06-06 cs.CV cs.AIcs.LG

Jiatao Gu , Tianrong Chen , David Berthelot , Huangjie Zheng , Yuyang Wang , Ruixiang Zhang , Laurent Dinh , Miguel Angel Bautista
show 2 more authors
Josh Susskind Shuangfei Zhai
This is my paper
classification cs.CVcs.AIcs.LG
keywords normalizingflowsautoregressivetransformerflowlatentspacemodelingtext-to-imagegenerationclassifier-freeguidancedeep-shallowarchitectureexactlikelihood
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

This paper claims that normalizing flows, a generative model family trained by exact maximum likelihood, can be scaled to high-resolution and text-conditioned image synthesis, a regime previously dominated by diffusion and discrete autoregressive models. It presents STARFlow, built from Transformer Autoregressive Flow (TARFlow), and identifies three changes that make the scaling work: a deep-shallow stack of flow blocks, learning in the latent space of a pretrained autoencoder with noise-injected latents and a fine-tuned decoder, and a new classifier-free guidance formula that stays stable at high guidance weights. On this recipe, a 1.4B-parameter model reports FID 2.40 on ImageNet-256 and 3.00 on ImageNet-512, and a 3.8B text-to-image model reports 9.1 zero-shot on COCO, approaching diffusion baselines while retaining an end-to-end invertible flow with exact likelihoods. The paper also proves that stacked autoregressive flows with three or more blocks and alternating orderings are universal approximators for continuous densities, which is offered as the theoretical reason the family can be scaled safely.

What carries the argument

The central object is Transformer Autoregressive Flow (TARFlow): a stack of invertible autoregressive blocks, each predicting per-coordinate mean and scale from previously processed coordinates under an alternating ordering, so the whole stack is a single end-to-end normalizing flow optimized by exact likelihood. STARFlow modifies this in three ways: a deep-shallow architecture with one deep $l$-layer block followed by $T-1$ two-layer shallow blocks; modeling of latents from a pretrained autoencoder, corrupted by Gaussian noise at $\sigma_L=0.3$ and decoded by a decoder fine-tuned with $L_2$, LPIPS, and GAN losses; and a guidance rule proved for Gaussian conditionals, $\tilde{\mu}_c = \mu_c + \frac{\omega s}{1+\omega-\omega s}(\mu_c-\mu_u)$ and $\tilde{\sigma}_c = \frac{\sigma_c}{\sqrt{1+\omega-\omega s}}$, with $s=\sigma_c^2/\sigma_u^2$ clipped to $[0,1]$. The universality proposition, that stacked autoregressive flows with $T \geq 3$ blocks are dense in $L^1(\mathbb{R}^D)$, is what licenses scaling the architecture rather than adding more flow blocks.

What would settle it

Decode 50K real ImageNet images from latents corrupted at $\sigma_L=0.3$ and compute rFID; the paper reports 2.73, while STARFlow's best gFID is 2.40. If this rFID is not reproduced, or if at 512x512 with an unfinetuned decoder the gFID/rFID gap inverts, the scaling claim would be shown to be decoder-limited rather than flow-limited.

Watch

Extended reading notes

Core claim

The paper's central claim is that the apparent small-scale ceiling of normalizing flows is not a fundamental limitation of the flow formulation. With a causal Transformer autoregressive flow performing exact maximum-likelihood training in continuous space, the reported results on class- and text-conditioned image synthesis approach state-of-the-art diffusion models at 256, 512, and 1024 resolutions and variable aspect ratios. The load-bearing components are a deep-shallow parameter allocation (one deep block near the prior plus shallow refinement blocks), latent-space modeling on a pretrained autoencoder's noisy latents with a decoder fine-tuned to invert the noise, and a score-derived guidance update that, for Gaussian conditionals, rescales the conditional mean and standard deviation by factors depending on the variance ratio. The paper treats these as the first successful demonstration of normalizing flows at this scale.

Load-bearing premise

The load-bearing premise is that a pretrained autoencoder's latent space, after Gaussian noise at $\sigma_L=0.3$ and decoder finetuning, remains a faithful, smooth proxy for the image distribution, so the reported FID is the flow's quality rather than the decoder's ceiling.

Editorial extensions

If this is right

  • Normalizing flows become a scalable third family for high-resolution generation, with exact densities available for downstream uses such as likelihood scoring and principled training objectives.
  • At ImageNet-256 the reported rFID of 2.73 for the decoder is already above the flow's gFID of 2.40, so further quality gains at this resolution are expected to come from better latent decoders, not bigger flows.
  • The model's end-to-end invertibility transfers to training-free inpainting via a Metropolis-Hastings sampler in latent space and to interactive editing after finetuning, because one forward pass encodes an image and one reverse pass decodes it.
  • The deep block can be initialized from a pretrained LLM and conditioned only there, while shallow blocks refine locally, so the image generator can be dropped into an LLM's semantic space without a separate text encoder.
  • The proposed guidance rule keeps sample quality nearly flat across a wide range of guidance weights, whereas the earlier AF guidance rule degrades quickly away from one optimal weight.

Reading between the lines

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

  • The same deep-shallow, latent-space recipe could plausibly transfer to video, audio, or 3D data, but the paper only demonstrates images, so that transfer is an untested extension.
  • One direct test this account suggests is training STARFlow on clean latents with a small denoising regularizer; if stable, it would remove the noise-injection and decoder-finetuning stages and lift the likelihood ceiling above the current decoder-limited rFID.
  • Because the guidance derivation only assumes Gaussian conditionals, the formula may be applicable to any flow or VAE whose per-variable predictive distributions are Gaussian, not only TARFlow-style models.
  • The gap between rFID 2.73 and gFID 2.40 indicates the flow is already working near the decoder's reconstruction boundary; at resolutions or aspect ratios where the decoder was not fine-tuned, the scaling behavior could differ from the headline numbers.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. STARFlow is a latent-space autoregressive normalizing flow for high-resolution image synthesis. It extends TARFlow with a deep-shallow stack of Transformer AF blocks, trains in the latent space of a frozen pretrained autoencoder corrupted by Gaussian noise at sigma_L = 0.3, and uses a finetuned pixel decoder. The paper also introduces a score-based classifier-free guidance rule (Prop. 2) with variance-ratio clipping and claims a universality result for stacked AFs (Prop. 1). Experiments report FID 2.40 on ImageNet-256 with a 1.4B model, FID 3.00 on ImageNet-512, and zero-shot COCO FID 9.1 with a 3.8B model, plus qualitative text-to-image samples, inpainting, and editing results.

Significance. If the results are taken at face value, this is a meaningful demonstration that exact-likelihood latent normalizing flows can be trained at billion-parameter scale and achieve competitive FIDs, while retaining exact maximum-likelihood training and invertibility. The deep-shallow architecture study, the comparison with TARFlow, and the CFG analysis are useful and largely internally consistent. However, the FID numbers are not cleanly attributable to the flow because of the finetuned decoder, and the headline breadth claims are stronger than the tables support: on ImageNet-256, FID 2.40 is close to DiT's 2.27 but with 1.4B versus 675M parameters; on ImageNet-512, FID 3.00 trails EDM2-XXL's 1.25; on zero-shot COCO, FID 9.1 trails Imagen, Parti, and eDiff-I. With proper decoder-controlled baselines and recalibrated claims, the contribution would be solid.

major comments (3)
  1. [§3.3, Appendix B.3, Tables 1–3] The FID attribution is confounded by the finetuned decoder. Appendix B.3 reports rFID ≈ 2.73 for real latents corrupted at sigma_L = 0.3 and decoded by the finetuned decoder, while Table 1 reports gFID 2.40; the text calls 2.73 an upper bound and a performance ceiling, which is internally inconsistent because the generated samples score better than the decoded noisy real latents. No analogous rFID is reported for ImageNet-512 or MS-COCO, where the headline scaling claims (FID 3.00 and 9.1) are made, and the decoder is finetuned only on 256x256 ImageNet. The authors should report rFID at 512 and COCO, and ideally a decoding-only baseline (e.g., decoding random or naive latents) to separate the flow prior's contribution to FID from the decoder's.
  2. [§3.1, Proposition 1, Appendix A.1] The universality result is not proven at the level claimed. Equation (5) shows that each conditional density is an infinite Gaussian mixture, but the mixture weights are p(y_{>d} | x_{<d}) produced by the previous flow block; the argument does not demonstrate that the set of achievable conditional densities is dense in L1(R^D), only that Gaussian mixtures are dense. The discussion for T ≥ 3 repeats the same reasoning, and no consistency argument is given for approximating the joint density by choosing conditionals independently. Please provide a rigorous proof or explicitly downgrade the statement to a conjecture.
  3. [Abstract, §4.2, Tables 1–3] The claim of 'approaching state-of-the-art diffusion models' is broader than the evidence. On ImageNet-256, STARFlow's FID 2.40 is close to DiT's 2.27 but with 1.4B parameters versus 675M; on ImageNet-512, STARFlow's FID 3.00 is far from EDM2-XXL's 1.25 at comparable model size; on zero-shot COCO, FID 9.1 trails Imagen (7.3), Parti (7.2), and eDiff-I (7.0). Given that the FID metric is decoder-confounded (see the first major comment), the 'approaching SOTA' wording should be recalibrated and the comparison should be placed on a parameter-, compute-, and decoder-matched footing.
minor comments (4)
  1. [§3.3, §4.1, Appendix B.1] There are small presentation errors: 'prertaiend' in §3.3, 'the the realism' in §4.1, and Appendix B.1 references 'Fig. 10(b–e)' while Figure 10 contains panels (a)–(f).
  2. [§3.4, Proposition 2] The clipping s = CLIP(s, 0, 1) is a heuristic that changes the guided distribution; the paper should state this explicitly and analyze the regime sigma_u < sigma_c, where the current 'mode-seeking' motivation is only intuitive.
  3. [Appendix D.1] The Metropolis–Hastings inpainting derivation should be more explicit: the proposal applies f^{-1} and then replaces observed pixels, and it is not immediate that this kernel is symmetric or that the acceptance ratio based on p_theta(x') yields p_theta(x_M | x_O) as the stationary distribution.
  4. [Appendix B.3] The statement that the finetuned decoder 'seamlessly generalizes to arbitrary resolutions, aspect ratios, and text-to-image domains' is supported only by qualitative samples; quantitative reconstruction/rFID numbers at other resolutions and domains should be reported.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: FID numbers come from sampling the trained model, and the two theoretical propositions are closed-form or standard density arguments rather than self-imported conclusions.

full rationale

The claimed derivation chain is not circular. Proposition 1 argues universality of alternating autoregressive flows via Gaussian-mixture density (citing an external textbook) plus neural-network expressivity; Proposition 2 is an explicit closed-form solution of the guided-score equation for Gaussian conditionals, verified by derivation in Appendix A.2. Neither proposition presupposes STARFlow's benchmark performance. The headline FID numbers (2.40 ImageNet-256, 3.00 ImageNet-512, 9.1 COCO) come from actual forward sampling of the trained model, not from fitting those benchmark numbers; the training losses in Eqs. (3), (6), and (7) do not contain FID as a target. The latent-space pipeline does rely on the authors' prior TARFlow (Zhai et al., 2024) for the flow formulation and noise-injection idea, and this self-citation is contextual, but the paper reproduces that baseline from the official codebase and contrasts it with external baselines (DiT, LlamaGen, GIVT), so the central scaling claim has independent support. The manuscript's own Appendix B.3 limitation (rFID 2.73 for decoded noisy real latents vs gFID 2.40) is an acknowledged reconstruction-ceiling concern about attributing quality to the flow versus the finetuned decoder; it is a substantive correctness/attribution caveat, not a definitional or fitted-input circularity. Overall circularity burden is low.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central results depend on a small number of tuned constants and several domain assumptions: the pretrained autoencoder/decoder pair is trusted to preserve information after noise injection, the flow's affine conditionals are treated as isotropic Gaussians for guidance, and the universality proof borrows an unproved Gaussian-mixture and neural approximation step. No new physical or mathematical entities are introduced.

free parameters (5)
  • sigma_L (latent noise scale) = 0.3
    Chosen by preliminary search in Section 3.3 and fixed across all experiments. Noise injection is cited as essential for stable training, so the choice directly affects all reported FID values.
  • deep-shallow block allocation l(N)-d = 18(6)-2048 for class models; 24(6)-3072 for text-to-image models
    Depth of the deep block and number of shallow blocks were selected via ablations shown in Figures 10e and 10f. The paper states this configuration plays a pivotal role in performance.
  • CFG variance-ratio clipping bounds = CLIP(s, 0, 1)
    Introduced in Section 3.4 to keep 1+omega-omega*s >= 1. It is a stability constraint with no derivation from the score-matching objective.
  • latent norm penalty weight = 1e-4
    Added in Appendix B.2 to stabilize log-sigma training. The paper gives no sweep or sensitivity analysis, and the soft-clipping amplitude a in f(x)=a*tanh(x/a) is left unspecified.
  • guidance weight omega = 3.0 for class-conditional, 4.0 for text-to-image
    Evaluation hyperparameter chosen per task. Reported FID values are at these weights, so cross-model comparisons assume the tuning is matched.
assumptions (5)
  • standard math The change-of-variable formula justifies maximum likelihood training of invertible flows.
    Foundation of all normalizing flow training, used throughout Section 2.
  • domain assumption Gaussian mixtures are dense in the space of continuous densities, and neural networks can approximate the required conditional parameters.
    Used in the Proposition 1 proof sketch in Section 3.1 and Appendix A.1. The paper cites Goodfellow et al. 2016 but does not prove the specific alternating AF parameterization is dense.
  • domain assumption A pretrained autoencoder provides a continuous latent space in which Gaussian-noise-corrupted inputs remain decodable after finetuning the decoder.
    Needed for Section 3.3 latent-space modeling and decoder finetuning in Eq. 7. The reported rFID of 2.73 in Appendix B.3 shows decoded real latents are imperfect, so the assumption is only approximately true.
  • domain assumption Each coordinate's conditional predictive distribution in the deep block is an isotropic Gaussian, so the score-based CFG derivation applies.
    An affine AF has Gaussian conditionals by construction, but the paper does not verify the empirical conditionals of the trained deep block satisfy the isotropic Gaussian assumption.
  • ad hoc to paper Clipping the variance ratio s to [0,1] preserves the intended guided distribution.
    Used in Section 3.4 to force numerical stability. No proof is given that s<=1 holds for the data or that the clipped update still corresponds to the score in Eq. 8.

how reviews work

0 comments
Cite this review

Pith. "Pith review of STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis." pith.science (2026). https://pith.science/paper/XJYHRPXO

@misc{pith2026250606276,
  author       = {Pith},
  title        = {Pith review of: STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XJYHRPXO}},
  note         = {Machine review of arXiv:2506.06276}
}
read the original abstract

We present STARFlow, a scalable generative model based on normalizing flows that achieves strong performance in high-resolution image synthesis. The core of STARFlow is Transformer Autoregressive Flow (TARFlow), which combines the expressive power of normalizing flows with the structured modeling capabilities of Autoregressive Transformers. We first establish the theoretical universality of TARFlow for modeling continuous distributions. Building on this foundation, we introduce several key architectural and algorithmic innovations to significantly enhance scalability: (1) a deep-shallow design, wherein a deep Transformer block captures most of the model representational capacity, complemented by a few shallow Transformer blocks that are computationally efficient yet substantially beneficial; (2) modeling in the latent space of pretrained autoencoders, which proves more effective than direct pixel-level modeling; and (3) a novel guidance algorithm that significantly boosts sample quality. Crucially, our model remains an end-to-end normalizing flow, enabling exact maximum likelihood training in continuous spaces without discretization. STARFlow achieves competitive performance in both class-conditional and text-conditional image generation tasks, approaching state-of-the-art diffusion models in sample quality. To our knowledge, this work is the first successful demonstration of normalizing flows operating effectively at this scale and resolution.

Figures

Figures reproduced from arXiv: 2506.06276 by the authors.

Figure 1
Figure 1. Text conditioned high-resolution samples of variable aspect ratios generated from our 3.8B STARFlow model. Resolutions are adjusted for the ease of visualization. Abstract We present STARFlow, a scalable generative model based on normalizing flows that achieves strong performance on high-resolution image synthesis. STARFlow’s main building block is Transformer Autoregressive Flow (TARFlow), which com￾bines normalizi… view at source ↗
Figure 2
Figure 2. Ex. of 2-block AFs. For the final coordinate d = D we have y>D = ∅. Eq. (5) re￾duces to a single Gaussian and the universality is lost. However, this restriction is lifted by extending additional flows (T ≥ 3) to re-introduce latent variables or appending an augmented Gaussian variable. Additional derivation details appear in the Appendix A. □ The preceding proposition clarifies why we can safely scale-up AFs on lar… view at source ↗
Figure 3
Figure 3. Top to bottom, guiding the first 0, 3, 8 flow blocks with a TARFlow model with 8 flow blocks. We see that guidance is only effective up to the top 3 blocks. The derivation in § 3.1 motivates a redesign of scalable AF architectures within realistic computational budgets, emphasizing that we need not greatly expand the number of flow blocks—indeed (even T = 2 often suffices). How￾ever, the remark leaves unresolved how… view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: An illustration of the autoregressive inference (left) and parallel training (right) process of our proposed model for text-to-image generation. The upward (green) and downward (purple) arrows refers to the inverse and forward AF step as shown in Eq. (2). shape from 25…
Figure 5
Figure 5. Figure 5: (a) Guidance from TARFlow (Zhai et al., 2024) (b) Proposed guidance on ImageNet 256 × 256. In the context of AFs, Zhai et al. (2024) made the first attempt to apply CFG by linearly extrapolating the mean and variance at each step (Eq. (2)): µ˜c = µc + ω(µc − µu) and σ˜…
Figure 6
Figure 6. Figure 6: (a) image inpainting (b) interactive editing. STARFlow is a versatile generative model that not only produces diverse, high-quality images under various conditions but also extends natu￾rally to downstream applications. We showcase two examples: image inpainting and ed…
Figure 7
Figure 7. Figure 7: Random samples of STARFlow on ImageNet 256 × 256 and 512 × 512 (ω = 3.0). Interactive Generation and Editing We finetune STARFlow on an image editing dataset (Fig. 6b), enabling joint modeling of generation and editing with a single conditional AF model. Its invertibil…
Figure 8
Figure 8. Figure 8: Selected samples of various aspect ratios from STARFlow on for text-to-image generation (ω = 4.0). Image resolutions are adjusted proportionally for the ease of visualization. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Example of Image editing using STARFlow. Given an input image and simple description, our model can seamlessly edit the contents based on various instruction using with the learned model prior [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: Experimental results of comprehensive ablation study Fig. 10a also compares inference throughput on a single H100 GPU for diffusion, AR, and STARFlow models. Diffusion’s wall-clock time grows linearly with its number of refinement steps—≈ 250 steps at best FID—so it’s…
Figure 11
Figure 11. Figure 11: (a) Direct generation results using the model’s latent samples without decoder fine-tuning or score-based denoising. (b) Results after applying decoder fine-tuning, effectively reducing latent-space noise. (c) and (d) provide comparisons of classifier-free guidance (C…
Figure 12
Figure 12. Figure 12: Demonstration of generation trajectories of inpainting output. [PITH_FULL_IMAGE:figures/full_fig_p022_12.png]
Figure 13
Figure 13. Figure 13: Interactive editing with STARFlow. Starting from an initial caption, STARFlow generates a base image. Given a subsequent user-provided editing instruction, the model then modifies the image accord￾ingly—without requiring re-encoding. Each example illustrates a generic…
Figure 14
Figure 14. Figure 14: Additional class-conditioned generation from STARFlows trained on 256 × 256 and 512 × 512, respectively. The classes are sulphur-crested cockatoo, Kakatoe galerita, Cacatua galerita, loggerhead, loggerhead turtle, Caretta caretta, and Siberian husky. 25 [PITH_FULL_IM…
Figure 15
Figure 15. Figure 15: Additional text-conditioned generation from STARFlows trained on 256 × 256 and 512 × 512, respectively. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_15.png]
Figure 16
Figure 16. Figure 16: Additional text-conditioned generation from STARFlows trained on 1024 × 1024. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_16.png]
Figure 17
Figure 17. Figure 17: Additional text-conditioned samples from STARFlows trained on various aspect ratios. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_17.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 6 Pith papers

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

  1. IB-Flow: Information Bottleneck-Guided CFG Distillation for Few-Step Text-to-Image Generation

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Information-Bottleneck closed forms for adaptive CFG supervisor timestep and strength yield SOTA 2-NFE text-to-image fidelity across FLUX, OpenUni, and Qwen-Image.

  2. MIMFlow: Integrating Masked Image Modeling with Normalizing Flows for End-to-End Image Generation

    cs.CV 2026-06 unverdicted novelty 6.0 of 10

    MIMFlow uses a VAE on masked images to feed semantic latents to a normalizing flow while a decoder handles high-frequency details, reporting FID 2.50 and 71.3% linear probing on ImageNet 256x256 with 128 tokens.

  3. SRC-Flow: Compact Semantic Representations Enable Normalizing Flows for Image Generation

    cs.CV 2026-05 unverdicted novelty 6.0 of 10

    SRC-Flow compresses RAE features via a Semantic Representation Compressor into a low-dimensional space, enabling normalizing flows to reach gFID 1.65 on ImageNet 256x256 and 2.07 on 512x512 while retaining exact likelihoods.

  4. Flow Stochastic Segmentation Networks

    cs.CV 2025-07 conditional novelty 6.0 of 10

    Flow-SSNs model high-rank pixel covariances for ambiguous medical image segmentation by mapping a learned diagonal-Gaussian prior through a lightweight flow, outperforming prior SOTA with fewer parameters.

  5. A Generative Model-Free Form Deformation Approach for the Generation of Mesh Motions with Applications to PDE

    math.NA 2026-07 conditional novelty 5.0 of 10

    A control-point ODE flow with a Bernstein-polynomial drift can match genus-0 meshes of different topology, and a POD+autoencoder+normalizing-flow generative model on the control trajectories yields lower-dimensional s...

  6. Fractal Flow: Hierarchical and Interpretable Normalizing Flow via Topic Modeling and Recursive Strategy

    stat.ML 2025-08 reject novelty 4.0 of 10

    Fractal Flow combines a Dirichlet-topic latent prior with recursive coupling layers in a normalizing flow, reporting lower bits-per-dim than a custom RealNVP baseline on MNIST and FashionMNIST.

Reference graph

Works this paper leans on

70 extracted references · 24 canonical work pages · cited by 6 Pith papers

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    ediffi: Text-to-image diffusion models with an ensemble of expert denoisers

    Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022

    arXiv 2022

  3. [3]

    Improving image generation with better captions

    Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al. Improving image generation with better captions. Computer Science. https://cdn. openai. com/papers/dall-e-3. pdf, 2 0 (3): 0 8, 2023

    2023

  4. [4]

    Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, ...

    work page 2020

  5. [5]

    P., Freeman, W

    Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.-H., Murphy, K. P., Freeman, W. T., Rubinstein, M., Li, Y., and Krishnan, D. Muse: Text-to-image generation via masked generative transformers. In Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., and Scarlett, J. (eds.), Proceedings of the 40th International Con...

    work page 2023

  6. [6]

    Conceptual 12M : Pushing web-scale image-text pre-training to recognize long-tail visual concepts

    Changpinyo, S., Sharma, P., Ding, N., and Soricut, R. Conceptual 12M : Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021

    work page 2021

  7. [7]

    Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis

    Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023

    arXiv

  8. [8]

    T., Rubanova, Y., Bettencourt, J., and Duvenaud, D

    Chen, R. T., Rubanova, Y., Bettencourt, J., and Duvenaud, D. K. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018

    2018

Show all 70 references
  1. [9]

    W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, E., Wang, X., Dehghani, M., Brahma, S., Webson, A., Gu, S

    Chung, H. W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, E., Wang, X., Dehghani, M., Brahma, S., Webson, A., Gu, S. S., Dai, Z., Suzgun, M., Chen, X., Chowdhery, A., Narang, S., Mishra, G., Yu, A., Zhao, V., Huang, Y., Dai, A., Yu, H., Petrov, S., Chi, E. H., Dean...

    2022 arXiv

  2. [10]

    ImageNet: A Large-scale Hierarchical Image Database

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. ImageNet: A Large-scale Hierarchical Image Database . IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 248--255, 2009

    2009

  3. [11]

    and Nichol, A

    Dhariwal, P. and Nichol, A. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34: 0 8780--8794, 2021

    2021

  4. [12]

    Cogview: Mastering text-to-image generation via transformers

    Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34: 0 19822--19835, 2021

    2021

  5. [13]

    Nice: Non-linear independent components estimation

    Dinh, L., Krueger, D., and Bengio, Y. Nice: Non-linear independent components estimation. arXiv preprint arXiv:1410.8516, 2014

    2014 arXiv

  6. [14]

    Density estimation using real nvp

    Dinh, L., Sohl-Dickstein, J., and Bengio, S. Density estimation using real nvp. arXiv preprint arXiv:1605.08803, 2016

    2016 arXiv

  7. [15]

    The llama 3 herd of models

    Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  8. [16]

    Tweedie’s formula and selection bias

    Efron, B. Tweedie’s formula and selection bias. Journal of the American Statistical Association, 106 0 (496): 0 1602--1614, 2011

    2011

  9. [17]

    Taming transformers for high-resolution image synthesis

    Esser, P., Rombach, R., and Ommer, B. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 12873--12883, 2021

    2021

  10. [18]

    Scaling rectified flow transformers for high-resolution image synthesis

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., M \"u ller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024

    2024

  11. [19]

    Fluid: Scaling autoregressive text-to-image generative models with continuous tokens

    Fan, L., Li, T., Qin, S., Li, Y., Sun, C., Rubinstein, M., Sun, D., He, K., and Tian, Y. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens. arXiv preprint arXiv:2410.13863, 2024

    2024 arXiv

  12. [20]

    Make-a-scene: Scene-based text-to-image generation with human priors

    Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., and Taigman, Y. Make-a-scene: Scene-based text-to-image generation with human priors. 2022. doi:10.48550/ARXIV.2203.13131. URL https://arxiv.org/abs/2203.13131

    2022 doi arXiv

  13. [21]

    Geneval: An object-focused framework for evaluating text-to-image alignment

    Ghosh, D., Hajishirzi, H., and Schmidt, L. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36: 0 52132--52152, 2023

    2023

  14. [22]

    Deep Learning

    Goodfellow, I., Bengio, Y., and Courville, A. Deep Learning. MIT Press, 2016. http://www.deeplearningbook.org

    2016

  15. [23]

    T., Bettencourt, J., Sutskever, I., and Duvenaud, D

    Grathwohl, W., Chen, R. T., Bettencourt, J., Sutskever, I., and Duvenaud, D. Ffjord: Free-form continuous dynamics for scalable reversible generative models. arXiv preprint arXiv:1810.01367, 2018

    2018 arXiv

  16. [24]

    Gu, J., Shen, Y., Zhai, S., Zhang, Y., Jaitly, N., and Susskind, J. M. Kaleido diffusion: Improving conditional diffusion models with autoregressive latent modeling. arXiv preprint arXiv:2405.21048, 2024 a

    2024 arXiv

  17. [25]

    Dart: Denoising autoregressive transformer for scalable text-to-image generation

    Gu, J., Wang, Y., Zhang, Y., Zhang, Q., Zhang, D., Jaitly, N., Susskind, J., and Zhai, S. Dart: Denoising autoregressive transformer for scalable text-to-image generation. arXiv preprint arXiv:2410.08159, 2024 b

    2024 arXiv

  18. [26]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  19. [27]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and Hochreiter, S. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017

    2017

  20. [28]

    and Salimans, T

    Ho, J. and Salimans, T. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021

    2021

  21. [29]

    Flow++: Improving flow-based generative models with variational dequantization and architecture design

    Ho, J., Chen, X., Srinivas, A., Duan, Y., and Abbeel, P. Flow++: Improving flow-based generative models with variational dequantization and architecture design. In International conference on machine learning, pp.\ 2722--2730. PMLR, 2019

    2019

  22. [30]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33: 0 6840--6851, 2020

    2020

  23. [31]

    P., Poole, B., Norouzi, M., Fleet, D

    Ho, J., Chan, W., Saharia, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D. P., Poole, B., Norouzi, M., Fleet, D. J., et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022 a

    arXiv

  24. [32]

    J., Norouzi, M., and Salimans, T

    Ho, J., Saharia, C., Chan, W., Fleet, D. J., Norouzi, M., and Salimans, T. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23: 0 47--1, 2022 b

    2022

  25. [33]

    Hutchinson, M. F. A stochastic estimator of the trace of the influence matrix for laplacian smoothing splines. Communications in Statistics-Simulation and Computation, 18 0 (3): 0 1059--1076, 1989

    1989

  26. [34]

    Scalable adaptive computation for iterative generation

    Jabri, A., Fleet, D., and Chen, T. Scalable adaptive computation for iterative generation. arXiv preprint arXiv:2212.11972, 2022

    2022 arXiv

  27. [35]

    Scaling up gans for text-to-image synthesis

    Kang, M., Zhu, J.-Y., Zhang, R., Park, J., Shechtman, E., Paris, S., and Park, T. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 10124--10134, 2023

    2023

  28. [36]

    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, 2020

    2001 arXiv

  29. [37]

    Guiding a diffusion model with a bad version of itself

    Karras, T., Aittala, M., Kynk \"a \"a nniemi, T., Lehtinen, J., Aila, T., and Laine, S. Guiding a diffusion model with a bad version of itself. Advances in Neural Information Processing Systems, 37: 0 52996--53021, 2024

    2024

  30. [38]

    Kingma, D. P. and Welling, M. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013

    2013 arXiv

  31. [39]

    P., Salimans, T., Jozefowicz, R., Chen, X., Sutskever, I., and Welling, M

    Kingma, D. P., Salimans, T., Jozefowicz, R., Chen, X., Sutskever, I., and Welling, M. Improved variational inference with inverse autoregressive flow. Advances in neural information processing systems, 29, 2016

    2016

  32. [40]

    Autoregressive image generation using residual quantization

    Lee, D., Kim, C., Kim, S., Cho, M., and Han, W.-S. Autoregressive image generation using residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 11523--11532, 2022

    2022

  33. [41]

    Autoregressive image generation without vector quantization

    Li, T., Tian, Y., Li, H., Deng, M., and He, K. Autoregressive image generation without vector quantization. arXiv preprint arXiv:2406.11838, 2024

    2024 arXiv

  34. [42]

    Fractal generative models

    Li, T., Sun, Q., Fan, L., and He, K. Fractal generative models. arXiv preprint arXiv:2502.17437, 2025

    2025 arXiv

  35. [43]

    A survey of multimodel large language models

    Liang, Z., Xu, Y., Hong, Y., Shang, P., Wang, Q., Fu, Q., and Liu, K. A survey of multimodel large language models. In Proceedings of the 3rd International Conference on Computer, Artificial Intelligence and Control Engineering, pp.\ 405--409, 2024

    2024

  36. [44]

    Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Doll \' a r, P., and Zitnick, C. L. Microsoft COCO: Common Objects in Context . European Conference on Computer Vision, pp.\ 740--755, 2014

    2014

  37. [45]

    Lipman, Y., Chen, R. T. Q., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=PqvMRDCJT9t

    2023

  38. [46]

    S., Boffi, N

    Ma, N., Goldstein, M., Albergo, M. S., Boffi, N. M., Vanden-Eijnden, E., and Xie, S. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024

    2024 arXiv

  39. [47]

    Gpt‑4o system card

    OpenAI. Gpt‑4o system card. https://openai.com/index/gpt-4o-system-card/, 2024. Accessed: April 12, 2025

    2024

  40. [48]

    Masked autoregressive flow for density estimation

    Papamakarios, G., Murray, I., and Pavlakou, T. Masked autoregressive flow for density estimation. In Guyon, I., von Luxburg, U., Bengio, S., Wallach, H. M., Fergus, R., Vishwanathan, S. V. N., and Garnett, R. (eds.), Advances in Neural Information Processing Systems 30: Annual...

    2017

  41. [49]

    and Xie, S

    Peebles, W. and Xie, S. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 4195--4205, 2023

    2023

  42. [50]

    Sdxl: improving latent diffusion models for high-resolution image synthesis

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., M \"u ller, J., Penna, J., and Rombach, R. Sdxl: improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023

    2023 arXiv

  43. [51]

    Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., and Liu, P. J. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21 0 (140): 0 1--67, 2020

    2020

  44. [52]

    Zero-shot text-to-image generation

    Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., and Sutskever, I. Zero-shot text-to-image generation. In International Conference on Machine Learning, pp.\ 8821--8831. PMLR, 2021

    2021

  45. [53]

    Hierarchical text-conditional image generation with clip latents

    Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., and Chen, M. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022

    2022 arXiv

  46. [54]

    and Mohamed, S

    Rezende, D. and Mohamed, S. Variational inference with normalizing flows. In Bach, F. and Blei, D. (eds.), Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pp.\ 1530--1538, Lille, France, 07--09 Jul 20...

    2015

  47. [55]

    High-resolution image synthesis with latent diffusion models

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10684--10695, 2022

    2022

  48. [56]

    Roformer: Enhanced transformer with rotary position embedding

    Su, J., Ahmed, M., Lu, Y., Pan, S., Bo, W., and Liu, Y. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568: 0 127063, 2024

    2024

  49. [57]

    Autoregressive model beats diffusion: Llama for scalable image generation

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

    2024 arXiv

  50. [58]

    Chameleon: Mixed-modal early-fusion foundation models, 2024

    Team, C. Chameleon: Mixed-modal early-fusion foundation models, 2024

    2024

  51. [59]

    G., Hardin, C., Bhupatiraju, S., Hussenot, L., Mesnard, T., Shahriari, B., Ram \'e , A., et al

    Team, G., Riviere, M., Pathak, S., Sessa, P. G., Hardin, C., Bhupatiraju, S., Hussenot, L., Mesnard, T., Shahriari, B., Ram \'e , A., et al. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118, 2024

    2024 arXiv

  52. [60]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction

    Tian, K., Jiang, Y., Yuan, Z., Peng, B., and Wang, L. Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2404.02905, 2024

    2024 arXiv

  53. [61]

    Givt: Generative infinite-vocabulary transformers

    Tschannen, M., Eastwood, C., and Mentzer, F. Givt: Generative infinite-vocabulary transformers. In European Conference on Computer Vision, pp.\ 292--309. Springer, 2024 a

    2024

  54. [62]

    S., and Kolesnikov, A

    Tschannen, M., Pinto, A. S., and Kolesnikov, A. Jetformer: An autoregressive generative model of raw images and text. arXiv preprint arXiv:2411.19722, 2024 b

    2024 arXiv

  55. [63]

    N., Kaiser, L., and Polosukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin, I. Attention Is All You Need . Advances in Neural Information Processing Systems, pp.\ 5998--6008, 2017

    2017

  56. [64]

    Emu3: Next-token prediction is all you need

    Wang, X., Zhang, X., Luo, Z., Sun, Q., Cui, Y., Wang, J., Zhang, F., Wang, Y., Li, Z., Yu, Q., et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024

    2024 arXiv

  57. [65]

    J., Wang, W., Lin, K

    Xie, J., Mao, W., Bai, Z., Zhang, D. J., Wang, W., Lin, K. Q., Gu, Y., Chen, Z., Yang, Z., and Shou, M. Z. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024

    2024 arXiv

  58. [66]

    Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B

    Yu, J., Xu, Y., Koh, J. Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B. K., et al. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022

    2022

  59. [67]

    A., Jaitly, N., and Susskind, J

    Zhai, S., Zhang, R., Nakkiran, P., Berthelot, D., Gu, J., Zheng, H., Chen, T., Bautista, M. A., Jaitly, N., and Susskind, J. Normalizing flows are capable generative models. arXiv preprint arXiv:2412.06329, 2024

    2024 arXiv

  60. [68]

    Fast training of diffusion models with masked transformers

    Zheng, H., Nie, W., Vahdat, A., and Anandkumar, A. Fast training of diffusion models with masked transformers. In Transactions on Machine Learning Research (TMLR), 2024 a

    2024

  61. [69]

    Learning stackable and skippable LEGO bricks for efficient, reconfigurable, and variable-resolution diffusion modeling

    Zheng, H., Wang, Z., Yuan, J., Ning, G., He, P., You, Q., Yang, H., and Zhou, M. Learning stackable and skippable LEGO bricks for efficient, reconfigurable, and variable-resolution diffusion modeling. In The Twelfth International Conference on Learning Representations, 2024 b ...

    2024

  62. [70]

    Transfusion: Predict the next token and diffuse images with one multi-modal model

    Zhou, C., Yu, L., Babu, A., Tirumala, K., Yasunaga, M., Shamis, L., Kahn, J., Ma, X., Zettlemoyer, L., and Levy, O. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.