Pith. sign in

REVIEW 4 major objections 4 minor 28 references

Fine-Tuning Next-Scale Visual Autoregressive Models with Group Relative Policy Optimization

T0 review · 4 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 The paper claims that GRPO fine-tuning lets a class-conditioned next-scale visual autoregressive model align outputs to aesthetic and text rewards, including painting styles absent from its ImageNet pre-training.

desk verdict A clean, honest GRPO-for-VAR application with a genuinely surprising out-of-distribution paint effect, but the style claim is measured only by the training reward itself. read the letter →

arxiv 2505.23331 v2 pith:XUWX5X7G submitted 2025-05-29 cs.CV cs.AIcs.LG

Matteo Gallici , Haitz Sáez de Ocáriz Borde This is my paper
classification cs.CVcs.AIcs.LG
keywords GRPOvisualautoregressivemodelsreinforcementlearningfine-tuningimagegenerationCLIPalignmentaestheticrewardhackingnext-scaleprediction
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

Group Relative Policy Optimization (GRPO) fine-tuning lets pre-trained next-scale visual autoregressive (VAR) image models follow scalar reward signals, and the paper argues this is the first demonstration of RL-based alignment for this model class. With an aesthetic-preference reward, both a 310M-parameter and a 2B-parameter VAR improve their aesthetic score by roughly one point while keeping ImageNet class accuracy above 90%. With a CLIP reward for the prompt "a painting", the model produces painterly images even though ImageNet deliberately contains photos only, so the paper concludes that RL-driven exploration can teach visual styles absent from pre-training. The claim matters because it suggests RL can expand what a generative model can do beyond its training distribution, and because VAR's fast inference makes online RL practical compared with diffusion models.

What carries the argument

The load-bearing object is the GRPO objective, which replaces PPO's learned value function with advantages computed from the mean and standard deviation of rewards within each group of G samples from the same label. This objective carries the argument because it converts a single image-level scalar reward into per-token policy updates through the clipped importance ratio, while the KL term prevents the model from collapsing into reward hacking. The other half of the machinery is the next-scale autoregressive formulation, in which the image is generated scale by scale from coarse to fine; because each scale emits a block of tokens rather than one token, the policy update treats the whole block as one action, and the frozen VQ-VAE decoder keeps the image space fixed.

What would settle it

Run a held-out human preference study or an independent style classifier on the fine-tuned samples: if the CLIP-reward gains on "a painting" do not transfer to human ratings of painting-ness, or if the aesthetic gains vanish under a different aesthetic predictor, the claim that RL taught new styles collapses. A direct check is whether the high CLIP scores persist when the reward model is frozen but the generation is evaluated with an out-of-distribution detector for paintings.

Watch

Extended reading notes

Core claim

The central discovery is that GRPO, originally developed for language-model reasoning, transfers directly to next-scale autoregressive image generation by treating each next-scale token block as one policy step. Rewards are computed on the decoded image, and the same group-relative advantage is applied to every token in the sample, with a per-token KL penalty that keeps the policy near the pre-trained model. Under this scheme, VAR-d16 and VAR-d30 improve aesthetic scores by about one point, and a CLIP reward conditioned on "a painting" yields paintings despite the photos-only ImageNet annotation policy. The paper interprets this as evidence that through RL exploration and exploitation the model learned artistic patterns with no demonstrations, a conclusion it flags as preliminary but supports with the observed increase in CLIP score and the visual presence of chiaroscuro-like lighting.

Load-bearing premise

The load-bearing premise is that the scalar rewards—the aesthetic-preference predictor and the CLIP text-image score—actually measure the improvements the paper claims, rather than being gamed by shortcuts the reward model happens to reward.

Editorial extensions

If this is right

  • Aesthetic-reward GRPO raises aesthetic scores by roughly one point while ResNet50 top-5 accuracy stays above 90%, so alignment does not destroy class conditioning.
  • CLIP-reward fine-tuning aligns generation to prompts like "a painting" even though ImageNet excluded paintings, implying RL can add styles that were never in the pre-training data.
  • The KL penalty is essential: without it the policy finds reward-hacking shortcuts and loses label conditioning.
  • Increasing group size during sampling improves performance monotonically up to 16 groups given a fixed compute budget.
  • Because the method treats VAR exactly like an LLM in discrete token space, the same recipe should transfer to text-to-image next-scale models.

Reading between the lines

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

  • If RL can invent styles absent from pretraining, the training distribution is not a hard ceiling for generative models; reward-driven search may discover compositions, lighting, or textures never seen by the model, which could matter for creative tools.
  • The paper's own observation that the model adopts a chiaroscuro-like lighting effect to inflate the CLIP score suggests the reward proxy is being gamed; a natural extension is to add human-preference or diversity rewards to suppress such shortcuts.
  • The group-relative advantage assigns the same credit to every token in an image, which is crude for long sequences; a per-token or per-scale credit assignment could improve sample efficiency and is a testable modification.
  • The same GRPO recipe for discrete-token generation should extend to text-to-image next-scale models and, plausibly, to video autoregressive models, where fast sampling also matters for online RL.
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

4 major / 4 minor

Summary. The paper proposes applying Group Relative Policy Optimization (GRPO) to fine-tune next-scale visual autoregressive (VAR) models. It reports three sets of experiments: a brightness toy task, fine-tuning with a LAION aesthetic predictor, and fine-tuning with a CLIP score against fixed text prompts. The central claim is that RL-driven exploration allows a class-conditioned VAR model to align with styles absent from ImageNet pretraining, in particular the prompt 'a painting', and that autoregressive models' fast inference makes such online RL efficient. The paper also includes ablations on the KL coefficient and the number of groups.

Significance. If the central claim is established, this would be the first demonstration of GRPO on next-scale VAR models and would provide evidence that RL can induce new visual styles without demonstrations. The paper deserves credit for applying the GRPO objective correctly, including a clean toy validation in Section 5.1, a useful label-fidelity control with ResNet50 accuracy in Table 1, and explicit ablations of regularization and group size. However, the headline out-of-distribution style claim currently rests almost entirely on reward curves and selected qualitative images, with the evaluation metric being the same reward used for training. The scientific significance of the paper will therefore depend on whether an independent evaluation can be added in revision.

major comments (4)
  1. [§5.3, Figure 9] The central claim concerning out-of-distribution painting styles is not independently established. The reported CLIP score is the same scalar used as the training reward in Eq. (1), so the upward curve in Figure 6 demonstrates that GRPO maximizes its training objective but cannot, by itself, show that the resulting images are humanly recognizable as paintings. The caption to Figure 9 explicitly describes a chiaroscuro-like dark-background shortcut, which is exactly the kind of reward-exploiting low-dimensional change known to inflate CLIP-style proxies. An independent style evaluation, such as a held-out human preference study, a painting/photograph classifier trained on data not used for reward computation, or a diversity metric over generated samples, is needed before the 'learned to paint' claim can be supported.
  2. [§5.3] No baseline is provided against which the RL fine-tuning is compared. The conclusion that GRPO is responsible for the observed style alignment would be strengthened by comparing GRPO fine-tuning with rejection sampling from the base model using the same CLIP reward, or with direct optimization of the prompt embedding on frozen VAE latents. Without such a comparison, the paper's claim that RL is an effective way to achieve this alignment, as opposed to other reward-guided search procedures, remains unsupported.
  3. [§5.2, Table 1] The aesthetic-score results are presented without variance or multiple-seed information. Since the AES predictor is used both as the training reward and as the evaluation metric, reporting standard errors across seeds and samples is necessary to establish that the reported ~1-point improvements are robust and not the result of a single optimization run. The ResNet50 accuracy control is useful and should be retained, but it is not a substitute for an independent measure of aesthetic quality.
  4. [Abstract and Section 5.2] The paper claims that fast inference of VAR models is particularly advantageous for online RL sampling compared with diffusion-based alternatives, but no wall-clock or sample-efficiency comparison with a diffusion-based RL method is provided. As stated, the efficiency advantage is plausible but unmeasured; a direct comparison with, for example, DDPO or Flow-GRPO on a matched reward would make the contribution more convincing.
minor comments (4)
  1. [Eq. (1)] The KL penalty term is written as βDKL without specifying the distribution over which the expectation is taken; the surrounding text clarifies that it is per-token, but the notation would benefit from an explicit expectation or subscript.
  2. [Figures 2, 4, 6, and 8] The reward curves lack error bars or shaded confidence intervals, which makes it difficult to judge the stability of the optimization across runs.
  3. [§5.3] The CLIP model used to compute the score is not specified (for example, ViT-B/32 versus other variants); specifying the exact checkpoint would improve reproducibility.
  4. [Throughout] Minor typographical issues appear throughout, such as 'Aaesthetic' in Figure 9, 'V oronov' and 'V Y' in the references, and inconsistent use of '1k' versus '1,000'.

Circularity Check

1 steps flagged · score 3.0 of 10

Main OOD style claim rests on CLIP score, which is also the optimized GRPO reward; no independent style metric is provided.

  1. self definitional [Section 4.2 (reward model) and Section 5.3, 'Learning to Paint with CLIP Alignment']
    "During training ... the outputs ... are subsequently decoded into images by the VQVAE and evaluated by a specific image reward model. ... In this experiment, our goal is to assess the ability of GRPO to align a generative model with respect to a different modality (text) ... while the CLIP Score is computed relative to a fixed text prompt. ... the policy finds ways to increase the 'painting' CLIP Score. This suggests that, at least in part, the model really learned something new through RL exploration and exploitation, i.e., it learned artistic patterns without demonstrations."

    The quantitative evidence for 'learned artistic patterns' is the increase in CLIP Score, but CLIP Score is exactly the reward signal that GRPO optimizes (Eq. 1 with rewards from the image reward model). Thus the reported improvement is the optimization objective itself, not an independent test of painting-likeness. The paper layers a semantic interpretation ('learned to paint') onto the optimized scalar, and the Figure 9 caption explicitly notes the model raises the score by adopting a chiaroscuro-like darkening shortcut, consistent with proxy optimization. No human evaluation, external painting classifier, or independent style metric is provided, so the central out-of-distribution style claim is not separately validated.

full rationale

The paper is an empirical RL study rather than a formal derivation, so there is no mathematical chain to reduce, and there is no load-bearing self-citation: VAR, GRPO, CLIP, and AES are all external references with independent provenance. The one genuine circularity-adjacent issue is evaluative: the headline 'learning to paint' is supported quantitatively by the CLIP score, which is the same scalar that GRPO is trained to maximize, making the metric self-referential rather than an independent confirmation of the claimed style. This is softened by the paper's own caveat that the hypothesis 'should be explored more deeply' and by the admission in Figure 9 that the model exploits a chiaroscuro-like shortcut, which actually supports a reward-hacking reading. ResNet50 accuracy provides independent evidence that class conditioning is preserved, and the qualitative figures give some non-scalar visual evidence, so the work is not wholly circular; nevertheless, the central out-of-distribution style claim lacks an external evaluation and is partly forced by the chosen reward.

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

The paper introduces no new entities, forces, or conserved quantities. Its central claims rest on reward-model validity and on standard GRPO machinery.

free parameters (6)
  • KL coefficient beta = 0.2
    Strength of the KL penalty to the pretrained model; authors found it essential for preserving generation quality, chosen by hand.
  • Sampling temperature tau = 0.7
    Exploration-exploitation tradeoff during training; chosen by hand to keep the sampling distribution tractable.
  • Number of groups G = 16
    Group size for advantage normalization; ablation shows more groups help, and 16 was selected.
  • Learning rate = 1e-4
    Standard optimization hyperparameter, chosen by hand.
  • Clipping epsilon = 0.2
    PPO clipping threshold, chosen by hand.
  • Aesthetic reward bonus weight
    In Section 5.3, the CLIP reward is combined with an aesthetic bonus, but the combination weight is not reported; it affects the observed style tradeoff.
assumptions (6)
  • standard math GRPO clipped surrogate objective with group advantage is a valid optimization target
    Imported from Shao et al. 2024 and used without modification.
  • domain assumption VQ-VAE decoder is frozen and reliably decodes token sequences for reward computation
    Section 4.2 relies on this to compute image rewards from sampled tokens.
  • domain assumption LAION aesthetic predictor V2 is a faithful proxy for human aesthetic preference
    Section 5.2 uses it as the main reward and evaluation metric.
  • domain assumption CLIP score measures semantic and stylistic alignment between generated image and fixed text prompt
    Section 5.3 uses CLIP score as reward and as evidence of 'learning to paint'.
  • domain assumption ImageNet contains no paintings or drawings
    Section 5.3 cites annotation instructions from Deng 2012 to argue painting-like outputs are out of distribution.
  • domain assumption Applying the same advantage to all tokens in a scale is a valid credit assignment approximation
    Section 4.2 states this approximation and uses it for per-token KL penalties.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fine-Tuning Next-Scale Visual Autoregressive Models with Group Relative Policy Optimization." pith.science (2026). https://pith.science/paper/XUWX5X7G

@misc{pith2026250523331,
  author       = {Pith},
  title        = {Pith review of: Fine-Tuning Next-Scale Visual Autoregressive Models with Group Relative Policy Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XUWX5X7G}},
  note         = {Machine review of arXiv:2505.23331}
}
read the original abstract

Fine-tuning pre-trained generative models with Reinforcement Learning (RL) has emerged as an effective approach for aligning outputs more closely with nuanced human preferences. In this paper, we investigate the application of Group Relative Policy Optimization (GRPO) to fine-tune next-scale visual autoregressive (VAR) models. Our empirical results demonstrate that this approach enables alignment to intricate reward signals derived from aesthetic predictors and CLIP embeddings, significantly enhancing image quality and enabling precise control over the generation style. Interestingly, by leveraging CLIP, our method can help VAR models generalize beyond their initial ImageNet distribution: through RL-driven exploration, these models can generate images aligned with prompts referencing image styles that were absent during pre-training. In summary, we show that RL-based fine-tuning is both efficient and effective for VAR models, benefiting particularly from their fast inference speeds, which are advantageous for online sampling, an aspect that poses significant challenges for diffusion-based alternatives.

Figures

Figures reproduced from arXiv: 2505.23331 by the authors.

Figure 1
Figure 1. Images from VAR-d30 (Tian et al., 2024) fine-tuned with group relative policy optimization. At the top, samples from the model fine-tuned with an aesthetic reward. In the middle, samples from the model fine-tuned to produce paintings optimized with CLIP score. At the bottom, the model is fine-tuned with a combination of the two rewards. Given the notable scarcity of artworks in the ImageNet dataset used for pre-trai… view at source ↗
Figure 3
Figure 3. Visual results of the toy experiment aligning the model to generate only bright or dark images. training on ImageNet. This amounted to 16 and 40 hours of training, respectively, using 8 NVIDIA H100 GPUs. As shown in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 2
Figure 2. Fine-Tuning VAR-d16 with GRPO on dark/bright align￾ment requires only 10 minutes with a single H100 GPU. To align the model to generate images with high brightness, we set up a scalar reward giving a value of 1 when brightness is greater than or equal to 0.8, and 0 otherwise. For aligning the model to generate dark images, we give a reward of 1 when brightness is lower than 0.2. As shown in [PITH_FULL_IMAGE:figures… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Fine-Tuning VAR-d16 and VAR-d30 with GRPO using Aesthetic Reward. In order to validate that the fine-tuned models are still ca￾pable of generating images corresponding to the ImageNet labels, we can no longer rely on the FID score, as the RL optimization has altered th…
Figure 6
Figure 6. Figure 6: CLIP Score measuring alignment with the prompt “A painting” during fine-tuning, with and without the application of the Aesthetic Reward bonus (excluded from the plots). 6 [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 9
Figure 9. Figure 9: Experiments conducted to align VAR with a fixed prompt using the CLIP Score. In this setup, VAR itself is not conditioned on the prompt; only the reward model is. Since VAR was trained on ImageNet, the training data predominantly consists of photographs (no paintings, …
Figure 8
Figure 8. Figure 8: Effect of varying the number of groups during the sam￾pling phase of GRPO (Aesthetic Reward optimization). 6. Conclusion In this work, we have demonstrated the effectiveness and efficiency of GRPO for aligning next-scale visual autoregres￾sive models. A natural next st…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 5 canonical work pages

  1. [1]

    Training diffusion models with reinforcement learning, 2024

    Black, K., Janner, M., Du, Y., Kostrikov, I., and Levine, S. Training diffusion models with reinforcement learning, 2024. URL https://arxiv.org/abs/2305.13301

    arXiv 2024

  2. [2]

    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, ...

    arXiv 2020

  3. [3]

    Large scale visual recognition

    Deng, J. Large scale visual recognition. PhD thesis, Princeton University, 2012

    work page 2012

  4. [4]

    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., Podell, D., Dockhorn, T., English, Z., and Rombach, R. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. URL https://openreview.net/forum?id=FPnUhsQJ5B

    work page 2024

  5. [5]

    J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y

    Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. Generative adversarial networks, 2014. URL https://arxiv.org/abs/1406.2661

    arXiv 2014

  6. [6]

    Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis, 2024

    Han, J., Liu, J., Jiang, Y., Yan, B., Zhang, Y., Yuan, Z., Peng, B., and Liu, X. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis, 2024. URL https://arxiv.org/abs/2412.04431

    arXiv 2024

  7. [7]

    and Salimans, T

    Ho, J. and Salimans, T. Classifier-free diffusion guidance, 2022. URL https://arxiv.org/abs/2207.12598

    arXiv 2022

  8. [8]

    Kingma, D. P. and Welling, M. Auto-encoding variational bayes. CoRR, abs/1312.6114, 2013. URL https://api.semanticscholar.org/CorpusID:216078090

    arXiv 2013

Show all 28 references
  1. [9]

    Lee, K., Liu, H., Ryu, M., Watkins, O., Du, Y., Boutilier, C., Abbeel, P., Ghavamzadeh, M., and Gu, S. S. Aligning text-to-image models using human feedback, 2023. URL https://arxiv.org/abs/2302.12192

    2023 arXiv

  2. [10]

    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

  3. [11]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Liu, X., Gong, C., and Liu, Q. Flow straight and fast: Learning to generate and transfer data with rectified flow. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=XVjTT1nw5z

    2023

  4. [12]

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P., Leike, J., and Lowe, R. Training language models to f...

    2022 arXiv

  5. [13]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., and Rombach, R. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. URL https://arxiv.org/abs/2307.01952

    2023 arXiv

  6. [14]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pp.\ 8748--8763. PmLR, 2021

    2021

  7. [15]

    Zero-shot text-to-image generation, 2021

    Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., and Sutskever, I. Zero-shot text-to-image generation, 2021. URL https://arxiv.org/abs/2102.12092

    2021 arXiv

  8. [16]

    High-resolution image synthesis with latent diffusion models, 2022

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models, 2022. URL https://arxiv.org/abs/2112.10752

    2022 arXiv

  9. [17]

    Laion-5b: An open large-scale dataset for training next generation image-text models

    Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35: 0...

    2022

  10. [18]

    Proximal policy optimization algorithms, 2017

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. Proximal policy optimization algorithms, 2017. URL https://arxiv.org/abs/1707.06347

    2017 arXiv

  11. [19]

    K., Wu, Y., and Guo, D

    Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y. K., Wu, Y., and Guo, D. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024. URL https://arxiv.org/abs/2402.03300

    2024 arXiv

  12. [20]

    N., Weiss, E

    Sohl-Dickstein, J. N., Weiss, E. A., Maheswaranathan, N., and Ganguli, S. Deep unsupervised learning using nonequilibrium thermodynamics. ArXiv, abs/1503.03585, 2015. URL https://api.semanticscholar.org/CorpusID:14888175

    2015 arXiv

  13. [21]

    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. 2024

    2024

  14. [22]

    Neural discrete representation learning

    van den Oord, A., Vinyals, O., and Kavukcuoglu, K. Neural discrete representation learning. ArXiv, abs/1711.00937, 2017. URL https://api.semanticscholar.org/CorpusID:20282961

    2017 arXiv

  15. [23]

    Switti: Designing scale-wise transformers for text-to-image synthesis

    Voronov, A., Kuznedelev, D., Khoroshikh, M., Khrulkov, V., and Baranchuk, D. Switti: Designing scale-wise transformers for text-to-image synthesis. arXiv preprint arXiv:2412.01819, 2024

    2024 arXiv

  16. [24]

    How to train state-of-the-art models using torchvision’s latest primitives

    Vryniotis, V. How to train state-of-the-art models using torchvision’s latest primitives. PyTorch, 2021

    2021

  17. [25]

    Simplear: Pushing the frontier of autoregressive visual generation through pretraining, sft, and rl

    Wang, J., Tian, Z., Wang, X., Zhang, X., Huang, W., Wu, Z., and Jiang, Y.-G. Simplear: Pushing the frontier of autoregressive visual generation through pretraining, sft, and rl. arXiv preprint arXiv:2504.11455, 2025

    2025 arXiv

  18. [26]

    Dancegrpo: Unleashing grpo on visual generation, 2025

    Xue, Z., Wu, J., Gao, Y., Kong, F., Zhu, L., Chen, M., Liu, Z., Liu, W., Guo, Q., Huang, W., and Luo, P. Dancegrpo: Unleashing grpo on visual generation, 2025. URL https://arxiv.org/abs/2505.07818

    2025 arXiv

  19. [27]

    Using human feedback to fine-tune diffusion models without any reward model, 2024

    Yang, K., Tao, J., Lyu, J., Ge, C., Chen, J., Li, Q., Shen, W., Zhu, X., and Li, X. Using human feedback to fine-tune diffusion models without any reward model, 2024. URL https://arxiv.org/abs/2311.13231

    2024 arXiv

  20. [28]

    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 gl...

Pith tools

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