Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Ultra-High-Resolution Image Synthesis: Data, Method and Evaluation

T0 review · 3 major / 5 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 off-the-shelf latent diffusion models can be upgraded to direct photorealistic 4K generation with a scale-consistent VAE and a wavelet-based fine-tuning loss, backed by a new dataset and local texture metrics.

desk verdict The dataset and SC-VAE are real contributions, but the headline WLF mechanism reduces exactly to the standard flow loss, so the central fine-detail claim is unsupported. read the letter →

arxiv 2506.01331 v1 pith:EHH3YIQX submitted 2025-06-02 cs.CV

Jinjin Zhang , Qiuyu Huang , Junjie Liu , Xiefan Guo , Di Huang This is my paper
classification cs.CV
keywords 4Kimagesynthesislatentdiffusionmodelsscale-consistentVAEwavelet-basedfine-tuningAesthetic-4Kdatasettext-to-imagegenerationtexturequalitymetricsGLCMscore
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 tries to establish that ultra-high-resolution 4K image synthesis does not require purpose-built architectures: an off-the-shelf latent diffusion model can be adapted to generate photorealistic $4096\times4096$ images directly. The two adaptations are a VAE that compresses visual tokens twice as aggressively (down-sampling factor $F=16$ instead of $F=8$) while being calibrated against the original VAE's feature maps so the diffusion model's latent space is undisturbed, and a wavelet-based fine-tuning loss meant to concentrate learning on high-frequency detail. Alongside the method, the paper supplies Aesthetic-4K, a public dataset of captioned 4K images with training and evaluation splits, and two patch-level metrics, the GLCM Score and JPEG Compression Ratio, aimed at quantifying texture richness that holistic scores like FID miss. If the claims hold, the payoff is a generally applicable recipe: any latent diffusion model can be upgraded to 4K detail using the same two components and benchmarked against standardized ground.

What carries the argument

The two load-bearing components are named objects. SC-VAE is a variational auto-encoder with down-sampling factor 16 instead of the usual 8; its encoder stays frozen and only the decoder is fine-tuned under a scale-consistency loss, Eq. (1), that matches the up-sampled penultimate decoder feature map to the corresponding feature map of the original $F=8$ teacher VAE, together with reconstruction, KL, LPIPS, and patch-adversarial terms. WLF is the training objective in Eq. (10), which replaces the standard rectified-flow loss $\mathbb{E}\|u_t(x_t)-v_\theta(x_t,t)\|^2$ with $\mathbb{E}\|f(u_t(x_t))-f(v_\theta(x_t,t))\|^2$, where $f$ is the Haar discrete wavelet transform with low-pass $L=\frac{1}{\sqrt{2}}[1,1]$ and high-pass $H=\frac{1}{\sqrt{2}}[-1,1]$ filters, decomposing each latent into one low-frequency and three high-frequency sub-bands. The supporting objects are the Aesthetic-4K benchmark (105,288 GPT-4o-captioned training images plus a 2,781-image evaluation set) and two local metrics, the GLCM Score, the negative mean entropy of gray-level co-occurrence matrices over $64\times64$ patches, and the Compression Ratio, the ratio of raw image size to JPEG-95 compressed size.

What would settle it

Run the two losses, Eq. (9) and Eq. (10), on a fixed batch of latents with the paper's stated $L=\frac{1}{\sqrt{2}}[1,1]$ and $H=\frac{1}{\sqrt{2}}[-1,1]$ filters: if they agree to floating-point precision, the Table 9 gains cannot come from frequency reweighting and must be traced to unstated implementation details or to the SC-VAE and data changes in the same experiments. A complementary check is an ablation that scales the high-frequency sub-bands by a factor $\lambda>1$: a genuine frequency mechanism should make quality improve with $\lambda$.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that direct photorealistic 4K text-to-image synthesis is within reach of existing latent diffusion models rather than requiring new architectures. The vehicle is Diffusion-4K, built from SC-VAE, which halves latent resolution to a down-sampling factor of $F=16$ and is fine-tuned by matching its penultimate decoder feature maps, up-sampled, against those of the original $F=8$ teacher VAE so the pretrained latent space is preserved, and WLF, a fine-tuning objective in which the rectified-flow velocity loss is applied after a discrete Haar wavelet transform splits each latent into low- and high-frequency sub-bands. Trained on the new Aesthetic-4K data, SD3-2B and Flux-12B generate images at $4096\times4096$ that the paper reports to be superior to its own prior work and to the 4K specialists PixArt-$\Sigma$ and Sana in structured textures and fine details, as measured by FID, Aesthetics, CLIPScore, and the proposed GLCM Score and Compression Ratio.

Load-bearing premise

The load-bearing premise is that the wavelet fine-tuning loss changes the training signal in favor of high-frequency detail, yet with the orthonormal Haar filters stated in the paper the loss in Eq. (10) equals the standard rectified-flow loss in Eq. (9), so the reported improvements rest on a mechanism the paper does not specify.

Editorial extensions

If this is right

  • SC-VAE brings 4K training within memory reach: SD3-2B uses 31.3 GB and Flux-12B 50.4 GB for a single $4096\times4096$ step on one A100, falling to 16.1 GB and 26.9 GB with CPU offload.
  • Because the encoder and latent space are never touched, the recipe should transfer to any latent diffusion model built on the same $F=8$ VAE without retraining from scratch.
  • Scaling the training set from 12,015 to 105,288 image-text pairs fails to move holistic scores like FID, Aesthetics, and CLIPScore, but consistently improves the two local metrics, indicating where 4K quality gains actually show up.
  • The new benchmark fills a genuine gap: only about 0.03% of the open LAION-Aesthetics V2 pool meets the 4K threshold, so Aesthetic-Eval gives direct-4K methods a shared yardstick for the first time.

Reading between the lines

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

  • The WLF objective as written in Eq. (10) is mathematically identical to the rectified-flow loss in Eq. (9): the stated Haar filters are orthonormal, so by Parseval's identity $\|f(u)-f(v)\|^2=\|u-v\|^2$, and the paper never identifies a deviation (sub-band weighting, coefficient cropping, or unnormalized filters) that would make the two losses differ. Any real gain from WLF therefore needs a mecha
  • If SC-VAE reconstructions at $F=16$ genuinely approach the $F=8$ teacher's quality, the same scale-consistency recipe should extend to 8K generation, where the teacher itself is computationally infeasible; this is a concrete stress test the paper does not run.
  • Since the two local metrics were the only ones that kept improving as training data grew, a natural extension is to use GLCM Score and Compression Ratio as early-stopping and data-curation signals in future ultra-high-resolution training runs, a use the paper does not explicitly propose.
  • The benchmark invites a controlled comparison the paper only partially performs: fine-tuning PixArt-$\Sigma$ and Sana on the same Aesthetic-Train-V2 data would isolate whether the reported advantage comes from the Diffusion-4K recipe or from the underlying base models and training data.
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 / 5 minor

Summary. The manuscript introduces Aesthetic-4K, a curated dataset of ultra-high-resolution image-text pairs for 4K image synthesis, and proposes Diffusion-4K, a fine-tuning framework combining SC-VAE (an F=16 VAE with scale-consistency regularization) and WLF (a wavelet-based objective for rectified-flow fine-tuning). The authors also propose two local evaluation metrics, GLCM Score and Compression Ratio, and evaluate fine-tuned SD3-2B and Flux-12B models at 4096x4096 resolution against prior direct 4K methods such as PixArt-Sigma and Sana. The central claim is that Diffusion-4K directly synthesizes photorealistic 4K images with richer textures and finer detail than existing approaches.

Significance. If the results held, the paper would provide both a public benchmark and a generally applicable fine-tuning recipe for 4K latent diffusion models. The SC-VAE component is supported by controlled ablations (Tab. 8) and appears to be a genuine improvement over the earlier partitioned VAE, and the Aesthetic-4K dataset is a potentially useful community resource. However, the WLF objective as written is mathematically identical to the standard rectified-flow loss, so the paper's principal methodological novelty and the Tab. 9 fine-detail gains have no stated mechanism. The quantitative support for the 4K superiority claim is also weak, since the main comparison against prior models is performed at 2048x2048 and the only 4K evaluation set contains 195 images.

major comments (3)
  1. [Section 4.2, Eq. (10); Tab. 9] The WLF objective is identical to the standard rectified-flow loss. With L=[1,1]/sqrt(2) and H=[-1,1]/sqrt(2), the 2D Haar DWT is an orthonormal linear map f, so ||f(u_t(x_t))-f(v_theta(x_t,t))||^2 = ||u_t(x_t)-v_theta(x_t,t)||^2 for every sample. Consequently L_wlf(theta) equals L_rf(theta) for all theta, and the gradients with respect to v_theta are identical; the described decomposition into LL, LH, HL, and HH sub-bands contains no frequency-dependent weighting. The improvements reported in Tab. 9 (e.g., GLCM Score 0.74 vs. 0.77) therefore cannot be mechanistically attributed to WLF and must stem from uncontrolled training differences or an unstated modification of Eq. (10). Please state the actual objective used, including any per-sub-band weights, and rerun the WLF ablations with that objective.
  2. [Section 6.2, Tab. 5; Section 3.2, Tab. 9] The claim of superior performance in ultra-high-resolution image synthesis is not supported by quantitative evidence at 4K. The only head-to-head comparison against PixArt-Sigma and Sana (Tab. 5) is performed at 2048x2048. The only 4K table (Tab. 9) compares SD3 fine-tuning variants without including these state-of-the-art baselines, and it uses Aesthetic-Eval@4096, which contains only 195 images. With such a small evaluation set, FID and CLIPScore estimates are noisy. Please provide 4K quantitative comparisons against all compared methods, or temper the conclusion that the method achieves superior 4K synthesis performance.
  3. [Section 5.1, Eq. (11); Tabs. 5 and 9] The GLCM Score as defined is s_glcm = -1/P * sum_p H(g_p), which is the negative of an average entropy and cannot produce the positive values (e.g., 0.75, 0.77) reported in the tables unless an additional normalization or a different sign convention is used. Since the GLCM Score is one of the two newly proposed metrics that underpin the evaluation claims, the definition must be corrected and the normalization stated; otherwise Tab. 2's correlation results cannot be interpreted.
minor comments (5)
  1. [Section 5.1] The word 'Regrading' should be 'Regarding'.
  2. [Section 6.4] 'To evaluation the effectiveness' should be 'To evaluate the effectiveness'.
  3. [Section 3.2] The sentence 'we deliberately exclude any samples collected from the Internet' is confusing because Aesthetic-Eval is itself sampled from the Internet-sourced LAION dataset; clarify that the exclusion refers to overlap with the training set.
  4. [Section 4.2] The Haar filters are written as row vectors L and H; please specify how they are applied in the 2D DWT (e.g., separable convolution with strides) so that the exact transform f is reproducible.
  5. [Section 6.2, Tab. 6] The column 'Time (s/step)' for inference is ambiguous; please state whether this is time per sampling step and report the total generation time.

Circularity Check

1 steps flagged · score 6.0 of 10

WLF objective Eq. (10) is mathematically identical to standard rectified-flow loss Eq. (9) because the Haar DWT is orthonormal, so Tab. 9's WLF gains have no mechanistic basis; dataset and SC-VAE remain independent.

  1. renaming known result [Sec. 4.2, Eqs. (9)-(10), page 6; ablation in Sec. 6.4, Tab. 9]
    "Lrf(θ) = E_{t,x0,ε} [ w_t || u_t(x_t) − v_θ(x_t,t)||^2 ], (9) ... Lwlf(θ) = E_{t,x0,ε} [ w_t || f(u_t(x_t)) − f(v_θ(x_t,t))||^2 ], (10) ... we utilize the Haar wavelet... L=1/√2[1,1] and H=1/√2[−1,1] denote the low-pass and high-pass filters, which are used to construct four kernels in DWT with a stride of 2, namely LL^T, LH^T, HL^T, HH^T."

    The filters L and H are orthonormal, so the 2D DWT f formed from LL^T, LH^T, HL^T, HH^T with stride 2 is an orthogonal linear map in the standard periodic-boundary implementation. Therefore ||f(a)-f(b)||^2 = ||a-b||^2 for any a,b. Substituting a=u_t(x_t) and b=v_θ(x_t,t) makes Eq. (10) exactly Eq. (9) for every sample and every θ; the gradients with respect to v_θ are identical. Because Eq. (10) applies the same f to both target and prediction and writes no per-subband weights, the described decomposition into LL/LH/HL/HH cannot change the optimization trajectory. Hence the WLF objective is the standard rectified-flow loss expressed in wavelet coordinates, and Tab.

full rationale

The dataset construction, SC-VAE scale-consistency loss, and the GLCM/Compression-Ratio metrics are not circular: the metrics are validated against independent human ratings (Tab. 2), and SC-VAE is compared against vanilla fine-tuning without scale consistency (Tab. 8). The one load-bearing reduction is WLF. The paper presents Eq. (10) as a new objective that 'emphasizes high-frequency components,' but because the Haar DWT is orthonormal, Eq. (10) is identical to the standard rectified-flow loss Eq. (9) by construction. Consequently the ablation in Tab. 9 cannot provide evidence for WLF as a distinct training signal; the difference between SD3-F16-FT and SD3-F16-WLF must be due to other factors. This is a partial, not total, circularity: SC-VAE and Aesthetic-4K remain independent contributions, so the overall score is 6 rather than 8 or 10.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central framework rests on hand-set loss weights, unvalidated metric design choices, and the false premise that the Haar-DWT loss is a new objective. No invented physical entities are introduced.

free parameters (4)
  • SC-VAE loss weights (lambda_sc, lambda_lpips, lambda_adv) = 1.0, 0.1, 0.05
    Set by hand in Sec 6.1; the reported reconstruction and generation improvements depend on these weights.
  • GLCM Score patch configuration = patch size 64, 64 gray levels, delta=[1,2,3,4], theta=[0,45,90,135]
    Metric design choices in Sec 5.1; no principled selection, affects all GLCM Score comparisons.
  • Compression Ratio JPEG quality = 95
    Chosen in Sec 5.1; the ratio is sensitive to this setting.
  • Aesthetic-Eval resolution threshold = short side > 2048 pixels
    Defines the benchmark subset in Sec 3.2; results on Aesthetic-Eval@2048 are the main comparisons.
assumptions (5)
  • standard math Haar DWT is an orthonormal transform and therefore norm-preserving
    Used implicitly in Eq (10) when replacing the flow-matching loss with a wavelet-domain loss; if the transform is orthonormal the two losses coincide, which the paper does not acknowledge.
  • domain assumption LAION-Aesthetics V2 6.5+ aesthetic scores are valid quality priors for selecting benchmark images
    Sec 3.2 uses this to build Aesthetic-Eval; errors in these scores would propagate to benchmark validity.
  • domain assumption Fine-tuning only the VAE decoder preserves latent-space consistency with the base diffusion model
    Sec 4.1 states this to ensure compatibility with various diffusion models; no empirical verification of negligible distribution shift is provided beyond reconstruction metrics.
  • ad hoc to paper GLCM entropy and JPEG compression ratio reflect human perception of texture richness and fine detail
    Sec 5.1 asserts this; validation uses only five participants and an unspecified patch sample, so the alignment claim is weakly supported.
  • domain assumption Human preference ratings from ten participants are representative
    Sec 6.2 preference study averages scores from ten raters; no inter-rater agreement is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ultra-High-Resolution Image Synthesis: Data, Method and Evaluation." pith.science (2026). https://pith.science/paper/EHH3YIQX

@misc{pith2026250601331,
  author       = {Pith},
  title        = {Pith review of: Ultra-High-Resolution Image Synthesis: Data, Method and Evaluation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EHH3YIQX}},
  note         = {Machine review of arXiv:2506.01331}
}
read the original abstract

Ultra-high-resolution image synthesis holds significant potential, yet remains an underexplored challenge due to the absence of standardized benchmarks and computational constraints. In this paper, we establish Aesthetic-4K, a meticulously curated dataset containing dedicated training and evaluation subsets specifically designed for comprehensive research on ultra-high-resolution image synthesis. This dataset consists of high-quality 4K images accompanied by descriptive captions generated by GPT-4o. Furthermore, we propose Diffusion-4K, an innovative framework for the direct generation of ultra-high-resolution images. Our approach incorporates the Scale Consistent Variational Auto-Encoder (SC-VAE) and Wavelet-based Latent Fine-tuning (WLF), which are designed for efficient visual token compression and the capture of intricate details in ultra-high-resolution images, thereby facilitating direct training with photorealistic 4K data. This method is applicable to various latent diffusion models and demonstrates its efficacy in synthesizing highly detailed 4K images. Additionally, we propose novel metrics, namely the GLCM Score and Compression Ratio, to assess the texture richness and fine details in local patches, in conjunction with holistic measures such as FID, Aesthetics, and CLIPScore, enabling a thorough and multifaceted evaluation of ultra-high-resolution image synthesis. Consequently, Diffusion-4K achieves impressive performance in ultra-high-resolution image synthesis, particularly when powered by state-of-the-art large-scale diffusion models (eg, Flux-12B). The source code is publicly available at https://github.com/zhang0jhon/diffusion-4k.

Figures

Figures reproduced from arXiv: 2506.01331 by the authors.

Figure 1
Figure 1. Histogram comparisons of image height and width [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustration of image-text samples in the Aesthetic-4K dataset, which includes high-quality images and precise text [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The framework of the proposed SC-VAE. Our method shares the same latent space as the pre-trained latent diffusion [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Qualitative analysis of GLCM Score↑ / Compression Ratio↓. The top and bottom images are generated using the same prompts and random seed, but with different mod￾els. Our indicators demonstrate a strong alignment with human-centric perceptual cognition of rich textures …
Figure 5
Figure 5. Figure 5: Qualitative reconstruction comparisons of ultra-high [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Qualitative generation comparisons of ultra-high [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Qualitative results synthesized by our Diffusion-4K, emphasizing exceptional fine details in the generated 4K images. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Qualitative results on Aesthetic-Eval@2048 at 2048 × 2048 resolution, including our proposed Diffusion-4K, PixArt [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: We present qualitative comparisons with PixArt- [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Human and GPT-4o preference evaluation. Our [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Qualitative scalability results on Aesthetic-Eval@2048 at [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Qualitative ablation study on WLF. The image in [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. UltraFlux: Data-Model Co-Design for High-quality Native 4K Text-to-Image Generation across Diverse Aspect Ratios

    cs.CV 2025-11 conditional novelty 5.0 of 10

    A data-model co-designed Flux DiT with resonance RoPE, post-trained F16 VAE, SNR-aware Huber wavelet loss, and aesthetic curriculum claims state-of-the-art open-source 4K multi-AR text-to-image generation.

Reference graph

Works this paper leans on

89 extracted references · 70 canonical work pages · cited by 1 Pith paper

  1. [1]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P . Abbeel, “Denoising diffusion probabilistic models,” inAdvances in Neural Information Processing Systems, 2020, pp. 6840–6851

    2020

  2. [2]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” inInternational Conference on Learning Representations, 2021, pp. 1–20

    2021

  3. [3]

    Score-based generative modeling through stochas- tic differential equations,

    Y. Song, J. Sohl-Dickstein, D. P . Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochas- tic differential equations,” inInternational Conference on Learning Representations, 2021, pp. 1–36

    2021

  4. [4]

    Improved denoising diffusion probabilistic models,

    A. Q. Nichol and P . Dhariwal, “Improved denoising diffusion probabilistic models,” inInternational Conference on Machine Learn- ing, 2021, pp. 8162–8171

    2021

  5. [5]

    Maximum like- lihood training of score-based diffusion models,

    Y. Song, C. Durkan, I. Murray, and S. Ermon, “Maximum like- lihood training of score-based diffusion models,” inAdvances in Neural Information Processing Systems, 2021, pp. 1415–1428

    2021

  6. [6]

    Score-based generative mod- eling in latent space,

    A. Vahdat, K. Kreis, and J. Kautz, “Score-based generative mod- eling in latent space,” inAdvances in Neural Information Processing Systems, 2021, pp. 11 287–11 302

    2021

  7. [7]

    Diffusion models beat gans on image synthesis,

    P . Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,” inAdvances in Neural Information Processing Systems, 2021, pp. 8780–8794

    2021

  8. [8]

    Elucidating the design space of diffusion-based generative models,

    T. Karras, M. Aittala, T. Aila, and S. Laine, “Elucidating the design space of diffusion-based generative models,” inAdvances in Neural Information Processing Systems, 2022, pp. 26 565–26 577

    2022

Show all 89 references
  1. [9]

    GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models,

    A. Q. Nichol, P . Dhariwal, A. Ramesh, P . Shyam, P . Mishkin, B. Mc- grew, I. Sutskever, and M. Chen, “GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models,” inInternational Conference on Machine Learning, 2022, pp. 16 784– 16 804

    2022

  2. [10]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,” 2022, arXiv:2207.12598

    2022 arXiv

  3. [11]

    Multimodal image synthesis and edit- ing: The generative ai era,

    F. Zhan, Y. Yu, R. Wu, J. Zhang, S. Lu, L. Liu, A. Kortylewski, C. Theobalt, and E. Xing, “Multimodal image synthesis and edit- ing: The generative ai era,”IEEE T ransactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 12, pp. 15 098–15 119, 2023

    2023

  4. [12]

    Diffusion models: A comprehensive survey of methods and applications,

    L. Yang, Z. Zhang, Y. Song, S. Hong, R. Xu, Y. Zhao, W. Zhang, B. Cui, and M.-H. Yang, “Diffusion models: A comprehensive survey of methods and applications,”ACM Computing Surveys, vol. 56, no. 4, pp. 1–39, 2023

    2023

  5. [13]

    Scalable diffusion models with transform- ers,

    W. Peebles and S. Xie, “Scalable diffusion models with transform- ers,” inInternational Conference on Computer Vision, 2023, pp. 4195– 4205

    2023

  6. [14]

    Diffusion models in vision: A survey,

    F.-A. Croitoru, V . Hondru, R. T. Ionescu, and M. Shah, “Diffusion models in vision: A survey,”IEEE T ransactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 9, pp. 10 850–10 869, 2023

    2023

  7. [15]

    A diffusion model translator for efficient image-to-image translation,

    M. Xia, Y. Zhou, R. Yi, Y.-J. Liu, and W. Wang, “A diffusion model translator for efficient image-to-image translation,”IEEE T ransactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 10 272–10 283, 2024

    2024

  8. [16]

    Photorealistic text-to-image diffusion models with deep lan- guage understanding,

    C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. L. Denton, K. Ghasemipour, R. Gontijo Lopes, B. Karagol Ayan, T. Salimans et al., “Photorealistic text-to-image diffusion models with deep lan- guage understanding,” inAdvances in Neural Information Processing Systems, 2022, ...

    2022

  9. [17]

    Imagen 3,

    J. Baldridge, J. Bauer, M. Bhutani, N. Brichtova, A. Bunner, L. Cas- trejon, K. Chan, Y. Chen, S. Dieleman, Y. Duet al., “Imagen 3,” 2024,arXiv:2408.07009

    2024

  10. [18]

    Hier- archical text-conditional image generation with clip latents,

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

    2022 arXiv

  11. [19]

    Improv- ing image generation with better captions,

    J. Betker, G. Goh, L. Jing, T. Brooks, J. Wang, L. Li, L. Ouyang, J. Zhuang, J. Lee, Y. Guoet al., “Improv- ing image generation with better captions,”Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, vol. 2, no. 3, p. 8, 2023

    2023

  12. [20]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P . Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” inIEEE Conference on Computer Vision and Pattern Recognition, 2022, pp. 10 684–10 695

    2022

  13. [21]

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

    D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. M ¨uller, J. Penna, and R. Rombach, “SDXL: Improving latent diffusion models for high-resolution image synthesis,” inInter- national Conference on Learning Representations, 2024, pp. 1–13

    2024

  14. [22]

    Scaling rectified flow transformers for high-resolution image synthesis,

    P . Esser, S. Kulal, A. Blattmann, R. Entezari, J. M ¨uller, H. Saini, Y. Levi, D. Lorenz, A. Sauer, F. Boeselet al., “Scaling rectified flow transformers for high-resolution image synthesis,” inInternational Conference on Machine Learning, 2024, pp. 12 606–12 633

    2024

  15. [23]

    PixArt-Σ: Weak-to-strong training of diffusion trans- former for 4k text-to-image generation,

    J. Chen, C. Ge, E. Xie, Y. Wu, L. Yao, X. Ren, Z. Wang, P . Luo, H. Lu, and Z. Li, “PixArt-Σ: Weak-to-strong training of diffusion trans- former for 4k text-to-image generation,” inEuropean Conference on Computer Vision, 2024, pp. 74–91

    2024

  16. [24]

    Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation,

    D. Li, A. Kamko, E. Akhgari, A. Sabet, L. Xu, and S. Doshi, “Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation,” 2024,arXiv:2402.17245

    2024 arXiv

  17. [25]

    Playground v3: Improving text-to- image alignment with deep-fusion large language models,

    B. Liu, E. Akhgari, A. Visheratin, A. Kamko, L. Xu, S. Shrirao, J. Souza, S. Doshi, and D. Li, “Playground v3: Improving text-to- image alignment with deep-fusion large language models,” 2024, arXiv:2409.10695

    2024 arXiv

  18. [26]

    Diffusion-4k: Ultra-high-resolution image synthesis with latent diffusion mod- els,

    J. Zhang, Q. Huang, J. Liu, X. Guo, and D. Huang, “Diffusion-4k: Ultra-high-resolution image synthesis with latent diffusion mod- els,” inIEEE Conference on Computer Vision and Pattern Recognition, 2025, pp. 1–17

    2025

  19. [27]

    (2024) Flux

    Black Forest Labs. (2024) Flux. [Online]. Available: https: //github.com/black-forest-labs/flux

    2024

  20. [28]

    Flow matching for generative modeling,

    Y. Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inInternational Conference on Learning Representations, 2023, pp. 1–28

    2023

  21. [29]

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

    X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,” inInternational Conference on Learning Representations, 2023, pp. 1–33

    2023

  22. [30]

    Building normalizing flows with stochastic interpolants,

    M. S. Albergo and E. Vanden-Eijnden, “Building normalizing flows with stochastic interpolants,” inInternational Conference on Learning Representations, 2023, pp. 1–29

    2023

  23. [31]

    Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers,

    N. Ma, M. Goldstein, M. S. Albergo, N. M. Boffi, E. Vanden- Eijnden, and S. Xie, “Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers,” in European Conference on Computer Vision, 2024, pp. 23–40

    2024

  24. [32]

    Sana: Efficient high-resolution image synthesis with lin- ear diffusion transformers,

    E. Xie, J. Chen, J. Chen, H. Cai, Y. Lin, Z. Zhang, M. Li, Y. Lu, and S. Han, “Sana: Efficient high-resolution image synthesis with lin- ear diffusion transformers,” inInternational Conference on Learning Representations, 2025, pp. 1–25

    2025

  25. [33]

    Sana 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer,

    E. Xie, J. Chen, Y. Zhao, J. Yu, L. Zhu, Y. Lin, Z. Zhang, M. Li, J. Chen, H. Caiet al., “Sana 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer,” 2025, arXiv:2501.18427

    2025 arXiv

  26. [34]

    GPT-4o system card,

    A. Hurst, A. Lerer, A. P . Goucher, A. Perelman, A. Ramesh, A. Clark, A. Ostrow, A. Welihinda, A. Hayes, A. Radfordet al., “GPT-4o system card,” 2024,arXiv:2410.21276

    2024 arXiv

  27. [35]

    Gans trained by a two time-scale update rule converge 15 to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochre- iter, “Gans trained by a two time-scale update rule converge 15 to a local nash equilibrium,” inAdvances in Neural Information Processing Systems, 2017, pp. 1–12

    2017

  28. [36]

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

    C. Schuhmann, R. Beaumont, R. Vencu, C. Gordon, R. Wightman, M. Cherti, T. Coombes, A. Katta, C. Mullis, M. Wortsmanet al., “Laion-5b: An open large-scale dataset for training next generation image-text models,” inAdvances in Neural Information Processing Systems, 2022, pp. 25...

    2022

  29. [37]

    Clip- score: A reference-free evaluation metric for image captioning,

    J. Hessel, A. Holtzman, M. Forbes, R. L. Bras, and Y. Choi, “Clip- score: A reference-free evaluation metric for image captioning,” in Empirical Methods in Natural Language Processing, 2021, pp. 7514– 7528

    2021

  30. [38]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y. Loet al., “Segment anything,” inInternational Conference on Computer Vision, 2023, pp. 4015–4026

    2023

  31. [39]

    Taming transformers for high-resolution image synthesis,

    P . Esser, R. Rombach, and B. Ommer, “Taming transformers for high-resolution image synthesis,” inIEEE Conference on Computer Vision and Pattern Recognition, 2021, pp. 12 873–12 883

    2021

  32. [40]

    Auto-encoding variational bayes,

    D. P . Kingma, M. Wellinget al., “Auto-encoding variational bayes,” inInternational Conference on Learning Representations, 2014, pp. 1– 14

    2014

  33. [41]

    Neural discrete representation learning,

    A. Van Den Oord, O. Vinyalset al., “Neural discrete representation learning,” inAdvances in Neural Information Processing Systems, 2017, pp. 1–10

    2017

  34. [42]

    PixArt-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis,

    J. Chen, J. Yu, C. Ge, L. Yao, E. Xie, Y. Wu, Z. Wang, J. Kwok, P . Luo, H. Luet al., “PixArt-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis,” inInternational Conference on Learning Representations, 2024, pp. 1–30

    2024

  35. [43]

    Diffit: Diffusion vision transformers for image generation,

    A. Hatamizadeh, J. Song, G. Liu, J. Kautz, and A. Vahdat, “Diffit: Diffusion vision transformers for image generation,” inEuropean Conference on Computer Vision, 2024, pp. 37–55

    2024

  36. [44]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Systems, 2017, pp. 1–11

    2017

  37. [45]

    Roformer: Enhanced transformer with rotary position embedding,

    J. Su, M. Ahmed, Y. Lu, S. Pan, W. Bo, and Y. Liu, “Roformer: Enhanced transformer with rotary position embedding,”Neuro- computing, vol. 568, p. 127063, 2024

    2024

  38. [46]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agar- wal, G. Sastry, A. Askell, P . Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” in International Conference on Machine Learning, 2021, pp. 8748–8763

    2021

  39. [47]

    Exploring the limits of transfer learning with a unified text-to-text transformer,

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y. Zhou, W. Li, and P . J. Liu, “Exploring the limits of transfer learning with a unified text-to-text transformer,”Journal of Machine Learning Research, vol. 21, no. 140, pp. 1–67, 2020

    2020

  40. [48]

    Gemma 2: Improving open language models at a practical size,

    G. Team, M. Riviere, S. Pathak, P . G. Sessa, C. Hardin, S. Bhu- patiraju, L. Hussenot, T. Mesnard, B. Shahriari, A. Ram ´eet al., “Gemma 2: Improving open language models at a practical size,” 2024,arXiv:2408.00118

    2024 arXiv

  41. [49]

    Generative adver- sarial nets,

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde- Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adver- sarial nets,” inAdvances in Neural Information Processing Systems, 2014, pp. 1–9

    2014

  42. [50]

    Progressive growing of gans for improved quality, stability, and variation,

    T. Karras, T. Aila, S. Laine, and J. Lehtinen, “Progressive growing of gans for improved quality, stability, and variation,” inInterna- tional Conference on Learning Representations, 2018, pp. 1–26

    2018

  43. [51]

    Large scale gan training for high fidelity natural image synthesis,

    A. Brock, J. Donahue, and K. Simonyan, “Large scale gan training for high fidelity natural image synthesis,” inInternational Confer- ence on Learning Representations, 2019, pp. 1–35

    2019

  44. [52]

    A style-based generator architec- ture for generative adversarial networks,

    T. Karras, S. Laine, and T. Aila, “A style-based generator architec- ture for generative adversarial networks,” inIEEE Conference on Computer Vision and Pattern Recognition, 2019, pp. 4401–4410

    2019

  45. [53]

    Analyzing and improving the image quality of stylegan,

    T. Karras, S. Laine, M. Aittala, J. Hellsten, J. Lehtinen, and T. Aila, “Analyzing and improving the image quality of stylegan,” inIEEE Conference on Computer Vision and Pattern Recognition, 2020, pp. 8110–8119

    2020

  46. [54]

    Alias-free generative adversarial networks,

    T. Karras, M. Aittala, S. Laine, E. H ¨ark¨onen, J. Hellsten, J. Lehti- nen, and T. Aila, “Alias-free generative adversarial networks,” in Advances in Neural Information Processing Systems, 2021, pp. 852– 863

    2021

  47. [55]

    Any- resolution training for high-resolution image synthesis,

    L. Chai, M. Gharbi, E. Shechtman, P . Isola, and R. Zhang, “Any- resolution training for high-resolution image synthesis,” inEuro- pean conference on computer vision, 2022, pp. 170–188

    2022

  48. [56]

    Scaling up gans for text-to-image synthesis,

    M. Kang, J.-Y. Zhu, R. Zhang, J. Park, E. Shechtman, S. Paris, and T. Park, “Scaling up gans for text-to-image synthesis,” inIEEE Conference on Computer Vision and Pattern Recognition, 2023, pp. 10 124–10 134

    2023

  49. [57]

    Generating diverse high-fidelity images with vq-vae-2,

    A. Razavi, A. Van den Oord, and O. Vinyals, “Generating diverse high-fidelity images with vq-vae-2,” inAdvances in Neural Informa- tion Processing Systems, 2019, pp. 1–11

    2019

  50. [58]

    Zero-shot text-to-image generation,

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

    2021

  51. [59]

    Muse: Text-to-image generation via masked generative trans- formers,

    H. Chang, H. Zhang, J. Barber, A. Maschinot, J. Lezama, L. Jiang, M.-H. Yang, K. Murphy, W. T. Freeman, M. Rubinsteinet al., “Muse: Text-to-image generation via masked generative trans- formers,” inInternational Conference on Machine Learning, 2023, pp. 4055–4075

    2023

  52. [60]

    Scaling autoregressive models for content-rich text-to-image generation,

    J. Yu, Y. Xu, J. Y. Koh, T. Luong, G. Baid, Z. Wang, V . Vasudevan, A. Ku, Y. Yang, B. K. Ayanet al., “Scaling autoregressive models for content-rich text-to-image generation,”T ransactions on Machine Learning Research, vol. 2, no. 3, p. 5, 2022

    2022

  53. [61]

    Visual au- toregressive modeling: Scalable image generation via next-scale prediction,

    K. Tian, Y. Jiang, Z. Yuan, B. Peng, and L. Wang, “Visual au- toregressive modeling: Scalable image generation via next-scale prediction,” inAdvances in Neural Information Processing Systems, 2024, pp. 84 839–84 865

    2024

  54. [62]

    Autoregressive image generation without vector quantization,

    T. Li, Y. Tian, H. Li, M. Deng, and K. He, “Autoregressive image generation without vector quantization,” inAdvances in Neural Information Processing Systems, 2024, pp. 56 424–56 445

    2024

  55. [63]

    Adversarial diffusion distillation,

    A. Sauer, D. Lorenz, A. Blattmann, and R. Rombach, “Adversarial diffusion distillation,” inEuropean Conference on Computer Vision, 2024, pp. 87–103

    2024

  56. [64]

    Fast high-resolution image synthesis with latent adversarial diffusion distillation,

    A. Sauer, F. Boesel, T. Dockhorn, A. Blattmann, P . Esser, and R. Rombach, “Fast high-resolution image synthesis with latent adversarial diffusion distillation,” inSIGGRAPH Asia, 2024, pp. 1–11

    2024

  57. [65]

    Multidiffusion: Fusing diffusion paths for controlled image generation,

    O. Bar-Tal, L. Yariv, Y. Lipman, and T. Dekel, “Multidiffusion: Fusing diffusion paths for controlled image generation,” inInter- national Conference on Machine Learning, 2023, pp. 1737–1752

    2023

  58. [66]

    Demo- fusion: Democratising high-resolution image generation with no $$$,

    R. Du, D. Chang, T. Hospedales, Y.-Z. Song, and Z. Ma, “Demo- fusion: Democratising high-resolution image generation with no $$$,” inIEEE Conference on Computer Vision and Pattern Recognition, 2024, pp. 6159–6168

    2024

  59. [67]

    Elasticdiffusion: Training-free arbitrary size image generation through global-local content separation,

    M. Haji-Ali, G. Balakrishnan, and V . Ordonez, “Elasticdiffusion: Training-free arbitrary size image generation through global-local content separation,” inIEEE Conference on Computer Vision and Pattern Recognition, 2024, pp. 6603–6612

    2024

  60. [68]

    W ¨urstchen: An efficient architecture for large-scale text-to-image diffusion models,

    P . Pernias, D. Rampas, M. L. Richter, C. J. Pal, and M. Aubreville, “W ¨urstchen: An efficient architecture for large-scale text-to-image diffusion models,” inInternational Conference on Learning Represen- tations, 2024, pp. 1–13

    2024

  61. [69]

    Deep compression autoencoder for efficient high- resolution diffusion models,

    J. Chen, H. Cai, J. Chen, E. Xie, S. Yang, H. Tang, M. Li, Y. Lu, and S. Han, “Deep compression autoencoder for efficient high- resolution diffusion models,” inInternational Conference on Learning Representations, 2025, pp. 1–22

    2025

  62. [70]

    The unreasonable effectiveness of deep features as a perceptual met- ric,

    R. Zhang, P . Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual met- ric,” inIEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 586–595

    2018

  63. [71]

    Image-to-image trans- lation with conditional adversarial networks,

    P . Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, “Image-to-image trans- lation with conditional adversarial networks,” inIEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 1125–1134

    2017

  64. [72]

    Wavelet score-based generative modeling,

    F. Guth, S. Coste, V . De Bortoli, and S. Mallat, “Wavelet score-based generative modeling,” inAdvances in Neural Information Processing Systems, 2022, pp. 478–491

    2022

  65. [73]

    Wavelet diffusion models are fast and scalable image generators,

    H. Phung, Q. Dao, and A. Tran, “Wavelet diffusion models are fast and scalable image generators,” inIEEE Conference on Computer Vision and Pattern Recognition, 2023, pp. 10 199–10 208

    2023

  66. [74]

    Imagereward: Learning and evaluating human preferences for text-to-image generation,

    J. Xu, X. Liu, Y. Wu, Y. Tong, Q. Li, M. Ding, J. Tang, and Y. Dong, “Imagereward: Learning and evaluating human preferences for text-to-image generation,” inAdvances in Neural Information Pro- cessing Systems, 2024, pp. 15 903–15 935

    2024

  67. [75]

    Textons, the elements of texture perception, and their interactions,

    B. Julesz, “Textons, the elements of texture perception, and their interactions,”Nature, vol. 290, no. 5802, pp. 91–97, 1981

    1981

  68. [76]

    Stockwell,T exture-a cognitive aesthetics of reading

    P . Stockwell,T exture-a cognitive aesthetics of reading. Edinburgh University Press, 2020

    2020

  69. [77]

    Early vision and texture percep- tion,

    J. R. Bergen and E. H. Adelson, “Early vision and texture percep- tion,”Nature, vol. 333, no. 6171, pp. 363–364, 1988

    1988

  70. [78]

    Image quality analysis using glcm,

    D. Gadkari, “Image quality analysis using glcm,” Master’s thesis, University of Central Florida, 2004

    2004

  71. [79]

    Textural features for image classification,

    R. M. Haralick, K. Shanmugam, and I. H. Dinstein, “Textural features for image classification,”IEEE T ransactions on Systems, Man, and Cybernetics, no. 6, pp. 610–621, 1973. 16

    1973

  72. [80]

    Musiq: Multi- scale image quality transformer,

    J. Ke, Q. Wang, Y. Wang, P . Milanfar, and F. Yang, “Musiq: Multi- scale image quality transformer,” inInternational Conference on Computer Vision, 2021, pp. 5148–5157

    2021

  73. [81]

    Maniqa: Multi-dimension attention network for no- reference image quality assessment,

    S. Yang, T. Wu, S. Shi, S. Lao, Y. Gong, M. Cao, J. Wang, and Y. Yang, “Maniqa: Multi-dimension attention network for no- reference image quality assessment,” inIEEE Conference on Com- puter Vision and Pattern Recognition, 2022, pp. 1191–1200

    2022

  74. [82]

    Pick-a-pic: An open dataset of user preferences for text-to-image generation,

    Y. Kirstain, A. Polyak, U. Singer, S. Matiana, J. Penna, and O. Levy, “Pick-a-pic: An open dataset of user preferences for text-to-image generation,” inAdvances in Neural Information Processing Systems, 2023, pp. 36 652–36 663

    2023

  75. [83]

    Human preference score: Better aligning text-to-image models with human prefer- ence,

    X. Wu, K. Sun, F. Zhu, R. Zhao, and H. Li, “Human preference score: Better aligning text-to-image models with human prefer- ence,” inIEEE Conference on Computer Vision and Pattern Recogni- tion, 2023, pp. 2096–2105

    2023

  76. [84]

    Rich human feedback for text-to-image generation,

    Y. Liang, J. He, G. Li, P . Li, A. Klimovskiy, N. Carolan, J. Sun, J. Pont-Tuset, S. Young, F. Yanget al., “Rich human feedback for text-to-image generation,” inIEEE Conference on Computer Vision and Pattern Recognition, 2024, pp. 19 401–19 411

    2024

  77. [85]

    Decoupled weight decay regulariza- tion,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regulariza- tion,” inInternational Conference on Learning Representations, 2019, pp. 1–18

    2019

  78. [86]

    Zero: Memory optimizations toward training trillion parameter models,

    S. Rajbhandari, J. Rasley, O. Ruwase, and Y. He, “Zero: Memory optimizations toward training trillion parameter models,” inIn- ternational Conference for High Performance Computing, Networking, Storage and Analysis, 2020, pp. 1–16

    2020

  79. [87]

    Zero-Offload: Democratizing billion- scale model training,

    J. Ren, S. Rajbhandari, R. Y. Aminabadi, O. Ruwase, S. Yang, M. Zhang, D. Li, and Y. He, “Zero-Offload: Democratizing billion- scale model training,” inUSENIX Annual T echnical Conference, 2021, pp. 551–564

    2021

  80. [88]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P . Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE T ransactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004

    2004

  81. [89]

    (2024) Sora

    OpenAI. (2024) Sora. [Online]. Available: https://openai.com/ index/video-generation-models-as-world-simulators

    2024

Pith tools

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