REVIEW 4 major objections 6 minor 3 cited by
Native-Resolution Image Synthesis
T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read The paper argues that training a diffusion transformer directly on images at their native resolutions and aspect ratios yields a single model that is state-of-the-art at both standard square benchmarks and capable of zero-shot generation…
desk verdict A real single-model resolution-generalization result, but the paper's explanation of why it works is ahead of its ablations. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is variable-length packed sequence processing inside a diffusion transformer. Each native-resolution image contributes a number of visual tokens equal to its latent height times width, and images are concatenated into one sequence with cumulative sequence lengths that tell a variable-length attention kernel to attend only within each image, eliminating padding and masking. Axial 2D rotary position embeddings factorize each token's position into separate height and width rotation frequencies, so relative 2D position is preserved across resolution changes, while Packed Adaptive Layer Normalization broadcasts per-image class or text conditioning to the right tokens. Together these make the model's effective computation independent of a fixed grid, which is what the paper claims lets it generate at unseen resolutions and aspect ratios.
What would settle it
Train the same NiT architecture on a fixed mixture of only 256x256 and 512x512 images for enough steps to match the native-resolution model's training budget; if that model also reaches FID near 4 on 1024x1024 or near 4 on 9:16, the resolution-generalization result is not caused by native-resolution data. Conversely, train a fixed-resolution architecture such as DiT with absolute positional embeddings on the native-resolution mixture; if it too generalizes, the effect is data-driven rather than architectural.
Extended reading notes
Core claim
On its own terms, the discovery is that a diffusion transformer trained exclusively on images at their native resolution learns a scale- and aspect-ratio-invariant visual distribution, so one model covers many output formats. NiT encodes each image with a 32x downsampling autoencoder, patches the latents into a token count set by the image's own height and width, packs all tokens in a batch up to a fixed token budget, and runs full attention inside each image through cumulative sequence lengths with a variable-length attention kernel. The 2D axial rotary position embedding gives tokens relative positional information along height and width separately, which is what lets the network keep spatial structure when the grid changes. The result is a single model with FID 2.03 at 256x256 and 1.45 at 512x512, and zero-shot FID 4.52 at 1024x1024 and 6.51 at 1536x1536, alongside strong scores on unseen aspect ratios; the authors read this as evidence that the model has internalized the intrinsic visual distribution rather than a fixed format.
Load-bearing premise
The claim rests on the assumption that native-resolution training data, rather than the new positional encoding and packing scheme, is what produces the model's ability to generalize to unseen resolutions.
Editorial extensions
If this is right
- A single NiT model can replace two resolution-specific models on the standard ImageNet benchmarks, cutting cumulative training compute: the 131B-token run reports an mFID of 1.86 across 256x256 and 512x512 while using about 9.17% of the token budget of the DiT-XL/SiT-XL pair.
- A model trained only on ImageNet's native resolutions generalizes to unseen 1024x1024 and 1536x1536 with FID 4.52 and 6.51, whereas resolution-expert baselines collapse (FID 40.74 for EDM2-L at 1024, and FiTv2 and SiT-REPA fail to produce valid 1024x1024 images).
- Unseen aspect ratios from 1:3 to 3:1 generate with FID between 3.41 and 16.85, with the best scores near square ratios, indicating the model does not inherit the truncation bias that cropping to square training images induces.
- The same architecture carries to text-to-image: NiT-T2I reports FID 9.18 and a CLIP score of 0.345 on zero-shot COCO, surpassing SD-v1.5 and SDXL-Turbo at a fraction of the model size.
Reading between the lines
- Editorial extension: the ablation in Table 4 changes the data mixture but keeps the NiT architecture fixed, so the paper does not isolate how much of the generalization comes from native-resolution data versus from 2D RoPE and packing; a factorial study with a fixed-resolution architecture on native data would separate the two.
- Editorial extension: because the model packs to a token budget rather than a batch of fixed images, the natural scaling currency for this paradigm is total tokens; this suggests resolution diversity in the training set, not just dataset size, may be the resource that buys zero-shot generalization.
- Editorial extension: the 32x downsampling autoencoder has a fixed receptive field per latent token, so the sharp FID degradation at 2048x2048 (24.76) may be an encoder or context limit rather than a fundamental limit of native-resolution training; testing with a higher-capacity encoder or sliding-window decoding would locate the bottleneck.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper introduces native-resolution image synthesis, a paradigm for diffusion transformers that trains directly on images at their original resolutions and aspect ratios without resizing, cropping, or padding. The proposed architecture, NiT, tokenizes variable-size latents into packed sequences, uses FlashAttention with cumulative sequence lengths for per-instance attention, applies an axial 2D Rotary Position Embedding, and employs a packed adaptive layer normalization. The authors report that a single NiT-XL model achieves ImageNet FID 2.03 at 256x256 and 1.45 at 512x512, outperforming several resolution-expert baselines while using less total training compute, and that it generalizes zero-shot to unseen resolutions up to 1536x1536 and to non-square aspect ratios, with FID 4.52 at 1024x1024 and 4.11 at 432x768. Ablation studies compare different data mixtures and attribute the generalization to native-resolution training. An appendix additionally reports text-to-image experiments on COCO.
Significance. If the empirical claims hold, this is a practically valuable contribution: a single diffusion-transformer model that is competitive or state-of-the-art on two standard ImageNet benchmarks, with credible zero-shot extrapolation to higher resolutions and varied aspect ratios, would address a real limitation of fixed-resolution generative models. The paper's strengths include a transparent token-budget accounting, comparisons against several modern baselines, explicit descriptions of the packing and attention mechanisms, and an honest discussion of limitations. However, the central mechanistic conclusion, that training on native resolutions is what enables the generalization, rests on an ablation that does not isolate the contribution of native-resolution data from other variables. In addition, several evaluation details that are load-bearing for the quantitative claims, such as how FID reference statistics are constructed at novel resolutions and how classifier-free-guidance hyperparameters were chosen, are not specified. These gaps prevent the paper, in its current form, from fully establishing the advertised mechanism, though the empirical system itself appears promising.
major comments (4)
- [Section 4.3, Tables 4 and 5] The ablation study varies the data mixture while keeping the NiT architecture fixed, but the comparison between arm (b), 'Native Resolution + 256 + 512', and arm (c), '256 + 512', changes two things at once: the presence of native-resolution images and the number and diversity of resolutions seen during training. Moreover, no ablation removes or replaces the two architectural innovations, 2D RoPE and packed attention without padding, which are known to be favorable for positional extrapolation and for avoiding padding artifacts. Consequently, the large improvement of (b) over (c) at 768x768 (FID 12.42 vs. 33.50) and the gains in aspect-ratio generalization cannot be attributed specifically to 'learning intrinsic visual distributions from native resolutions.' The conclusion in the 'Insights' paragraph overstates what the experiments establish. Please add ablations that vary the data mixture while holding the architecture fixed and that also vary the architecture (e.g., with and without 2D RoPE, with and without packing) to disentangle these factors.
- [Section 4.1 and Tables 2-3] The paper reports FID values at resolutions and aspect ratios that do not exist natively in ImageNet (e.g., 1024x1024, 1536x1536, 320x960), but it never describes how the reference feature statistics are constructed for these evaluation settings. If the reference set is created by resizing or cropping ImageNet images to each target resolution, the resulting FID values depend on that preprocessing choice and are not directly comparable across resolutions or to the baseline numbers, which may have been computed on different reference sets. This is load-bearing for the zero-shot FID claims. Please describe the reference set construction explicitly, state the number of samples, and clarify whether the same reference set is used for all methods at each resolution and aspect ratio.
- [Table 7 and Section 4.2] Table 7 reports that the CFG scale and CFG interval are tuned separately for every evaluation setting: for example, 256x256 uses scale 2.25 and interval [0.0, 0.7], 1536x1536 uses scale 1.5 and interval [0.0, 1.0], and 960x320 uses scale 4.5 and interval [0.0, 0.9]. The headline comparisons in Tables 1-3 therefore mix model capability with per-setting guidance optimization. It is not stated whether the baseline models were tuned to the same degree for each resolution and aspect ratio. Please either report results with a fixed CFG policy across all settings or provide an explicit upper envelope over a grid of CFG values, and state the baseline tuning procedure.
- [Section 4.1 and Table 1] No confidence intervals, multiple evaluation runs, or different random seeds are reported for any FID value, and the code and checkpoints are not released. Given that some of the headline improvements are numerically small (e.g., 256x256 FID 2.03 vs. 2.00 for FlowDCN, and 512x512 FID 1.45 vs. 1.88 for EDM2-L in the same parameter class), the reader cannot assess whether the reported ordering is stable. Please report variance over at least a few evaluation runs or seeds and, if possible, release code and trained checkpoints to make the empirical claims reproducible.
minor comments (6)
- [Table 1] The column header line 'mFID↓FID↓sFID↓IS↑Prec.↑Rec.↑' is malformed, with the metric names and arrows run together; this should be split into separate columns for clarity.
- [Table 2] There are numerical formatting inconsistencies, such as '9.817' and '41.170' for FlowDCN, which contain an extra digit compared with the other entries; please standardize the number of decimal places.
- [Section 4.2] The text states that NiT-XL achieves an FID of 4.07 on 768x768 resolution, but Table 2 reports 4.05 for this setting; please correct the text to match the table.
- [Table 3] The footnote explains that SiT-REPA was evaluated at different resolutions (160x480, 216x384, etc.) because only the 256-resolution model is open-sourced, while all other methods are evaluated at 320x960, 432x768, etc.; this makes the comparison at those aspect ratios unmatched. Please state explicitly whether this difference is expected to bias the comparison and, if possible, obtain matched-resolution numbers for SiT-REPA.
- [Abstract and Section 1] The term 'zero-shot' is used for resolutions and aspect ratios that may partially overlap with the native-resolution training distribution of ImageNet; please define precisely what counts as an unseen resolution or aspect ratio relative to the training data and clarify the status of settings such as 768x768 and 4:3.
- [Table 5] Table 5 compares only arms (a) and (b), but the table does not include a row for arm (c) and the caption does not explain the omission; in the text, the comparison is described in Section 4.3, yet the table header 'Data' is less informative than the full data-mixture labels used in Table 4. Please align the notation and caption with the body text.
Circularity Check
No significant circularity: NiT's benchmark results are externally measured, and the paper's self-citations appear only as baselines or routine engineering precedents, not as premises that force the conclusions.
full rationale
NiT is an empirical system paper; there is no formal derivation whose inputs are equivalent to its outputs. The headline SOTA claims are evaluated against fixed external benchmarks (FID/sFID/IS on ImageNet 256x256 and 512x512), so they cannot reduce by construction to the method's own design choices. The ablation in Section 4.3 compares data mixtures (native-only, native+256+512, 256+512) and concludes that 'training in native-resolution to learn a resolution- and aspect-ratio invariant visual distribution' is the enabler of generalization. This experiment is confounded: the architecture, including 2D RoPE and packed attention, is held fixed while data diversity changes, so the mechanistic attribution is under-supported. That is, however, an experimental-design or correctness concern, not circularity, because the inputs (data mixture) and the measured outputs (FID at 768x768 and at varied aspect ratios) are distinct quantities. Self-citations to the authors' prior FiT and FiTv2 work appear only as baselines in Tables 1-3 and as background on padding artifacts; neither is invoked as an unverified premise that forces NiT's design or outcomes. The core architectural ingredients are grounded in external, non-overlapping prior work (FlashAttention-2 [15], histogram packing [39], RoPE [68], and native-resolution VLM processing [4, 70, 72, 75]). The paper's own limitation statement, 'its generalization ability on extremely high-resolution and aspect ratios is still not satisfactory,' further indicates that the generalization claim is presented as a bounded empirical result rather than as a definitional consequence. Score 1 reflects only the presence of minor, non-load-bearing self-citations.
Assumptions & free parameters
free parameters (3)
- CFG scale and interval per evaluation setting =
Per-setting values in Table 7 (e.g., 256x256: 2.25; 2048x2048: 1.5; 960x320: 4.5)
- Logit-Normal time distribution parameters (P_mean, P_std) =
Not specified numerically; described as manually selected coefficients in Section 3.1
- 2D RoPE base theta =
10000
assumptions (4)
- standard math Flow matching with linear path and EDM logit-normal time distribution is a valid training objective for variable-resolution data.
- domain assumption FID computed at each evaluated resolution and aspect ratio is a valid and comparable measure of generation quality.
- domain assumption ImageNet native-resolution distribution is diverse enough to support extrapolation to unseen resolutions and aspect ratios.
- domain assumption The DC-AE autoencoder remains accurate at 32x downsampling for resolutions far outside its typical training range.
Cite this review
Pith. "Pith review of Native-Resolution Image Synthesis." pith.science (2026). https://pith.science/paper/NF4Z7NYG
@misc{pith2026250603131,
author = {Pith},
title = {Pith review of: Native-Resolution Image Synthesis},
year = {2026},
howpublished = {\url{https://pith.science/paper/NF4Z7NYG}},
note = {Machine review of arXiv:2506.03131}
}
read the original abstract
We introduce native-resolution image synthesis, a novel generative modeling paradigm that enables the synthesis of images at arbitrary resolutions and aspect ratios. This approach overcomes the limitations of conventional fixed-resolution, square-image methods by natively handling variable-length visual tokens, a core challenge for traditional techniques. To this end, we introduce the Native-resolution diffusion Transformer (NiT), an architecture designed to explicitly model varying resolutions and aspect ratios within its denoising process. Free from the constraints of fixed formats, NiT learns intrinsic visual distributions from images spanning a broad range of resolutions and aspect ratios. Notably, a single NiT model simultaneously achieves the state-of-the-art performance on both ImageNet-256x256 and 512x512 benchmarks. Surprisingly, akin to the robust zero-shot capabilities seen in advanced large language models, NiT, trained solely on ImageNet, demonstrates excellent zero-shot generalization performance. It successfully generates high-fidelity images at previously unseen high resolutions (e.g., 1536 x 1536) and diverse aspect ratios (e.g., 16:9, 3:1, 4:3), as shown in Figure 1. These findings indicate the significant potential of native-resolution modeling as a bridge between visual generative modeling and advanced LLM methodologies.
Figures
Figures from the paper (22 more)
Forward citations
Cited by 3 Pith papers
-
Transition Models: Rethinking the Generative Learning Objective
TiM trains a single diffusion-type model on arbitrary time-interval transitions, achieving strong one-step and multi-step text-to-image generation with 865M parameters.
-
PixNerd: Pixel Neural Field Diffusion
PixNerd is a single-stage pixel-space diffusion transformer that uses predicted neural field weights to decode large patches, reaching 2.15 FID on ImageNet 256 without a VAE.
-
Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing
A compact 4B image generation/editing system with a fast one-step VAE, native-resolution packing, RL alignment, and 4-step distillation reports competitive benchmarks against 6B–80B open models.
Reference graph
Works this paper leans on
-
[1]
Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023
Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023
arXiv 2023
-
[2]
Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions.arXiv preprint arXiv:2303.08797, 2023
arXiv 2023
-
[3]
Building normalizing flows with stochastic interpolants
Michael S Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. arXiv preprint arXiv:2209.15571, 2022
arXiv 2022
-
[4]
Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923, 2025
arXiv 2025
-
[5]
Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers.arXiv preprint arXiv:2211.01324, 2022
arXiv 2022
-
[6]
Language models are few-shot learners
Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020
2020
-
[7]
Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers.arXiv preprint arXiv:2301.00704, 2023
arXiv 2023
-
[8]
Maskgit: Masked generative image transformer
Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022
2022
Show all 87 references
-
[9]
Clex: Continuous length extrapolation for large language models, 2024.URL https://arxiv
Guanzheng Chen, Xin Li, Zaiqiao Meng, Shangsong Liang, and Lidong Bing. Clex: Continuous length extrapolation for large language models, 2024.URL https://arxiv. org/abs/2310.16450
2024 arXiv
-
[10]
Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation
Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. InEuropean Conference on Computer Vision, pages 74–91. Spr...
2024
-
[11]
Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis.arXiv preprint arXiv:2310.00426, 2023
Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis.arXiv preprint arXiv:2310.00426, 2023
-
[12]
Deep compression autoencoder for efficient high-resolution diffusion models.arXiv preprint arXiv:2410.10733, 2024
Junyu Chen, Han Cai, Junsong Chen, Enze Xie, Shang Yang, Haotian Tang, Muyang Li, Yao Lu, and Song Han. Deep compression autoencoder for efficient high-resolution diffusion models.arXiv preprint arXiv:2410.10733, 2024
2024 arXiv
-
[13]
Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling.arXiv preprint arXiv:2412.05271, 2024
Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling.arXiv preprint arXiv:2412.05271, 2024
2024 arXiv
-
[14]
Ramadge, and Alexander Rudnicky
Ta-Chung Chi, Ting-Han Fan, Peter J. Ramadge, and Alexander Rudnicky. Kerple: Kernelized relative positional embedding for length extrapolation. InNeurIPS, 2022
2022
-
[15]
Flashattention-2: Faster attention with better parallelism and work partitioning.arXiv preprint arXiv:2307.08691, 2023
Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning.arXiv preprint arXiv:2307.08691, 2023
2023 arXiv
-
[16]
Patch n’pack: Navit, a vision transformer for any aspect ratio and resolution.arXiv preprint arXiv:2307.06304, 2023
Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim Alabdulmohsin, et al. Patch n’pack: Navit, a vision transformer for any aspect ratio and resolution.arXiv preprint arXiv...
2023 arXiv
-
[17]
Imagenet: A large-scale hierarchical image database
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. InCVPR, 2009
2009
-
[18]
Diffusion models beat gans on image synthesis.NeurIPS, 2021
Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis.NeurIPS, 2021
2021
-
[19]
Cogview2: Faster and better text-to-image generation via hierarchical transformers.Advances in Neural Information Processing Systems, 35:16890–16902, 2022
Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers.Advances in Neural Information Processing Systems, 35:16890–16902, 2022. 26
2022
-
[20]
Longrope: Extending llm context window beyond 2 million tokens.arXiv preprint arXiv:2402.13753, 2024
Yiran Ding, Li Lyna Zhang, Chengruidong Zhang, Yuanyuan Xu, Ning Shang, Jiahang Xu, Fan Yang, and Mao Yang. Longrope: Extending llm context window beyond 2 million tokens.arXiv preprint arXiv:2402.13753, 2024
2024 arXiv
-
[21]
An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020
Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:...
2010 arXiv
-
[22]
The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024
Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024
2024 arXiv
-
[23]
Scaling rectified flow transformers for high-resolution image synthesis
Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. 2024
2024
-
[24]
Make-a-scene: Scene-based text-to-image generation with human priors
Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. InEuropean Conference on Computer Vision, pages 89–106. Springer, 2022
2022
-
[25]
Masked diffusion transformer is a strong image synthesizer.arXiv preprint arXiv:2303.14389, 2023
Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Masked diffusion transformer is a strong image synthesizer.arXiv preprint arXiv:2303.14389, 2023
2023 arXiv
-
[26]
Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025
Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025
2025 arXiv
-
[27]
Lm-infinite: Zero-shot extreme length generalization for large language models.arXiv preprint arXiv:2308.16137, 2023
Chi Han, Qifan Wang, Hao Peng, Wenhan Xiong, Yu Chen, Heng Ji, and Sinong Wang. Lm-infinite: Zero-shot extreme length generalization for large language models.arXiv preprint arXiv:2308.16137, 2023
2023 arXiv
-
[28]
Greg Heinrich, Mike Ranzinger, Yin Hongxu, Yao Lu, Jan Kautz, Andrew Tao, Bryan Catanzaro, and Pavlo Molchanov. Radiov2. 5: Improved baselines for agglomerative vision foundation models. InProc. CVPR, volume 2, page 6, 2025
2025
-
[29]
Gans trained by a two time-scale update rule converge to a local nash equilibrium.NeurIPS, 2017
Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium.NeurIPS, 2017
2017
-
[30]
Denoising diffusion probabilistic models.NeurIPS, 2020
Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.NeurIPS, 2020
2020
-
[31]
Lora: Low-rank adaptation of large language models
Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. InICLR, 2022
2022
-
[32]
Estimation of non-normalized statistical models by score matching
Aapo Hyvärinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. 2005
2005
-
[33]
Springer Science & Business Media, 2005
Bernd Jähne.Digital image processing. Springer Science & Business Media, 2005
2005
-
[34]
Scaling laws for neural language models.arXiv preprint arXiv:2001.08361, 2020
Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.arXiv preprint arXiv:2001.08361, 2020
2001 arXiv
-
[35]
Elucidating the design space of diffusion-based generative models.NeurIPS, 2022
Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models.NeurIPS, 2022
2022
-
[36]
Analyzing and improving the training dynamics of diffusion models
Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24174–24184, 2024
2024
-
[37]
The impact of positional encoding on length generalization in transformers.Advances in Neural Information Processing Systems, 36:24892–24928, 2023
Amirhossein Kazemnejad, Inkit Padhi, Karthikeyan Natesan Ramamurthy, Payel Das, and Siva Reddy. The impact of positional encoding on length generalization in transformers.Advances in Neural Information Processing Systems, 36:24892–24928, 2023
2023
-
[38]
Segment anything
Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. InProceedings of the IEEE/CVF international conference on computer vision, pages 4015–4026, 2023
2023
-
[39]
Efficient sequence packing without cross-contamination: Accelerating large language models without impacting performance.arXiv preprint arXiv:2107.02027, 2021
Mario Michael Krell, Matej Kosec, Sergio P Perez, and Andrew Fitzgibbon. Efficient sequence packing without cross-contamination: Accelerating large language models without impacting performance.arXiv preprint arXiv:2107.02027, 2021. 27
2021 arXiv
-
[40]
Kynkäänniemi, T
T. Kynkäänniemi, T. Karras, S. Laine, and T Lehtinen, J.and Aila. Improved precision and recall metric for assessing generative models.NeurIPS, 2019
2019
-
[41]
Microsoft coco: Common objects in context
Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InECCV, 2014
2014
-
[42]
Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022
Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022
2022 arXiv
-
[43]
Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024
Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024
2024 arXiv
-
[44]
Flow straight and fast: Learning to generate and transfer data with rectified flow
Xingchao Liu, Chengyue Gong, and qiang liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. InICLR, 2023
2023
-
[45]
Ntk-aware scaled rope allows llama models to have extended (8k+) context size with- out any fine-tuning and minimal perplexity degradation
LocalLLaMA. Ntk-aware scaled rope allows llama models to have extended (8k+) context size with- out any fine-tuning and minimal perplexity degradation. https://www.reddit.com/r/LocalLLaMA/ comments/14lz7j5/ntkaware_scaled_rope_allows_llama_models_to_have/. Accessed: 2024- 2-1
2024
-
[46]
Fit: Flexible vision transformer for diffusion model
Zeyu Lu, ZiDong Wang, Di Huang, Chengyue Wu, Xihui Liu, Wanli Ouyang, and LEI BAI. Fit: Flexible vision transformer for diffusion model. 2024
2024
-
[47]
Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers.arXiv preprint arXiv:2401.08740, 2024
Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers.arXiv preprint arXiv:2401.08740, 2024
2024 arXiv
-
[48]
Sdedit: Guided image synthesis and editing with stochastic differential equations.arXiv preprint arXiv:2108.01073, 2021
Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations.arXiv preprint arXiv:2108.01073, 2021
2021 arXiv
-
[49]
The llama 4 herd: The beginning of a new era of natively multimodal ai innovation, april 2025, 2025
AI Meta. The llama 4 herd: The beginning of a new era of natively multimodal ai innovation, april 2025, 2025
2025
-
[50]
C. Nash, J. Menick, S. Dieleman, and P. W Battaglia. Generating images with sparse representations. arXiv preprint arXiv:2103.03841, 2021
2021 arXiv
-
[51]
Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741, 2021
Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741, 2021
2021 arXiv
-
[52]
Improved denoising diffusion probabilistic models
Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162–8171. PMLR, 2021
2021
-
[53]
Pearson Educación, 1997
Alan V Oppenheim, Alan S Willsky, and Syed Hamid Nawab.Signals & systems. Pearson Educación, 1997
1997
-
[54]
Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023
Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023
2023 arXiv
-
[55]
Scalable diffusion models with transformers
William Peebles and Saining Xie. Scalable diffusion models with transformers. InICCV, 2023
2023
-
[56]
Yarn: Efficient context window extension of large language models.arXiv preprint arXiv:2309.00071, 2023
Bowen Peng, Jeffrey Quesnelle, Honglu Fan, and Enrico Shippole. Yarn: Efficient context window extension of large language models.arXiv preprint arXiv:2309.00071, 2023
2023 arXiv
-
[57]
Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023
Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023
2023 arXiv
-
[58]
Train short, test long: Attention with linear biases enables input length extrapolation
Ofir Press, Smith Noah, and Lewis Mike. Train short, test long: Attention with linear biases enables input length extrapolation. InICLR, 2021
2021
-
[59]
Hierarchical text-conditional image generation with clip latents.arXiv preprint arXiv:2204.06125, 1(2):3, 2022
Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents.arXiv preprint arXiv:2204.06125, 1(2):3, 2022
2022 arXiv
-
[60]
Zero-shot text-to-image generation
Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. InInternational conference on machine learning, pages 8821–8831. Pmlr, 2021. 28
2021
-
[61]
High-resolution image synthesis with latent diffusion models
Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. InCVPR, 2022
2022
-
[62]
Salimans, I
T. Salimans, I. Goodfellow, W. Zaremba, V . Cheung, A. Radford, and X Chen. Improved techniques for training gans.NeurIPS, 2016
2016
-
[63]
Adversarial diffusion distillation
Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. InEuropean Conference on Computer Vision, pages 87–103. Springer, 2024
2024
-
[64]
Generative modeling by estimating gradients of the data distribution
Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, 2019
2019
-
[65]
Improved techniques for training score-based generative models.NeurIPS, 2020
Yang Song and Stefano Ermon. Improved techniques for training score-based generative models.NeurIPS, 2020
2020
-
[66]
Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020
Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020
2011 arXiv
-
[67]
Springer, 2013
Milan Sonka, Vaclav Hlavac, and Roger Boyle.Image processing, analysis and machine vision. Springer, 2013
2013
-
[68]
Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 2024
Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 2024
2024
-
[69]
Autore- gressive model beats diffusion: Llama for scalable image generation.arXiv preprint arXiv:2406.06525, 2024
Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autore- gressive model beats diffusion: Llama for scalable image generation.arXiv preprint arXiv:2406.06525, 2024
2024 arXiv
-
[70]
Seed1.5-vl technical report.arXiv preprint arXiv:2505.07062, 2025
ByteDance Seed Team. Seed1.5-vl technical report.arXiv preprint arXiv:2505.07062, 2025
2025 arXiv
-
[71]
Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 2023
Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 2023
2023 arXiv
-
[72]
Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.arXiv preprint arXiv:2403.05530, 2024
Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.arXiv preprint arXiv:2403.05530, 2024
2024 arXiv
-
[73]
Gemma 3 technical report.arXiv preprint arXiv:2503.19786, 2025
Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, et al. Gemma 3 technical report.arXiv preprint arXiv:2503.19786, 2025
2025 arXiv
-
[74]
Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, et al. Kimi k1. 5: Scaling reinforcement learning with llms.arXiv preprint arXiv:2501.12599, 2025
2025 arXiv
-
[75]
Kimi-vl technical report.arXiv preprint arXiv:2504.07491, 2025
Kimi Team, Angang Du, Bohong Yin, Bowei Xing, Bowen Qu, Bowen Wang, Cheng Chen, Chenlin Zhang, Chenzhuang Du, Chu Wei, et al. Kimi-vl technical report.arXiv preprint arXiv:2504.07491, 2025
2025 arXiv
-
[76]
Visual autoregressive modeling: Scalable image generation via next-scale prediction.Advances in neural information processing systems, 37:84839–84865, 2024
Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction.Advances in neural information processing systems, 37:84839–84865, 2024
2024
-
[77]
Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023a
Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, and Baptiste Rozière et al. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023a
-
[78]
Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023b
Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, and Nikolay Bashlykov et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023b
-
[79]
Flowdcn: Exploring dcn-like architectures for fast image generation with arbitrary resolution.arXiv preprint arXiv:2410.22655, 2024
Shuai Wang, Zexian Li, Tianhui Song, Xubin Li, Tiezheng Ge, Bo Zheng, and Limin Wang. Flowdcn: Exploring dcn-like architectures for fast image generation with arbitrary resolution.arXiv preprint arXiv:2410.22655, 2024
2024 arXiv
-
[80]
Fitv2: Scalable and improved flexible vision transformer for diffusion model.arXiv preprint arXiv:2410.13925, 2024
ZiDong Wang, Zeyu Lu, Di Huang, Cai Zhou, Wanli Ouyang, and LEI BAI. Fitv2: Scalable and improved flexible vision transformer for diffusion model.arXiv preprint arXiv:2410.13925, 2024. 29
2024 arXiv
-
[81]
Sana: Efficient high-resolution image synthesis with linear diffusion transformers.arXiv preprint arXiv:2410.10629, 2024
Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution image synthesis with linear diffusion transformers.arXiv preprint arXiv:2410.10629, 2024
-
[82]
An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report.arXiv preprint arXiv:2412.15115, 2024
2024 arXiv
-
[83]
Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024
Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024
2024 arXiv
-
[84]
Minicpm-v: A gpt-4v level mllm on your phone.arXiv preprint arXiv:2408.01800, 2024
Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone.arXiv preprint arXiv:2408.01800, 2024
2024 arXiv
-
[85]
Scaling autoregressive models for content-rich text-to-image generation.arXiv preprint arXiv:2206.10789, 2(3):5, 2022
Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation.arXiv preprint arXiv:2206.10789, 2(3):5, 2022
2022 arXiv
-
[86]
Representation alignment for generation: Training diffusion transformers is easier than you think
Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024
-
[87]
Pose: Efficient context window extension of llms via positional skip-wise training.arXiv preprint arXiv:2309.10400, 2023
Dawei Zhu, Nan Yang, Liang Wang, Yifan Song, Wenhao Wu, Furu Wei, and Sujian Li. Pose: Efficient context window extension of llms via positional skip-wise training.arXiv preprint arXiv:2309.10400, 2023. 30
2023 arXiv
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.