Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Scale-DiT generates coherent 4K images from a 1K-pretrained diffusion model with only local windows and a low-res anchor.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 09:14 UTC pith:JVYF24XJ

load-bearing objection A practical efficiency recipe for FLUX-to-4K that deserves a serious referee, but the unspecified low-res guidance pathway and the overclaiming abstract must be fixed first. the 4 major comments →

arxiv 2510.16325 v4 pith:JVYF24XJ submitted 2025-10-18 cs.CV

UltraImageGen: Efficient Ultra-High-Resolution Image Generation with Hierarchical Local Attention

Yuyao Zhang , Yu-Wing Tai This is my paper
classification cs.CV
keywords text-to-image diffusionultra-high-resolution generationlocal window attentionlow-resolution global guidanceLoRA adaptationflow matchingnear-linear attention scalingdiffusion transformer
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Diffusion text-to-image models are stuck near 1K resolution because full attention cost grows quadratically with image size, and genuinely high-resolution training data is scarce. This paper claims both problems can be sidestepped: instead of retraining on 4K data, keep the pretrained model and change the attention pattern. High-resolution latents are split into fixed-size windows that attend only within the window, so cost grows nearly linearly with pixel count. A small low-resolution latent with scaled positional encodings is added as a global anchor to preserve structure across windows, and only the projections that process this low-resolution branch are adapted with LoRA, trained on 10,000 synthetic 1K images. If the claim holds, a standard pretrained model can generate 4K images at over twice the speed and lower memory than dense attention, matching methods trained on native 4K data.

Core claim

The central claim is that a diffusion transformer pretrained at 1K can reliably generate 4K×4K images without any native 4K training data, by decomposing attention into local windows plus a low-resolution global anchor. The original Q, K, V projections are reused unchanged for high-resolution window tokens; only the projections applied to the low-resolution latent are LoRA-adapted, with training done entirely on 1K-resolution synthetic images (10,000 examples, 20,000 steps). The paper reports that at 4K the method reaches FID 67.03, Inception Score 17.21, and CLIP score 0.3231, matching or outperforming baselines including methods trained on native 4K data, while running in 395 seconds versu

What carries the argument

The mechanism is hierarchical local attention with low-resolution global guidance. High-resolution latent X is partitioned into non-overlapping windows of 16×16 tokens (256×256 pixels), and each window computes self-attention only within itself plus a small neighboring band, reducing complexity from O(N^2) to O(N·l^2). A low-resolution latent X_lr at 1/4 scale has its positions scaled by ratio ρ=4, projecting it onto the high-resolution grid; each high-res window attends to its local tokens and to the corresponding scaled low-res region, while low-res tokens attend globally among themselves and to text. Because scaling positions changes the frequencies in the rotary position embeddings (RoPE

Load-bearing premise

The load-bearing premise is that the pretrained Q, K, V projections and rotation-frequency statistics learned at 1K transfer unchanged to arbitrary 4K local windows, and that the low-resolution guidance latents can be reliably produced and aligned during joint denoising even though the paper does not specify how that is done.

What would settle it

A direct falsifier is to reproduce the method on a different 1K-pretrained diffusion transformer with the same 10,000-image LoRA budget: if it produces structural artifacts or requires far more adaptation data, the claim that no high-resolution training data are needed is false. A cheaper internal check: log X_lr during the 395-second generation; if the low-resolution latent is not a proper noised flow-matching latent (e.g., it is clean or must be denoised separately), the reported latency omits a required component and the efficiency claim weakens.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Scaling to 4K no longer requires collecting or synthesizing native 4K training data; commodity 1K data suffices for the LoRA adaptation.
  • Inference efficiency grows with resolution: doubling resolution roughly doubles cost rather than quadrupling it, and the speed/memory gap over dense attention widens from 1K to 4K.
  • The design is recursive—a generated high-resolution image can be downsampled and used as low-resolution guidance for an even higher resolution, so the same model can reach 8K and beyond with stable cost.
  • Output quality is claimed to be resolution-invariant because every local window operates within the pretrained positional distribution; GenEval numbers stay essentially unchanged between 1K and 4K.
  • The approach can be plugged into an existing pretrained multimodal diffusion transformer rather than requiring training from scratch.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the pattern transfers, the same local-window plus low-res-anchor recipe should work on other 1K-pretrained diffusion transformers with minimal re-tuning, since it does not modify the pretrained generation path.
  • A cheap testable extension: sweep the low-res anchor resolution instead of fixing a 4:1 ratio; the theory predicts global coherence is governed by the anchor's resolution, so raising it should fix remaining long-range failures without enlarging windows.
  • The paper does not specify how X_lr is produced or noised during joint denoising; the reported 395-second latency may omit the cost of creating the low-resolution latent, so readers reproducing the number should check whether an extra forward pass is required.
  • Because the claimed speedup is against dense attention on the same backbone, a fair comparison should also measure a dense-attention baseline that itself uses Hilbert permutation and block-sparse kernels; part of the gain may come from the implementation rather than the attention pattern.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The manuscript (listed on arXiv under the title "UltraImageGen", internally titled "Scale-DiT") proposes a method to extend a pretrained text-to-image diffusion transformer, FLUX.1-dev, to 4K×4K resolution without native high-resolution training data. The high-resolution latent is partitioned into fixed-size local windows so that self-attention scales near-linearly, while a low-resolution latent X_lr with scaled RoPE anchors provides global context. Only the Q/K/V projections for X_lr are adapted via LoRA, trained on 10,000 generated 1024×1024 images. The paper reports FID 67.03, IS 17.21, CLIP 0.3231 at 4K, about 2× lower latency and lower memory than dense-attention FLUX, and consistent GenEval scores from 1K to 4K. A theoretical appendix argues that local-window quality is resolution-invariant. The central claim is that this architecture "reliably scales to 4K×4K without requiring additional high-resolution training data."

Significance. The core idea—reusing the pretrained local attention path unchanged while adding a low-resolution global anchor—is attractive and, if fully specified, could offer a practical way to push T2I models beyond their training resolutions with modest compute. The efficiency analysis (near-linear scaling with fixed window size), the separation of frozen high-resolution projections from LoRA-adapted low-resolution projections, and the broad comparison against 4K baselines are useful contributions. The memory/latency comparison on common hardware is also a strength. However, the central conditioning signal X_lr is not specified algorithmically, the quantitative evidence lacks uncertainty quantification, and the appendix's "resolution-invariant" guarantee is an assumption rather than a proof. These issues are load-bearing for the paper's main claim and need to be addressed before the result can be relied upon.

major comments (4)
  1. [Sections 3 and 4.1, loss equation in "Parameter-Efficient Joint Denoising"] The low-resolution guidance latent X_lr is never specified as an algorithm. The text says "we generate a low-resolution guidance image X_lr" and the flow-matching loss conditions on X_lr, but it does not state how X_lr is obtained at inference from pure noise, whether it is noised or downsampled per step, or whether it is produced by a separate FLUX pass. If X_lr is a separately generated 256×256 image, the end-to-end latency and memory should include that pass; if it is a downsampled version of the current X, it cannot provide global guidance at high noise levels; if it is derived from ground truth, it violates the no-high-resolution-data claim. The same issue applies to training: for a 1024×1024 training image, is X_lr its low-resolution sibling or a synthetic low-res output? Without this, Table 3 and the FID/IS results are not reproducible.
  2. [Table 4; Tables 1, 2, 5] The claim that window-size and ratio differences are "not statistically significant" is unsupported: no test, confidence interval, seed count, or variance estimate is reported. In Table 4, the FID spread is 67.83 vs. 66.31, which may be practically meaningful. More generally, Tables 1, 2, and 5 report single-run metrics with no error bars; in Table 1 the claimed FID lead over URAE is only 0.36, so without repeated runs or intervals it is impossible to assess whether the ranking is stable. This is particularly important because the paper's central claim is that quality is maintained at 4K.
  3. [Appendix H.2, Eq. (3)] The "resolution-invariant quality" guarantee is asserted, not proved. The argument assumes that every local window's relative position range falls entirely inside the pretrained distribution D_train, but the attention actually mixes local tokens with X_lr tokens at scaled/out-of-distribution RoPE positions and with text tokens. Thus the effective attention context at 4K is not the same as at 1K. As a concrete test, the authors should measure per-window distribution shift or ablate scaled positional anchors; until then, Eq. (3) is a plausible assumption rather than a theoretical guarantee.
  4. [Abstract and Table 3] The arXiv abstract states that the method scales "to resolutions higher than 8K with more than 10× speed up," while the manuscript's own abstract, introduction, and conclusion claim 4K generation with "more than 2×" speedup, and Table 3 reports about 2× speedup over FLUX. This is a material overclaim/mismatch and must be reconciled. If 8K and 10× results exist, they should be reported in the body; otherwise the abstract should be corrected.
minor comments (6)
  1. [Title/arXiv metadata] The arXiv listing is titled "UltraImageGen" but the manuscript is titled "Scale-DiT" throughout. Please unify the name and ensure the abstract matches the body claims.
  2. [Appendix H.1, Eq. (1)] The complexity formula appears to have an error: the number of windows should involve division by w^2, giving (HW/(s^2 w^2)) * (w^2)^2 = HW w^2 / s^2, not HW/s^2 * w^2 * w^4. The paper's overall linear-scaling conclusion is plausible, but the equation as written is not dimensionally correct.
  3. [Section 4.1, metrics] "FID patch" and "IS patch" are not defined. Please specify patch size, stride, and whether patches are extracted from generated images only or also from the reference set.
  4. [Table 3] The caption says "time to generate one 4K latents (65536 tokens)" which is ambiguous: is this total end-to-end generation time, or one denoising iteration? Please clarify and state whether the low-resolution guidance pass is included.
  5. [Section 4.1] The kernel sizes are stated as "Q-block=128 and K-block=64 (128 for blocksparse Flash Attention)", which is inconsistent. Please clarify the actual block sizes used in the reported experiments.
  6. [Related work and references] The reference "Zhang et al." appears without a year or venue in the related-work discussion and in the linear-attention comparison. Please complete the citation.

Circularity Check

1 steps flagged

Core benchmarks are external and not fitted; the only circular passage is the appendix's 'resolution-invariant quality' guarantee, which restates its assumption.

specific steps
  1. self definitional [Appendix H.2 (Quality Preservation across Resolutions)]
    "Since every local window operates on the same relative position range R_local = {(∆i,∆j) : |∆i|,|∆j| ≤ w−1}, and this range falls entirely within the pretrained distribution D_train, each window achieves identical quality. ... Quality per unit area = Q(R_local)/(w^2/s^2) = constant ... The quality is resolution-invariant because the fundamental building blocks (local spatial relationships) remain within the learned parameter space."

    The appendix presents this as a 'theoretical guarantee' that per-unit quality is constant across resolutions. But the asserted equality rests entirely on the premise that 'each window achieves identical quality'; that premise is the same proposition as resolution-invariant quality. The formula merely writes Q(R_local) as a constant rather than deriving it from the architecture or experiments. If local-window quality at 4K were actually worse than at 1K, the formula would still hold only because Q(R_local) is defined as the constant. It is a tautology, not an independent derivation.

full rationale

The main quantitative claims (FID 67.03, IS 17.21, CLIP 0.3231 at 4K; GenEval at 2K/3K/4K; memory and latency in Table 3) are evaluated against external reference sets (LAION-HR, GenEval) and against other 4K methods, and the LoRA parameters are trained only on 1024x1024 synthetic images from FLUX.1-dev. Those numbers are not fitted constants re-announced as predictions, and the method is not justified by a self-citation chain. The largest reproducibility gap — how X_lr is produced, noised, and aligned at inference, which Sections 3 and 4.1 do not specify — is a completeness/correctness problem rather than circularity, because the paper does not define X_lr in terms of the reported outcome. The only step that is circular is Appendix H.2, where 'resolution-invariant quality' is asserted from the assumption that every local window has identical quality; this is a restatement rather than a derivation. That tautology is not load-bearing for the headline empirical results, hence the low score.

Axiom & Free-Parameter Ledger

3 free parameters · 7 axioms · 2 invented entities

The central empirical claim rests on standard diffusion/attention machinery plus several domain assumptions: local windows preserve detail, the unadapted high-resolution branch transfers to 4K, a self-generated 10,000-image training set is sufficient, and the benchmark protocol is fair. The paper's 'theoretical guarantee' of resolution-invariant quality in Appendix H.2 is an assumption in disguise, and the X_lr pathway is underspecified.

free parameters (3)
  • window_size_l = 16 tokens (256×256 pixel windows)
    Chosen by ablation as the smallest window that balances quality and GPU utilization; no theoretical derivation.
  • resolution_ratio_rho = 4
    Empirically set as the best balance between global-guidance quality and efficiency; no independent justification.
  • LoRA_rank = 16
    Selected without a systematic sweep; rank is a critical capacity hyperparameter for the guidance-branch adaptation.
axioms (7)
  • standard math Softmax attention and RoPE positional encoding operate exactly as in the pretrained FLUX/MMDiT architecture.
    The method reuses pretrained Q,K,V for high-res tokens and relies on the standard attention formulation in Section 3.
  • domain assumption Local windows of 16×16 tokens capture all fine texture detail, while global structure is adequately carried by the low-resolution guidance latent.
    This is the core hypothesis of the design; it is tested only by ablations, not proven.
  • domain assumption Pretrained Q,K,V projections for the high-resolution branch transfer to 4K without any adaptation, despite RoPE extrapolation via NTK scaling.
    Section 3 keeps original Q,K,V for X; the paper provides no direct evidence that unadapted high-res attention remains valid at 4K.
  • domain assumption A 10,000-image synthetic dataset generated by FLUX.1-dev at 1024×1024 is sufficient and unbiased for training the LoRA guidance pathway.
    Section 4.1 describes the dataset; there is no analysis of distribution bias or coverage of high-resolution compositional scenarios.
  • domain assumption LAION-High-Resolution is an appropriate FID reference set for 4K text-to-image generations.
    Section 4.1 computes FID against 10,000 LAION-HR images without demonstrating that the prompt distribution and image domain align with the 1,000 GPT-4o prompts.
  • ad hoc to paper Each local window operates within the pretrained distribution Dtrain, therefore per-unit-area quality is resolution-invariant.
    Appendix H.2 states this as a 'theoretical guarantee' but it is an assumption in disguise: it assumes the very thing it claims to prove.
  • domain assumption The low-resolution guidance latent X_lr is a faithful condition that can be reliably generated and aligned with the high-resolution latent throughout denoising.
    The paper never specifies how X_lr is produced (clean image, noisy latent, or recursive output), nor how it is noised in the flow-matching objective.
invented entities (2)
  • Low-resolution guidance latent X_lr no independent evidence
    purpose: Provides global semantic context to every local window through scaled positional anchors.
    Its utility is shown only by internal ablations; no external benchmark or independent evaluation validates the necessity of this new latent pathway.
  • Scaled RoPE positional anchors for X_lr no independent evidence
    purpose: Maps the low-resolution grid onto the high-resolution coordinate system so attention can align local and global structure.
    The scaling ratio rho=4 is a hand-set design choice; the paper provides no independent measurement or theoretical justification for this value.

pith-pipeline@v1.3.0-alltime-deepseek · 14102 in / 12698 out tokens · 108763 ms · 2026-08-04T09:14:38.513173+00:00 · methodology

0 comments
read the original abstract

Ultra-high-resolution text-to-image generation is increasingly vital for applications requiring fine-grained textures and global structural fidelity, yet state-of-the-art text-to-image diffusion models such as FLUX and SD3 remain confined to sub 2MP (< $1K\times2K$) resolutions due to the quadratic complexity of attention mechanisms and the scarcity of high-quality high-resolution training data. We present UltraImageGen, a novel framework that introduces hierarchical local attention with low-resolution global guidance, enabling efficient, scalable, and semantically coherent image synthesis at ultra-high resolutions. Specifically, high-resolution latents are divided into hardware aligned fixed-size local windows to reduce attention complexity from quadratic to near-linear, while a low-resolution latent equipped with scaled positional embeddings injects global semantics as an anchor. A lightweight LoRA adaptation bridges global and local pathways during denoising, ensuring consistency across structure and detail. To maximize inference efficiency and achieve scalable ultra-high-resolution generation, we repermute token sequence in window-first order, so that the GPU-friendly dense local blocks in attention calculation equals to the fixed-size local window in 2D regardless of resolution. Together ourwork reliably scales the pretrained model to resolutions higher than $8K$ with more than $10\times$ speed up and significantly lower memory usage. Extensive experiments demonstrate that ourwork achieves superior quality while maintaining computational efficiency, establishing a practical paradigm for advancing ultra-high-resolution image generation.

Figures

Figures reproduced from arXiv: 2510.16325 by Yu-Wing Tai, Yuyao Zhang.

Figure 1
Figure 1. Figure 1: Ultra-high-resolution images generated by [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Schematic of Scale-DiT’s attention block modifications. The left column illustrates that high-resolution image latents X are partitioned into local windows (red) in Hilbert-Curve order that attend to each other within their window. Simultaneously, a low-resolution guidance latent Xlr (yellow) provides global context to each window via position scaling. The right column shows the joint-denoising process and… view at source ↗
Figure 3
Figure 3. Figure 3: 4K comparison with leading baselines. Zoom in to observe the fine details. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Memory and speed statistics when scaling from 1K to 4K comparing with native FLUX.1 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Ablation Study on attention scale. Images from left to right corresponds to cases that 1) [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: More 4K results 15 [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: More 4K results 16 [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: More 4K results 17 [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: More 2K results. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Complete comparison against the 9 baselines on 4K resolution. Zoom in to view the [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Complete comparison against the 9 baselines on 4K resolution. Zoom in to view the [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Complete comparison against the 9 baselines on 2K resolution. Zoom in to view the [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. HierEdit: Region-Aware Hierarchical Diffusion for Efficient High-Resolution Editing

    cs.CV 2026-05 unverdicted novelty 6.0

    HierEdit enables efficient 4K image editing via low-resolution proxy localization followed by hierarchical local-window diffusion that reuses unaltered regions as conditioning.

Reference graph

Works this paper leans on

22 extracted references · 21 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Longformer: The long-document transformer

    Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150,

    Pith/arXiv arXiv 2004

  2. [4]

    Hidream-i1: A high-efficient image generative foundation model with sparse diffusion transformer.arXiv preprint arXiv:2505.22705,

    Qi Cai, Jingwen Chen, Yang Chen, Yehao Li, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Yiheng Zhang, Fengbin Gao, Peihan Xu, et al. Hidream-i1: A high-efficient image generative foundation model with sparse diffusion transformer.arXiv preprint arXiv:2505.22705,

    Pith/arXiv arXiv

  3. [5]

    Pixart-α: Fast training of diffusion transformer for photore- alistic text-to-image synthesis.arXiv preprint arXiv:2310.00426,

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photore- alistic text-to-image synthesis.arXiv preprint arXiv:2310.00426,

    Pith/arXiv arXiv

  4. [8]

    Flexprefill: A context-aware sparse attention mechanism for efficient long-sequence inference.arXiv preprint arXiv:2502.20766,

    Xunhao Lai, Jianqiao Lu, Yao Luo, Yiyuan Ma, and Xun Zhou. Flexprefill: A context-aware sparse attention mechanism for efficient long-sequence inference.arXiv preprint arXiv:2502.20766,

    Pith/arXiv arXiv

  5. [9]

    Playground v2

    Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2. 5: Three insights towards enhancing aesthetic quality in text-to-image generation.arXiv preprint arXiv:2402.17245, 2024a. Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et a...

    Pith/arXiv arXiv

  6. [11]

    Polaformer: Polarity-aware linear attention for vision transformers.arXiv preprint arXiv:2501.15061,

    Weikang Meng, Yadan Luo, Xin Li, Dongmei Jiang, and Zheng Zhang. Polaformer: Polarity-aware linear attention for vision transformers.arXiv preprint arXiv:2501.15061,

    Pith/arXiv arXiv

  7. [12]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952,

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952,

    Pith/arXiv arXiv

  8. [13]

    Freescale: Unleashing the resolution of diffusion models via tuning-free scale fusion.arXiv preprint arXiv:2412.09626,

    Haonan Qiu, Shiwei Zhang, Yujie Wei, Ruihang Chu, Hangjie Yuan, Xiang Wang, Yingya Zhang, and Ziwei Liu. Freescale: Unleashing the resolution of diffusion models via tuning-free scale fusion.arXiv preprint arXiv:2412.09626,

    Pith/arXiv arXiv

  9. [15]

    Sparse videogen: Accelerating video diffusion transformers with spatial-temporal sparsity.arXiv preprint arXiv:2502.01776,

    Haocheng Xi, Shuo Yang, Yilong Zhao, Chenfeng Xu, Muyang Li, Xiuyu Li, Yujun Lin, Han Cai, Jintao Zhang, Dacheng Li, et al. Sparse videogen: Accelerating video diffusion transformers with spatial-temporal sparsity.arXiv preprint arXiv:2502.01776,

    Pith/arXiv arXiv

  10. [16]

    Enze Xie, Junsong Chen, Yuyang Zhao, Jincheng Yu, Ligeng Zhu, Chengyue Wu, Yujun Lin, Zhekai Zhang, Muyang Li, Junyu Chen, et al

    URLhttps://arxiv.org/abs/2410.10629. Enze Xie, Junsong Chen, Yuyang Zhao, Jincheng Yu, Ligeng Zhu, Chengyue Wu, Yujun Lin, Zhekai Zhang, Muyang Li, Junyu Chen, et al. Sana 1.5: Efficient scaling of training-time and inference- time compute in linear diffusion transformer.arXiv preprint arXiv:2501.18427,

    Pith/arXiv arXiv

  11. [17]

    Xattention: Block sparse attention with antidiagonal scoring.arXiv preprint arXiv:2503.16428,

    Ruyi Xu, Guangxuan Xiao, Haofeng Huang, Junxian Guo, and Song Han. Xattention: Block sparse attention with antidiagonal scoring.arXiv preprint arXiv:2503.16428,

    Pith/arXiv arXiv

  12. [18]

    Sparse videogen2: Accelerate video generation with sparse attention via semantic-aware permutation.arXiv preprint arXiv:2505.18875,

    Shuo Yang, Haocheng Xi, Yilong Zhao, Muyang Li, Jintao Zhang, Han Cai, Yujun Lin, Xiuyu Li, Chenfeng Xu, Kelly Peng, et al. Sparse videogen2: Accelerate video generation with sparse attention via semantic-aware permutation.arXiv preprint arXiv:2505.18875,

    Pith/arXiv arXiv

  13. [19]

    Under review

    12 Preprint. Under review. Ruonan Yu, Songhua Liu, Zhenxiong Tan, and Xinchao Wang. Ultra-resolution adaptation with ease.arXiv preprint arXiv:2503.16322,

    Pith/arXiv arXiv

  14. [20]

    Ditfastattnv2: Head-wise attention compression for multi-modality diffusion transformers.arXiv preprint arXiv:2503.22796, 2025a

    Hanling Zhang, Rundong Su, Zhihang Yuan, Pengtao Chen, Mingzhu Shen Yibo Fan, Shengen Yan, Guohao Dai, and Yu Wang. Ditfastattnv2: Head-wise attention compression for multi-modality diffusion transformers.arXiv preprint arXiv:2503.22796, 2025a. Jinjin Zhang, Qiuyu Huang, Junjie Liu, Xiefan Guo, and Di Huang. Diffusion-4k: Ultra-high- resolution image synt...

    Pith/arXiv arXiv

  15. [21]

    The hedgehog & the porcupine: Expressive linear attentions with softmax mimicry, 2024b.URL https://arxiv

    Michael Zhang, Kush Bhatia, Hermann Kumbong, and Christopher R ´e. The hedgehog & the porcupine: Expressive linear attentions with softmax mimicry, 2024b.URL https://arxiv. org/abs/2402.04347. Qingping Zheng, Yuanfan Guo, Jiankang Deng, Jianhua Han, Ying Li, Songcen Xu, and Hang Xu. Any-size-diffusion: Toward efficient text-driven synthesis for any-size h...

    Pith/arXiv arXiv

  16. [22]

    Our method is in leading position across all metrics

    2K resolution comparison are demonstrated in Figure 12 F QUANTITATIVECOMPARISON ON2K RESOLUTION Table 5 demonstrates the quantitative results for2K×2Kresolution image generation. Our method is in leading position across all metrics. Table 5: Quantitative comparison at 2K × 2K resolution. The best result is highlighted in bold, while the second-best result...

    2025

  17. [2020]

    Flux.1 kontext: Flow matching for in-context image generation and editing in latent space.arXiv preprint arXiv:2506.15742,

    Black Forest Labs. Flux.1 kontext: Flow matching for in-context image generation and editing in latent space.arXiv preprint arXiv:2506.15742,

    Pith/arXiv arXiv

  18. [2021]

    Demofusion: Democratising high-resolution image generation with no $$$

    Ruoyi Du, Dongliang Chang, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. Demofusion: Democratising high-resolution image generation with no $$$. InCVPR, 2024a. Ruoyi Du, Dongyang Liu, Le Zhuo, Qin Qi, Hongsheng Li, Zhanyu Ma, and Peng Gao. I-max: Maximize the resolution potential of pre-trained rectified flow transformers with projected flow. 2024b. Yu ...

    Pith/arXiv arXiv

  19. [2022]

    Linfusion: 1 gpu, 1 minute, 16k image.arXiv preprint arXiv:2409.02097, 2024a

    Songhua Liu, Weihao Yu, Zhenxiong Tan, and Xinchao Wang. Linfusion: 1 gpu, 1 minute, 16k image.arXiv preprint arXiv:2409.02097, 2024a. Xinyu Liu, Yingqing He, Lanqing Guo, Xiang Li, Bu Jin, Peng Li, Yan Li, Chi-Min Chan, Qifeng Chen, Wei Xue, et al. Hiprompt: Tuning-free higher-resolution generation with hierarchical mllm prompts.arXiv preprint arXiv:2409...

    Pith/arXiv arXiv

  20. [2023]

    Pixart-σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation, 2024a

    Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation, 2024a. Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart-δ: Fast and controllable image gene...

    Pith/arXiv arXiv

  21. [2024]

    Relay diffusion: Unifying diffusion process across resolutions for image synthesis.arXiv preprint arXiv:2309.03350,

    Jiayan Teng, Wendi Zheng, Ming Ding, Wenyi Hong, Jianqiao Wangni, Zhuoyi Yang, and Jie Tang. Relay diffusion: Unifying diffusion process across resolutions for image synthesis.arXiv preprint arXiv:2309.03350,

    Pith/arXiv arXiv

  22. [2025]

    Hiflow: Training-free high-resolution image generation with flow-aligned guidance.arXiv preprint arXiv:2504.06232,

    Jiazi Bu, Pengyang Ling, Yujie Zhou, Pan Zhang, Tong Wu, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Dahua Lin, and Jiaqi Wang. Hiflow: Training-free high-resolution image generation with flow-aligned guidance.arXiv preprint arXiv:2504.06232,

    Pith/arXiv arXiv