Pith. sign in

REVIEW 3 major objections 5 minor 42 references

DenseAR turns a single latent grid into coarse-to-fine image generation by predicting tokens in denser and denser strides.

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 · grok-4.5

2026-07-14 14:46 UTC pith:WDGLKT35

load-bearing objection Solid methods paper: single-grid coarse-to-fine AR order that is cleaner than multi-scale stacks, with a real multi-task medical demo and matched ablations. the 3 major comments →

arxiv 2607.09892 v1 pith:WDGLKT35 submitted 2026-07-10 eess.IV cs.AI

Next-Dense-Stride Prediction for Multimodal Autoregressive Visual Modeling

Chicago Y. Park , Jialin Mao , Xiaojian Xu , Taha Kass-Hout , Ulugbek S. Kamilov , Cao Xiao This is my paper
classification eess.IV cs.AI
keywords autoregressive image generationnext-dense-stride predictionsingle-scale tokenizercoarse-to-fine generationmultimodal medical imagingunified multi-task modelvector-quantized tokensclassifier-free guidance
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.

Autoregressive image models have been stuck between two bad options: raster order is slow because it emits one token at a time, and multi-scale next-scale order is coarse-to-fine but inflates the token sequence and memory. DenseAR claims a third path. It keeps one compact single-scale token grid and simply changes the visit order: first a sparse set of widely spaced tokens that fix global layout, then progressively denser strides that fill in detail, with all tokens inside a stride stage predicted in parallel. The same short-sequence design lets one backbone treat several imaging modalities and tasks as different choices of which grids are given and which are generated. On multi-contrast brain MRI the single model handles cross-modal translation, modality-conditioned generation, and tumor segmentation competitively with specialists; on ImageNet it improves FID and IS over a matched single-grid baseline without stride ordering and uses far less memory than multi-scale competitors of similar quality.

Core claim

Traversing a single-scale latent token grid with a decreasing sequence of strides (next-dense-stride prediction) yields a coarse-to-fine generative order without multi-resolution token stacks or token-by-token raster decoding. Sparse stages set global structure; denser stages add local detail; tokens within each stage can be sampled in parallel from the already-committed coarser context. The resulting short, single-grid sequences make multi-modal, multi-task unification practical inside one autoregressive backbone.

What carries the argument

Next-dense-stride order: a stride pyramid on one G imes G grid that nests sub-grids DK ⊃ DK/2 ⊃ … ⊃ D1, commits stages coarse-to-fine, trains with random within-stage order, and at inference predicts each stage’s new tokens in parallel from a shared KV-cached context.

Load-bearing premise

That tokens inside one stride stage are weakly enough dependent, given only the coarser committed tokens, that sampling them independently in parallel does not materially hurt image quality.

What would settle it

Train an otherwise identical model with the same step budget but force fully sequential (token-causal) decoding inside each stride stage; if FID/IS or medical metrics stay the same, the parallel-stage assumption holds; if they improve, the independence claim fails.

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

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

3 major / 5 minor

Summary. The paper introduces DenseAR, an autoregressive paradigm that generates images by next-dense-stride prediction on a single-scale VQ latent grid: sparse strides fix global structure, then progressively denser strides add detail, with tokens within each stage predicted in parallel. This is intended to retain the compact sequence of raster-order AR while obtaining a coarse-to-fine prior without multi-scale token stacks. The same backbone is extended to a unified multimodal multi-task model via modality/task markers and task-conditional classifier-free guidance. On ImageNet, DenseAR improves FID/IS over a matched single-grid random-order baseline (ARPG) and reports better quality–memory tradeoffs than multi-scale VAR. On BraTS-2023, one model handles cross-modal translation, modality-conditioned generation, and tumor segmentation, remaining competitive with task-specific and prior generalist methods.

Significance. If the results hold, DenseAR is a useful middle path between slow raster AR and memory-heavy next-scale AR, especially when several modalities must share one sequence. The matched-capacity ARPG comparison and the BraTS with/without-stride control (Table 4) give a clean attribution of gains to stride ordering. The unified medical backbone—generation, translation, and segmentation through one token interface without text encoders, multi-scale stacks, or in-context example pairs—is of practical interest, and the planned open-source release of training and inference pipelines is a concrete strength. The work is relevant to both visual generative modeling and multi-task medical imaging.

major comments (3)
  1. [§3.1, Eq. (5)] The efficiency and multi-task sequence design rest on sampling an entire stride stage in parallel from the same committed coarser context (Eq. 5), with safety attributed to random within-stage order during training. This independence premise is load-bearing but only weakly characterized: there is no diagnostic of residual within-stage spatial dependence, nor an ablation that varies the number of tokens decoded per step (or stage size) while holding total steps and capacity fixed, relative to fully causal within-stage decoding. The ARPG and Table 4 results show that stride order helps overall, but they do not isolate the quality cost of the parallel relaxation itself. A short ablation or correlation diagnostic would substantially strengthen the central claim.
  2. [§4.1, Table 2] BraTS evaluation uses an author re-split (1000/11/240) because the official test set lacks segmentation masks, and segmentation metrics exclude near-empty tumor slices (mask area threshold of 2,000 pixels). Both choices are understandable but affect auditability of the “competitive with specialists” claim in Table 2. Please (i) release subject IDs or the split seed, (ii) report generation/translation metrics on the same filtered vs unfiltered slice sets, and (iii) state explicitly that specialist and generalist baselines were trained and scored under identical filtering and preprocessing. Without this, cross-method margins are harder to interpret.
  3. [§3.2, Eq. (7); Appendix A.3] Task-conditional CFG weights are swept per task on validation metrics (Appendix A.3: roughly w=4 generation, w=6 segmentation, w=15 translation). Headline unified-model numbers in Table 2 therefore reflect per-task guidance selection. Report at least one shared-w (or small fixed set of w) result alongside the per-task numbers so the cost of unification without task-specific tuning is visible, and confirm that the sweep used only the validation split.
minor comments (5)
  1. [Figure 1] The caption correctly warns that the figure strides in pixel space for intuition while DenseAR strides on the latent grid; a small latent-grid panel would reduce the risk of over-interpretation.
  2. [Table 5] Memory and step counts are valuable; please confirm peak memory for VAR vs DenseAR under identical batch size, precision, and stack (batch 64, bfloat16 as stated).
  3. [Table 1; §2.3] The “first AR model to span all three” claim is strong; the discussion of MedVAR and AR-Seg already distinguishes single-task next-scale work—make that distinction explicit in the Table 1 caption as well.
  4. [References] LlamaGen is listed twice with the same title ([5] and [13]); consolidate the bibliography.
  5. [§3.1] A short pseudocode block for sampling the stride order π at training time would aid reimplementation of Eqs. (3)–(4).

Circularity Check

0 steps flagged

No significant circularity: empirical AR ordering method trained by standard MLE and evaluated on external benchmarks against independent baselines.

full rationale

DenseAR defines a stride pyramid on a single-scale VQ grid (Eq. 3), trains with ordinary next-token MLE under randomly sampled within-stage orders (Eq. 4), and at inference predicts parallel groups from the same committed coarser context (Eq. 5). The reported FID/IS gains on ImageNet and the multi-task BraTS numbers are measured against held-out data and against independent baselines (LlamaGen, DiT, VAR, ARPG, nnU-Net, etc.). The only mild self-selection is the per-task CFG weight sweep (Fig. 6 / App. A.3), which is ordinary hyper-parameter tuning on validation metrics and does not make any reported number tautological by construction. No uniqueness theorem, self-citation chain, or definitional identity equates the method’s outputs to its inputs. The parallel-within-stage independence premise is an empirical modeling assumption, not a circular derivation.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The work is an empirical generative-modeling paper. Load-bearing content is the stride-order design, the shared VQ token domain across modalities (including masks), and standard AR/CFG machinery. Free parameters are schedule and guidance choices that affect reported metrics; axioms are standard ML and imaging assumptions; invented entities are the ordering scheme and the unified marker-based task interface rather than new physical objects.

free parameters (5)
  • Inference step/group schedule (e.g., 64 steps on ImageNet)
    Number of parallel groups and tokens per step is chosen at inference and directly sets the efficiency–quality operating point reported in Table 5.
  • Task-conditional CFG weight w (generation≈4, segmentation≈6, translation≈15)
    Swept 0.5–15 per task and selected to optimize each task metric (Appendix A.3); main BraTS numbers use these tuned values.
  • Source-dropout probability p_drop for CFG training
    Controls the source-free prior used in task-conditional guidance; set during joint multi-task training.
  • Coarsest stride K and stage pyramid on the G×G grid
    Defines the coarse-to-fine nesting D_K ⊂ … ⊂ D_1; chosen relative to tokenizer grid size (G=16 in experiments).
  • BraTS subject re-split (1000/11/240 train/val/test)
    Official test masks unavailable; authors define their own split, so absolute medical metrics depend on this partition.
axioms (5)
  • domain assumption A single-scale VQ tokenizer yields a discrete G×G grid whose joint distribution can be factorized autoregressively (standard VQ-VAE + AR setup).
    Background in §2.1–2.2; all generation quality is bounded by tokenizer fidelity.
  • ad hoc to paper Within a stride stage, tokens are sufficiently weakly dependent given coarser committed context that parallel independent sampling (Eq. 5) is a valid inference relaxation of fully causal training.
    Core methodological premise of next-dense-stride parallel decoding in §3.1.
  • ad hoc to paper Segmentation masks can share the same VQ codebook and token interface as continuous MRI contrasts without a task-specific head.
    Stated in §4.1 implementation; enables the unified medical backbone claim.
  • domain assumption Classifier-free guidance extends from class labels to task-and-source conditioning via source dropout and the log-linear combination in Eq. 7.
    Builds on Ho & Salimans CFG; task-conditional form is the paper's extension in §3.2.
  • domain assumption Standard perceptual and overlap metrics (FID/IS, PSNR/SSIM/LPIPS, Dice/IoU) are adequate proxies for generation and segmentation quality on ImageNet and BraTS.
    Evaluation protocol in §4.2–4.4.
invented entities (2)
  • Next-dense-stride prediction / stride pyramid order independent evidence
    purpose: Impose coarse-to-fine generation on a single latent grid by nested sub-grids D_k with parallel within-stage decoding.
    Central design object of the paper; not a physical entity but a new ordering construct relative to raster and next-scale orders.
  • Task-conditional classifier-free guidance with modality markers independent evidence
    purpose: Control source adherence per task inside one multi-task AR sequence without extra modules.
    Extension of standard CFG to multi-source medical tasks via [TASK]/[SRC]/[TGT] markers (§3.2).

pith-pipeline@v1.1.0-grok45 · 24726 in / 3632 out tokens · 37438 ms · 2026-07-14T14:46:58.272796+00:00 · methodology

0 comments
read the original abstract

We introduce DenseAR, a new generative paradigm that reformulates autoregressive image generation as coarse-to-fine next-dense-stride prediction using a compact single-scale tokenizer. Our key insight is that traversing a single-scale latent grid with progressively denser strides naturally captures the transition from global structure to fine detail. This addresses two limitations of existing autoregressive models at once: the slow inference of raster-order autoregression, which DenseAR avoids by predicting multiple tokens in parallel, and the heavy cost of multi-scale approaches, which need long, multi-resolution token sequences to achieve coarse-to-fine prediction. Building on our efficient framework and the flexibility of autoregressive modeling, we further extend DenseAR to a unified model that handles multiple modalities and imaging tasks within a single backbone. We validate DenseAR on both medical and natural images. On multi-contrast brain MRI, a single DenseAR model unifies cross-modal translation, modality-conditioned generation, and tumor segmentation, while remaining competitive with task-specific methods. On ImageNet, DenseAR improves class-conditional generation quality (FID and IS) over both a single-grid baseline without stride ordering and a multi-scale tokenizer-based baseline.

Figures

Figures reproduced from arXiv: 2607.09892 by Cao Xiao, Chicago Y. Park, Jialin Mao, Taha Kass-hout, Ulugbek S. Kamilov, Xiaojian Xu.

Figure 1
Figure 1. Figure 1: Intuition for how stride density controls coarse-to [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visual illustration of DenseAR’s next-dense-stride prediction. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Unified multi-task inference. (a) Multi-task inference. A task is one token sequence: a task marker ([TASK]), zero or more source modalities each with a class marker ([SRC]), and a target after [TGT]. This prefix plus the target tokens decoded so far forms the committed context, which the model reads from the key–value (KV) cache to predict the next group of target tokens in parallel; token colors match (b… view at source ↗
Figure 4
Figure 4. Figure 4: Visual comparison of translation and segmentation results for representative methods from Table [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Generated samples from DenseAR-XL with 64 steps. More samples are in Appendix [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Effect of the task-conditional guidance weight [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: All-pairs single-source translation with a single DenseAR model on BraTS-2023. Each 4 × 4 block shows one subject: rows index the source contrast and columns the target contrast (T1, T1ce, T2, FLAIR). The cell marked Source on the diagonal is the input given to the model; every off-diagonal cell is DenseAR’s prediction of that target contrast. A single model produces all twelve source–target translations. … view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison across methods on BraTS dataset. Generated multi-contrast MRI samples from LlamaGen [13] (generation-only, raster autoregressive), DiT [32] (generation-only, latent diffusion), and DenseAR (unified, next-dense-stride autoregressive). 15 [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Tumor segmentation across methods and input modalities on BraTS-2023. Each block group’s examples are shown by the input contrast (T1, T1ce, T2, FLAIR); columns show the input, MedSegDiff-V2 [37] , nnU-Net [36], DenseAR, and the ground truth, with Dice reported per example. The pink mask is the predicted segmentation and the green mask is the ground truth. DenseAR produces masks comparable to the task-spec… view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparison across methods on ImageNet. Class-conditional samples from LlamaGen-XL [5] (raster order), VAR-d20 [6] (next-scale order), and DenseAR-XL (next-dense-stride order). 17 [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Effect of model scale. Class-conditional samples from DenseAR-L and -XL under identical class labels. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Step-by-step inference of DenseAR (illustration only). Here a 4 × 4 token grid (16 tokens) is generated over four steps, predicting 3, 4, 4, and 5 tokens per step. The number of steps and the tokens predicted at each are fixed by the inference schedule; these particular values and the 4 × 4 grid are illustrative, not the sizes used in practice. (a) Full token generation. At each step, the model predicts t… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

42 extracted references · 12 linked inside Pith

  1. [1]

    Denoising diffu- sion probabilistic models,

    Jonathan Ho, Ajay Jain, and Pieter Abbeel, “Denoising diffu- sion probabilistic models,” inNeurIPS, 2020

    2020

  2. [2]

    Score-based generative modeling through stochastic differential equations,

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole, “Score-based generative modeling through stochastic differential equations,” inICLR, 2021

    2021

  3. [3]

    Ran- dom walks with Tweedie: A unified view of score-based diffusion models [in the spotlight],

    Chicago Y . Park, Michael T. McCann, Cristina Garcia- Cardona, Brendt Wohlberg, and Ulugbek S. Kamilov, “Ran- dom walks with Tweedie: A unified view of score-based diffusion models [in the spotlight],”IEEE Signal Processing Magazine, vol. 42, no. 3, pp. 40–51, 2025

    2025

  4. [4]

    Deep parameter interpolation for scalar conditioning,

    Chicago Y . Park, Michael T. McCann, Cristina Garcia- Cardona, Brendt Wohlberg, and Ulugbek S. Kamilov, “Deep parameter interpolation for scalar conditioning,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026, pp. 3884–3892

    2026

  5. [6]

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

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Li- wei Wang, “Visual autoregressive modeling: Scalable image generation via next-scale prediction,” inNeurIPS, 2024

    2024

  6. [7]

    Randomized autoregressive visual generation,

    Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang- Chieh Chen, “Randomized autoregressive visual generation,” inICCV, 2025

    2025

  7. [8]

    Tam- ing transformers for high-resolution image synthesis,

    Patrick Esser, Robin Rombach, and Bjorn Ommer, “Tam- ing transformers for high-resolution image synthesis,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 12868–12878

    2021

  8. [9]

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

    Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu, “Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis,” 2024

    2024

  9. [10]

    Cyclevar: Repurposing autoregressive model for unsuper- vised one-step image translation,

    Yi Liu, Shengqian Li, Zuzeng Lin, Feng Wang, and Si Liu, “Cyclevar: Repurposing autoregressive model for unsuper- vised one-step image translation,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025, pp. 15194–15204

    2025

  10. [11]

    Visual autoregres- sive modeling for instruction-guided image editing,

    Qingyang Mao, Qi Cai, Yehao Li, Yingwei Pan, Mingyue Cheng, Ting Yao, Qi Liu, and Tao Mei, “Visual autoregres- sive modeling for instruction-guided image editing,”arXiv preprint arXiv:2508.15772, 2025

    arXiv 2025

  11. [12]

    Neural discrete representation learning,

    Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu, “Neural discrete representation learning,” inAdvances in Neu- ral Information Processing Systems, 2017, vol. 30

    2017

  12. [13]

    Autoregressive model beats diffusion: Llama for scalable image generation,

    Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan, “Autoregressive model beats diffusion: Llama for scalable image generation,”arXiv preprint arXiv:2406.06525, 2024

    Pith/arXiv arXiv 2024

  13. [14]

    Maskgit: Masked generative image transformer,

    Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman, “Maskgit: Masked generative image transformer,” inCVPR, 2022

    2022

  14. [15]

    Parallelized autoregressive visual generation,

    Yuqing Wang, Shuhuai Ren, Zhijie Lin, Yujin Han, Haoyuan Guo, Zhenheng Yang, Difan Zou, Jiashi Feng, and Xihui Liu, “Parallelized autoregressive visual generation,”arXiv preprint arXiv:2412.15119, 2024

    Pith/arXiv arXiv 2024

  15. [16]

    Autore- gressive image generation with randomized parallel decoding,

    Haopeng Li, Jinyue Yang, Guoqi Li, and Huan Wang, “Autore- gressive image generation with randomized parallel decoding,” inThe Fourteenth International Conference on Learning Rep- resentations, 2026

    2026

  16. [17]

    Adding conditional control to text-to-image diffusion models,

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala, “Adding conditional control to text-to-image diffusion models,” in ICCV, 2023

    2023

  17. [18]

    T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models,

    Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan, “T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models,”arXiv preprint arXiv:2302.08453, 2023

    Pith/arXiv arXiv 2023

  18. [19]

    Composer: Creative and controllable image synthesis with composable conditions,

    Lianghua Huang, Di Chen, Yifei Liu, Yujun Shen, Deli Zhao, and Jingren Zhou, “Composer: Creative and controllable image synthesis with composable conditions,”arXiv preprint arXiv:2302.09778, 2023

    Pith/arXiv arXiv 2023

  19. [20]

    UniVG: A generalist diffusion model for unified image generation and editing,

    Tsu-Jui Fu, Yusu Qian, Chen Chen, Wenze Hu, Zhe Gan, and Yinfei Yang, “UniVG: A generalist diffusion model for unified image generation and editing,” inICCV, 2025

    2025

  20. [21]

    MedM2G: Unifying medical multi-modal generation via cross-guided diffusion with visual invariant,

    Chenlu Zhan, Yu Lin, Gaoang Wang, Hongwei Wang, and Jian Wu, “MedM2G: Unifying medical multi-modal generation via cross-guided diffusion with visual invariant,” inCVPR, 2024

    2024

  21. [22]

    Adaptive latent diffusion model for 3D medical image to image translation: Multi- modal magnetic resonance imaging study,

    Jonghun Kim and Hyunjin Park, “Adaptive latent diffusion model for 3D medical image to image translation: Multi- modal magnetic resonance imaging study,” inWACV, 2024

    2024

  22. [23]

    A unified framework for multimodal image reconstruction and synthesis using denoising diffusion models,

    Weijie Gan, Xucheng Wang, Tongyao Wang, Wenshang Wang, Chunwei Ying, Yuyang Hu, Yasheng Chen, Hongyu An, and Ulugbek S. Kamilov, “A unified framework for multimodal image reconstruction and synthesis using denoising diffusion models,”arXiv preprint arXiv:2602.08249, 2026

    arXiv 2026

  23. [24]

    Editar: Unified conditional generation with autoregressive models,

    Jiteng Mu, Nuno Vasconcelos, and Xiaolong Wang, “Editar: Unified conditional generation with autoregressive models,” arXiv preprint arXiv:2501.04699, 2025

    Pith/arXiv arXiv 2025

  24. [25]

    Medvar: Towards scalable and efficient medical image generation via next-scale autoregressive prediction,

    Zhicheng He, Yunpeng Zhao, Junde Wu, Ziwei Niu, Zijun Li, Bohan Li, Lanfen Lin, and Yueming Jin, “Medvar: Towards scalable and efficient medical image generation via next-scale autoregressive prediction,” 2026

    2026

  25. [26]

    Autoregressive medical image segmentation via next-scale mask prediction,

    Tao Chen, Chenhui Wang, Zhihao Chen, and Hongming Shan, “Autoregressive medical image segmentation via next-scale mask prediction,” inMICCAI, 2025

    2025

  26. [27]

    Medical vision generalist: Unifying medical imaging tasks in context,

    Sucheng Ren, Xiaoke Huang, Xianhang Li, Junfei Xiao, Jieru Mei, Zeyu Wang, Alan Yuille, and Yuyin Zhou, “Medical vision generalist: Unifying medical imaging tasks in context,” arXiv preprint arXiv:2406.05565, 2024

    Pith/arXiv arXiv 2024

  27. [28]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    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:2010.11929, 2020

    Pith/arXiv arXiv 2010

  28. [29]

    Visual autoregressive modeling for image super-resolution,

    Yunpeng Qu, Kun Yuan, Jinhua Hao, Kai Zhao, Qizhi Xie, Ming Sun, and Chao Zhou, “Visual autoregressive modeling for image super-resolution,”arXiv preprint arXiv:2501.18993, 2025

    Pith/arXiv arXiv 2025

  29. [30]

    Restorevar: Visual autoregressive generation for all-in-one image restoration,

    Sudarshan Rajagopalan, Kartik Narayan, and Vishal M. Patel, “Restorevar: Visual autoregressive generation for all-in-one image restoration,” 2025

    2025

  30. [31]

    Classifier-free diffusion guidance,

    Jonathan Ho and Tim Salimans, “Classifier-free diffusion guidance,” inNeurIPS Workshop, 2021

    2021

  31. [32]

    Scalable diffusion models with transformers,

    William Peebles and Saining Xie, “Scalable diffusion models with transformers,” inProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, 2023, pp. 4195–4205

    2023

  32. [33]

    Image-to-image translation with conditional adversarial net- works,

    Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros, “Image-to-image translation with conditional adversarial net- works,”CVPR, 2017

    2017

  33. [34]

    Unpaired image-to-image translation using cycle-consistent adversarial networks,

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” inComputer Vision (ICCV), 2017 IEEE International Conference on, 2017

    2017

  34. [35]

    Multiscale structure-guided latent diffusion for multimodal mri translation,

    Jianqiang Lin, Zhiqiang Shen, Peng Cao, Jinzhu Yang, Os- mar R. Zaiane, and Xiaoli Liu, “Multiscale structure-guided latent diffusion for multimodal mri translation,”arXiv preprint arXiv:2603.12581, 2026

    arXiv 2026

  35. [36]

    nnu-net: a self-configuring method for deep learning-based biomedical image segmenta- tion,

    Fabian Isensee, Paul F. Jaeger, Simon A. A. Kohl, Jens Pe- tersen, and Klaus H. Maier-Hein, “nnu-net: a self-configuring method for deep learning-based biomedical image segmenta- tion,”Nature Methods, vol. 18, no. 2, pp. 203–211, 2021

    2021

  36. [37]

    Medsegdiff-v2: Diffusion based medi- cal image segmentation with transformer,

    Junde Wu, Wei Ji, Huazhu Fu, Min Xu, Yueming Jin, and Yanwu Xu, “Medsegdiff-v2: Diffusion based medi- cal image segmentation with transformer,”arXiv preprint arXiv:2301.11798, 2023. 10

    Pith/arXiv arXiv 2023

  37. [38]

    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,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2009, pp. 248–255

    2009

  38. [39]

    The multimodal brain tumor image segmentation benchmark (brats),

    Bjoern H. Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya Bur- ren, et al., “The multimodal brain tumor image segmentation benchmark (brats),”IEEE Transactions on Medical Imaging, vol. 34, pp. 1993–2024, 2015

    1993

  39. [40]

    Identifying the best machine learning algorithms for brain tumor segmentation, progression assessment, and overall survival prediction in the BRATS challenge,

    Spyridon Bakas, Mauricio Reyes, Andras Jakab, Stefan Bauer, Markus Rempfler, Alessandro Crimi, Russell Takeshi Shino- hara, et al., “Identifying the best machine learning algorithms for brain tumor segmentation, progression assessment, and overall survival prediction in the BRATS challenge,”arXiv preprint arXiv:1811.02629, 2018

    Pith/arXiv arXiv 2018

  40. [41]

    Autoregressive image generation without vector quantization,

    Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He, “Autoregressive image generation without vector quantization,” inNeurIPS, 2024

    2024

  41. [42]

    Neighboring autoregres- sive modeling for efficient visual generation,

    Yefei He, Yuanyu He, Shaoxuan He, Feng Chen, Hong Zhou, Kaipeng Zhang, and Bohan Zhuang, “Neighboring autoregres- sive modeling for efficient visual generation,”arXiv preprint arXiv:2503.10696, 2025

    Pith/arXiv arXiv 2025

  42. [43]

    Randar: Decoder-only autoregressive visual generation in random orders,

    Ziqi Pang, Tianyuan Zhang, Fujun Luan, Yunze Man, Hao Tan, Kai Zhang, William T. Freeman, and Yu-Xiong Wang, “Randar: Decoder-only autoregressive visual generation in random orders,”arXiv preprint arXiv:2412.01827, 2024. A Additional Results We provide additional results across all four tasks: modality- conditioned medical image generation, cross-modal me...

    Pith/arXiv arXiv 2024