Next-Dense-Stride Prediction for Multimodal Autoregressive Visual Modeling

One autoregressive backbone for multimodal image generation, cross-modal translation, and segmentation.
Chicago Y. Park1,* · Jialin Mao2 · Xiaojian Xu2 · Taha Kass-Hout2 · Ulugbek S. Kamilov1 · Cao Xiao2
1University of Wisconsin–Madison  ·  2GE HealthCare *This work was done during an internship at GE HealthCare.
Coarse-to-fine generation via progressively denser strides
Motivation. Sparse strides capture global structure, and progressively denser strides add fine detail. DenseAR follows this principle, autoregressively predicting tokens in this coarse-to-fine order — on a single compact latent grid, with no multi-scale token inflation.

We introduce DenseAR, which reformulates autoregressive image generation as coarse-to-fine next-dense-stride prediction on a compact single-scale latent grid.

Traversing the grid with progressively denser strides captures the transition from global structure to fine detail — avoiding both the slow inference of raster-order autoregression (DenseAR decodes multiple tokens in parallel) and the long multi-resolution sequences of multi-scale approaches.

On brain MRI, a single DenseAR model unifies cross-modal translation, modality-conditioned generation, and tumor segmentation. On natural images, DenseAR improves class-conditional ImageNet generation over single-grid and multi-scale-tokenizer baselines.

What makes DenseAR different

Coarse-to-Fine

Next-dense-stride prediction: coarse-to-fine generation on a compact single-scale latent grid.

Fast

Parallel decoding: emits multiple tokens per step, avoiding slow raster-order autoregression.

Compact

No sequence inflation: none of the long multi-resolution token sequences that multi-scale (VAR) models rely on.

Unified

One model, many tasks: MRI cross-modal translation, modality-conditioned generation, and tumor segmentation.

Strong

Competitive: competitive with task-specific methods on brain MRI, and better class-conditional ImageNet (FID/IS) than single-grid and multi-scale baselines.

How DenseAR works

DenseAR encodes an image into a single-scale grid of latent tokens and then visits those tokens in stride order: a sparse set first, followed by progressively denser strides that fill in the gaps. At each step, the model predicts multiple tokens in the current stride group in parallel.

Next-dense-stride prediction: training, inference, and grid visualization
(a) Training — masked positions act as placeholders for tokens not yet generated, and the model learns to predict the next, denser stride group given the tokens already committed. (b) Inference — committed tokens are stored in a KV cache, so each new stride group is decoded efficiently without recomputing the past. (c) Stride order — repeating this over increasingly dense strides reconstructs the full grid coarse-to-fine.
Training

Sampled stride orders

DenseAR keeps the standard autoregressive likelihood, but trains over a randomly sampled stride order rather than a fixed raster path. Tokens within each stride stage are shuffled, so the model learns to predict any position in the next, denser stage from any coarser context — and never sees a sequence longer than the single G×G grid.

Inference

Few-step parallel decoding

At test time DenseAR walks the same coarse-to-fine order. A full grid is produced in a handful of parallel steps instead of hundreds of raster steps — the random within-stage training is exactly what makes this grouping safe.

One model, many tasks

Because DenseAR is autoregressive, every task is just one token sequence assembled from a few special markers — so a shared backbone covers generation, translation, and segmentation without task-specific heads. A [TASK] marker selects the task, a class token precedes each image to name its modality/class, source images are appended as [SRC] blocks, and a [TGT] marker requests the desired output.

Unified multi-task interface: task, source, and target tokens
(a) Multi-task inference. The markers, class tokens, and source-image tokens form the committed prompt held in the KV cache; conditioned on it, the model generates the target image tokens in next-dense-stride order, decoding multiple tokens per step in parallel. (b) Token sequence layout. A single sequence concatenates the task token, source/target class tokens, and source-image tokens, followed by the predicted target-image tokens emitted over progressively denser strides.

One backbone for generation, translation, and segmentation

A single DenseAR model handles three multi-contrast brain-MRI tasks — generation, translation, and segmentation — within one backbone, with no task-specific retraining. Evaluated on BraTS-2023 against strong specialists and unified baselines.

Unified evaluation on BraTS-2023. Metrics per task among unified (generalist) models. Best in bold, second best underlined. DenseAR leads translation and generation, and stays competitive on segmentation — while being the model that covers all three tasks.
Method Translation Generation Segmentation
PSNR↑SSIM↑LPIPS↓ FID↓ Dice↑
Generalists
EditAR20.780.8110.11215.88
MVG20.460.8250.1140.872
DenseAR (Ours)21.830.8330.0998.500.851

Modality-Conditioned Generation

Given a target contrast, DenseAR synthesizes realistic T1, T1ce, T2, and FLAIR scans, with FID competitive with strong generative baselines (LlamaGen, DiT).

Modality-conditioned brain MRI generation

Any-to-Any Translation

From any single input contrast, DenseAR translates to every other contrast, preserving anatomy and lesions across modalities.

Any-to-any cross-modal MRI translation

Tumor Segmentation

The same model segments tumors from any contrast, reaching Dice scores on par with specialized methods (MedSegDiff, nnU-Net).

Brain tumor segmentation across modalities

Class-conditional natural-image generation

Class-conditional samples generated by DenseAR across diverse categories, improving over single-grid and multi-scale-tokenizer baselines.

Class-conditional ImageNet samples
@article{park2026DenseAR,
  title   = {Next-Dense-Stride Prediction for Multimodal Autoregressive Visual Modeling},
  author  = {Park, Chicago Y. and Mao, Jialin and Xu, Xiaojian and Kass-Hout, Taha and Kamilov, Ulugbek S. and Xiao, Cao},
  journal = {arXiv:2607.09892},
  year    = {2026}
}