DenseAR speeds up autoregressive image generation with stride prediction

DenseAR tackles two key bottlenecks in autoregressive visual modeling: the slow sequential inference of raster-order prediction and the high cost of multi-scale tokenizers. By introducing next-dense-stride prediction, the model traverses a single-scale latent grid with progressively denser strides, capturing global structure first then fine details. This allows parallel token prediction, dramatically speeding generation while keeping a compact representation.

The model's flexibility extends to multiple modalities and tasks within a single backbone. On multi-contrast brain MRI, a single DenseAR model handles cross-modal translation, modality-conditioned generation, and tumor segmentation—matching task-specific models. On ImageNet class-conditional generation, DenseAR achieves better FID and IS than both a single-grid baseline without stride ordering and a multi-scale tokenizer baseline. This breakthrough could accelerate medical imaging workflows and general image synthesis.