dNaViT: Discrete Codes are General-Purpose Visual Representations
One discrete visual tokenizer matched or beat task-specific baselines across understanding, generation and image editing, the authors report.
dNaViT builds on a frozen pretrained image encoder, adding lightweight adjustments to preserve both semantic meaning and pixel-level detail. Grouped product quantization expands its discrete vocabulary’s capacity without lengthening spatial token sequences. The same representation supports downstream visual tasks and high-fidelity image decoding. Evaluations cover 12 visual understanding benchmarks, four generation benchmarks and one image editing benchmark.
HF Daily Papers' note
dNaViT builds on a frozen pretrained image encoder, adding lightweight adjustments to preserve both semantic meaning and pixel-level detail. Grouped product quantization expands its discrete vocabulary’s capacity without lengthening spatial token sequences. The same representation supports downstream visual tasks and high-fidelity image decoding. Evaluations cover 12 visual understanding benchmarks, four generation benchmarks and one image editing benchmark.
HF Daily Papers' note
score 5