- Who it helps
- Researchers building models that both interpret and generate images
- Availability
- November 9 paper and method page are public. Exact downloadable artifacts and their deployment requirements were not verified.
The representation
A small residual modulator adjusts the frozen features. Grouped quantization divides those features into groups and maps each group to a learned codebook, producing discrete visual tokens.
Evidence and scope
The authors report evaluations spanning understanding, generation and editing. These are reported research results; the inspected text does not establish performance on every image type or a ready-to-use editing product.
What you can do
- Read the grouped quantization method and confirm downloadable artifacts before planning an experiment. Open reference ↗
Where the evidence stops
Primary method text and abstract checked November 11, 2026. No training, inference, artifact installation or benchmark reproduction performed.
Sources read. Public source text checked November 11, 2026 UTC. No third-party project or complete workflow executed.
Sources & further reading
Paper · paper
dNaViT paper ↗
November 9 research abstract describes a shared discrete visual representation.
Xiaotong Li and coauthors · page checked 2026-11-11 · Published 2026-11-09
Web · official
dNaViT method ↗
Frozen encoder, residual modulator and grouped quantization explained; reported benchmarks not reproduced.
Meituan LongCat · page checked 2026-11-11
More sources reviewed (1)
Hugging Face · discussion
dNaViT paper discussion ↗
Submitter comment read; release links and execution were not verified.
XiaotongLi97 · page checked 2026-11-11
Coverage note: Primary method text and abstract checked November 11, 2026. No training, inference, artifact installation or benchmark reproduction performed.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
dNaViT studies a shared discrete image representation for visual understanding, generation and editing. It adapts a frozen image encoder instead of training the entire encoder again.