Uno studies diffusion augmentation of an autoregressive backbone
A relationship to earlier approaches does not establish redundancy without a technical comparison.
Uno adds trainable diffusion components to a pretrained autoregressive backbone and studies a distribution-preserving sampler. Its reported speedups belong to the paper’s model and decoding comparisons. Keeping the backbone does not mean no trainable components are added.
The original article suggested that similarity to Orthrus undermined the work’s originality, relying on a community comment rather than an explicit comparison of architecture, training and sampling assumptions. That evidence was insufficient for the accusation.
For a decoding system, compare the two original methods on the same target distribution, hardware, acceptance procedure and quality metric before declaring one redundant or interchangeable. Here the useful contribution is a specific augmentation and sampler, not an unsupported judgment about priority.
Editorial correction (2026-10-13 UTC): Removed the unsubstantiated originality accusation and clarified added trainable components.