Nemotron-Labs-Diffusion-Uni: Elastic Self-Refining Unified Diffusion LLM with Cross-Paradigm On-Policy Distillation
NLD-Uni lets image tokens undergo progressive refinement instead of replacing them outright with blank masks during training.
The model combines visual understanding and image generation in a single discrete diffusion language model. Its distillation framework transfers capabilities from autoregressive vision-language models and continuous flow models. Gradual token corruption during training enables intermediate token commitments and further refinement at inference. The authors report state-of-the-art results among unified diffusion language models across visual question answering, multimodal reasoning, object grounding, text-to-image generation and image editing.
HF Daily Papers' note
The model combines visual understanding and image generation in a single discrete diffusion language model. Its distillation framework transfers capabilities from autoregressive vision-language models and continuous flow models. Gradual token corruption during training enables intermediate token commitments and further refinement at inference. The authors report state-of-the-art results among unified diffusion language models across visual question answering, multimodal reasoning, object grounding, text-to-image generation and image editing.
HF Daily Papers' note
score 5