Scale-Wise VAR
A theoretical connection between scale-wise autoregressive generation and discrete diffusion motivates iterative visual refinement.
Scale-wise generation as discrete diffusion
Scale-wise VAR generates visual tokens from coarse scales to finer scales. This work shows that, under a Markovian attention mask, its generation process is mathematically equivalent to a discrete diffusion process.
The attention-mask condition is central to the equivalence. It provides a principled way to connect autoregressive token prediction with diffusion-style refinement.
How it works
- Express next-scale prediction as a transition between discrete visual-token representations.
- Apply the Markovian attention mask and relate the resulting transitions and training objective to discrete diffusion.
- Use this interpretation in Scalable Visual Refinement with Discrete Diffusion (SRDD) to bring iterative refinement into the scale-wise generation framework.
Results
The main theoretical result is the equivalence under the specified attention mask.
The preprint reports improvements in generation and efficiency, and demonstrates iterative refinement and zero-shot reconstruction through the diffusion interpretation.
Contributions
- Established an equivalence between scale-wise VAR and discrete diffusion under a Markovian attention mask.
- Used the connection to develop Scalable Visual Refinement with Discrete Diffusion (SRDD).
- Co-first author with Amandeep Kumar.
Scale-Wise VAR is Secretly Discrete Diffusion. arXiv preprint 2025.