LaViDa

Summary

LaViDa is a family of masked discrete diffusion vision-language models introduced by LaViDa. It combines a vision encoder with LLaDA-8B or Dream-7B diffusion language backbones and generates text through iterative masked-token refinement.

Family Lineage

flowchart LR
  L[LaViDa<br/>visual understanding and infilling]
  O[LaViDa-O<br/>understanding + generation + editing]
  S[Sparse-LaViDa<br/>cached sparse decoding]
  R[LaViDa-R1<br/>multitask reasoning post-training]
  L --> O
  O --> S
  O --> R

Sparse-LaViDa and LaViDa-R1 are separate continuations of LaViDa-O: the former changes the sampling parameterization for efficiency, while the latter changes post-training for reasoning.

Official Artifacts

• Preprint: arXiv 2505.16839
• Official project page: LaViDa
• Official code: jacklishufan/LaViDa
• Official Hugging Face collection: LaViDa-1.0
• Official training data: jacklishufan/lavida-train

Role In The Wiki

Use this page as the object card for the LaViDa family. The source pages carry paper-level evidence and limitations. The lineage is an upstream multimodal and inference-efficiency signal, not direct evidence for numeric time series or action-conditioned world models.

Related Pages

• LaViDa-O
• Sparse-LaViDa
• LaViDa-R1
• Diffusion Language Models
• Unified Multimodal Models
• Vision-Language Models
• Inference Dynamics