LaViDa-O
Summary
LaViDa-O is the unified understanding-and-generation continuation of LaViDa. It represents output images with discrete VQ tokens and couples an 8B understanding branch to a 2.4B generation branch through Elastic-MoT.
Interface
- Inputs: text prompts plus semantic image embeddings and, for editing, image VQ tokens.
- Outputs: text tokens, bounding-box tokens, or image VQ tokens.
- Tasks: image understanding, object grounding, text-to-image generation, image editing, planning, and self-reflection.
- Compute routing: 8B for understanding, 6.4B active parameters for image generation, and 10.4B for interleaved tasks in the reported design.
Official Artifacts
- Preprint: arXiv 2509.19244
- Official project page: LaViDa-O
- Official code: adobe-research/LaVida-O
- Official model: jacklishufan/LaViDa-O-v1.0
Role In The Wiki
LaViDa-O is a useful masked-diffusion counterpart to autoregressive and continuous-flow unified multimodal models. Its planning/reflection loop is inference-dynamics evidence, while its Elastic-MoT is a task-dependent compute-allocation pattern. Neither should be treated as direct evidence for action-conditioned world models.