ByteDance open-sources Lance for six image and video tasks
Lance handles text-to-image and text-to-video generation, editing, and understanding in one framework. Its dual-stream MoE uses modality-aware rotary positional encoding and has 3B active parameters. ByteDance reported scores of 85.11 on VBench, 0.90 on GenEval, and 62.0 on MVBench, leading comparable unified models in its tests. Weights are on Hugging Face under Apache 2.0.
Visit original source