ByteDance Lance unifies image and video generation, editing, and understanding
ByteDance Research released Lance, a multimodal model with 3B active parameters based on Qwen2.5-VL-3B-Instruct. It generates, edits, and interprets images and video in one framework. Lance scored 0.90 on GenEval and 85.11 on VBench, competitive with models of 7B parameters or more. It was trained on 128 A100 GPUs and released under Apache 2.0.
Visit original source