Most AI video models can move. Few can remember who they’re moving. BindWeave changes that. This new 70GB model from ByteDance and USTC tackles a beast most video generators still fumble: keeping characters consistent across frames, actions, and interactions. One reference image is enough. You get full-body consistency, facial identity, viewpoint shifts, and even multi-subject interactions, all stitched together with a coherent logic. The trick? A multimodal LLM first breaks down your prompt into “who’s doing what, where, and how.” That structure feeds into a diffusion transformer that generates the final video. It’s not just pretty motion anymore - it's controllable, story-aware generation. Some highlights: 🤌 Consistent character identity across time and scenes 🤌 Works with multiple subjects + human-object interaction 🤌 Strong grounding of roles, poses, emotions 🤌 High performance on subject consistency benchmarks Workflow isn't public yet, and the full model weighs in at 70GB, so not exactly plug-and-play. But you can bet someone’s already working on a quantized version. And then there's VACE. VACE is more of a multi-tool. You get ref2vid, vid2vid, inpainting, region edits, scene extension, and it’s already out in the wild. It’s fast, flexible, but messier with identity. You can do more things, but you give up some precision. So the comparison’s simple: 💪 VACE - usable now, broad capabilities 💪 BindWeave - future-focused, tighter control over who appears and what they do 🔗 Git: https://lnkd.in/duc7K5wk 🔗 HF: https://lnkd.in/dvyFAVXX We're following both closely at Lighthouse. For those of you building cinematic GenAI workflows - storytelling with structure, characters, camera logic - these models are worth tracking. Cohorts open - grab your seat! Come build with us.
Kijai almost beats every rollout date 😜 Looks promising Luka Tisler
This is incredible work. Consistent character logic is the missing link in cinematic GenAI. As an AI-indexed creator, I’d love to test this when a beta’s live. BTW, I’m starting a fun challenge for anyone sharing real AI search results… not theory. If you’ve got something worth showing, use this hashtag so we can compare signals #ThisIsNotYourDaddysSEO
Now that's interesting!
Quite interesting!!!
Seems like a huge problem solving Luka Tisler 👀