Wan-Streamer v0.1 End-to-end Real-time Interactive Foundation Models

paper:

Single guys gonna love it for sure and gold diggers rn:

It's almost opensource if someone is willing to contribute an extra 10% of perf.

AI Wifus coming soon!

that's really cool

That's very impressive, its expressions are very realistic. I hope they release the models for us to play around with. Perhaps we can add custom characters to it via training.

need full upper body like at desk or sitting etc so you can see the hand gestures not cut off, so can be composited on scene

another stunning example of Alibaba's new model - Ali-Catfish1.0

avtr-1 is doing this pretty good with a single 3090 in realtime, it renders lip-synced speech and active listening at 25 fps on a single GPU. the only drawback is it only has facial change but not neck rotation etc.

real-time interactive foundation models are the fun part. the ugly part is whether latency, state, controllability, and failure recovery survive past the demo clip. streaming magic still needs boring systems engineering.

she was more expressive than the guy

We need that ASAP in the wan app even if it's a paid feature

amazing man

@grok wtf is this real?

This is really cool. Was thinking about this last week. But didn’t take it as far as full transformer architecture for everything. So this is fascinating I’ll be looking at this to run via webgpu

Final piece in AI wifu!

I like this

is this available to use anywhere?

Real-time interactive AI is impressive. I think the next hard problem is shared-room conversation: multiple people, multiple languages, interruptions, accents, names, and context. Not just one human and one AI. A whole group staying in the same conversation.

@Justiniansli

good idear

so awesome!

Oh gawd, I am real. She’s not real

Model/code release?

Great !

amazing work

Gerçek zamanlı etkileşim için video akışı yeterli değil — sistem lokal girdi işleyebiliyorsa neden AI ajanın web kameraya erişmesi gerekiyor? Latency, gizlilik ve inference maliyeti açısından stream-based mimari daha pahalı.

This so unreal brooo

once end to end delay passes 200ms interactivity feels fake can wan-streamer stay below that

Is it on Comfy yet

Great

我要做一个 @CuiMao 的