These look.. great, by and large. Hands are super natural, coherency is really high. Showing off piano chord blocking is a huge flex.
I’d like to play with this! No code, but bytedance often releases models, so I’m hopeful. It’s significantly better than vasa, and looks likely to be an iteration of that architecture.
Github is overflowing with Tencent, Alibaba, and Ant Group models. Typically licensed as Apache 2, and replete with pretrained weights and fine tuning scripts.
Yes, the Llava model can encode image, and you can encode image into 3D vae space. Without fine-tune the model though, you are not going to have fidelity to original (if only use Llava's SigLIP to encode), or end up with image with limited motion (3D vae encoded latents as the first frame then doing vid2vid).
The hands do look good, but the spacing of the black and white keys on the piano keyboard is off. Similar problem with the strings and fretboard on the guitar. Such glitches won’t be noticed by as many people as weird hands are, though.
First video: Disappearing and appearing shirt buttons. Disappearing, appearing, and shapeshifting rings. Ear appears to be bluescreened despite the rest of the person appearing to be in front of a real background. Belt buckle slides unnaturally.
Second video: Shadows reveal inconsistent lighting direction. Disappearing and appearing studs on the watch strap. It also has bizarre clothing design with buttons on a non-opening shirt and what seems to be a printed fake weaving pattern that doesn't actually correspond to real weaving, but this could theoretically be made in reality.
What are the tells in most of these videos? I can't point at any in many of them. Hands, teeth, lip sync, body and should movement all look correct. Specially the TED talk like presentation examples near bottom.
This looks better than EMO (also closed source by Alibaba group https://humanaigc.github.io/emote-portrait-alive/). See the rap example on their page. They apparently have EMO2 now which doesn't look as believable to me.
EMO covers head + shoulders while this OmniHuman-1 is covering full body and its looking even better. I would have easily mistaken these for real (specially while doom scrolling) if I was not looking for AI glitches.
UPDATE: Googling animate bytedance site:github.io returns many in the same domain (all proprietry). Found a few good ones.
This is very good attempt with people playing musical instruments.
But, there are some subtle timing tells, that this is AI generated. Take a look at the singer playing the piano. Timing of the hands with the singer is slightly off. The same goes with the singer and the guitar. I'm not a guitar player or piano player, but I do play a lot of different musical instruments at a high level, and the timing looks off, slightly ahead or behind the actual piece of audio of the piece of music.
They have cracked lip sync. AI can now realistically show singers strumming away out of time on guitars that are not plugged in or drummers playing the incorrect fills.
> The images and audios used in these demos are from public sources or generated by models, and are solely used to demonstrate the capabilities of this research work. If there are any concerns, please contact us (jianwen.alan@gmail.com) and we will delete it in time.
Ethical concerns with this technology have nothing to do with videos on a demo page, and everything to do with what can be generated later.
I don’t know if they have a profound lack of understanding of the ethical implications or are purposefully trying to pretend, but neither is good.
this could be used as an incredible low bitrate codec for some streaming use cases.
(video conferencing/podcasts on <3G for ex, just use some keyframes + the audio).
Modern operating systems should include by default a very simple private/public key system to sign arbitrary files. I think it should not be very complicated? We badly need this in the age of AI.
I’d like to play with this! No code, but bytedance often releases models, so I’m hopeful. It’s significantly better than vasa, and looks likely to be an iteration of that architecture.