doogyhatts
Arch-Supremacy Member
- Joined
- Feb 13, 2018
- Messages
- 13,740
- Reaction score
- 3,949
I shifted all the video generation work onto Wan2GP app.nope haven't tried that.. is it any good? I have one workflow that uses the old model. not bad but the new one supposedly has better prompt adherence and movement. but I can't get it to work properly.
Much faster in loading the model and start denoising, compared to ComfyUI.
Hunyuan Video Avatar was the first model I used on the app.
And the output is very good, the lip-sync is accurate and the body motion is sufficiently good, but not Veo3 standard.
Initially, no one wanted to optimize the vram usage, so I had to contact the guy behind Wan2GP and convince him to do it. (Israeli ML expert)
Before that, I had to rent a 5090 to run it, while a 4090 will OOM easily.
Now, I can run it on a 5080 at 832x624 resolution, but the output will be a bit blurry.
After that, I upscale and sharpen it using Topaz Starlight Mini.