**looks like youtube age restricted it, click through to view it**
but, the current tech doesn't get much better than this. So I'll keep an eye on new things that come out in this space but for now it is a rather time consuming process for a kinda mediocre result. So I'll likely be sticking with still images for now.
Could you tell how much of it is input graphics and how much is really generated out of nothing? (Just a subjective percentage guess)
I'm so excited how this will develop. We should give this Video a very special residence on this page or at some other noticable place. This is history
Blackcat23 said: Could you tell how much of it is input graphics and how much is really generated out of nothing? (Just a subjective percentage guess)
The short answer is about 30% input and 70% AI output (...ish)
The long nerdy answers is ..... the character is completely AI generated (as is the gunge). However it still requires reference images for the controlnet part of stable diffusion to be able to create an increasing amount of gunge pouring down. In this circumstance I took an existing gunge video I already shot and took out about 300 frames. This was fed into controlnet to allow it to make reference depth maps. Then I generate the character through AI and take that as the reference image for the rest of the animation. The system then uses the ai prompt, the initial ai character and the reference frames to produce the video. Total ouput was around 700 frames I think.
It's not an ideal process yet and needs some actual AI video models and tools to be developed before this becomes common. Currently it's like AI image generation 700 times over with hacks to try to keep things as similar looking between frames as possible.
Good. But it still looks more like a toon than a real video. I still dream with the day I'll be able to take a pic of a stewardess, type "pie in her face" and -splat! - the stewardess appears pied in the face...
rt2018 said: Good. But it still looks more like a toon than a real video. I still dream with the day I'll be able to take a pic of a stewardess, type "pie in her face" and -splat! - the stewardess appears pied in the face...
The anime style was actually a choice. I didn't think the shakiness would suit a photorealistic style of character
SlimyDiffusion said: Fantastic work! The coherence is especially impressive, which model did you use for the frames?
I use my own trained ckpt and lora files in stable diffusion. There was also a little from a japanese girl character from civitai mixed in there as well.
Love you, too
