As always, if you enjoy this content please kindly like and share!
Firstly, I rotoscoped the subject using After effects. This isn’t necessary but I needed a mask for a later step anyways so - why not right?
I then fed the video with this image into ViggleAI
It gave me this video.
The next steps involved compositing back into the source footage.
This was the step that I need to improve on and have since come up with a better method for doing it involving in painting and gen fill
I then took the composited video and ran it through AnimateDiff using the LCM motion model, photon checkpoint, and Differential diffusion with a combined mask of the joker + the mask I made in the earlier step. I also used a Joker LoRa.
No controlnets.. very simple workflow.
This ended up cleaning it to about 70% of the quality but I had some issues with the jacket morphing into the wrong perspective.
Easy fix though I just cropped a good frame of the jacket, tracked it loosely to the footage, then reran again through ADiff.
The biggest take away is I need to clean the initial composite a bit better. You can see some artifacts of the original actor that I composited over. I’ve since come up with a new method to improve on this that’ll I’ll discuss in the future.
You can use Viggle for FREE in their discord. The crazy thing is this is a 27s video and it legit took less than 5 mins to get this result 🤯 @ViggleAI
Disclaimer: This is more of a traditional animator pipeline. It involves pre-rendering the mouth shapes and then manually placing them in post (Will talk on this later).
This works best for fixed perspective shots which are present in the majority of anime.
2/17
The first step is to choose a still from your animation or just start with a still image. I used the attached image to start.
Bring the image into ComfyUI and simply use a standard img2img workflow with inpainting. Just make a small mask on the mouth area.