Wan 2.2:
This was trained using ai-toolkit on Wan 2.2 I2V 14B using 20 81-frame 16-fps videos.
Wan 2.1:
This model was trained using diffusion-pipe on the Wan I2V 14B 720P checkpoint using ~30 2-second videos from various anal sex positions at 24fps. I used an Nvidia A6000.
I'm aware that it's the general convention to train on the T2V Wan checkpoint, but I get horrendous results (with both T2V and I2V) when I do so with this dataset. My theory is Wan isn't as aware of this type of action as other ones, but I'm not really sure. I just know that I've tried a bunch of times to train this model in T2V to no avail while I've had success training on T2V with other datasets.
Description
FAQ
Comments (5)
Great concept! How do generate the base photo?
I was going through some of my old generated images to find some of this concept, I think I used this to create them and they have a similar look to the ones in his examples. Works with photorealistic PonyV6 checkpoints.
@Degenerator123 This is also what I used. Base model was Pony Diffusion V6 XL. Check the metadata on this image for the full prompt and other loras used: https://civitai.com/images/70872685
Maybe it's my settings and input images, v0.1 is a bit hit-or-miss for me with the amount of motion. It seems like it's possible to combine it with this (POV of girl grinding viewer), has useful motion for this concept, and possibly with this (female orgasm). I haven't had much time to experiment with strengths and other settings. Looking forward to trying out the retrained version you mentioned!
Yeah it takes many generations to get good results with v0.1. Really hoping this next round of training significantly improves the quality!