ALPHA 2.0
I never released this model since I don't really know, its another thing I've made out of chaos and its good but also bad
not every gen will be a good gen, and your mileage may vary, its about tuning and luck.
Play with it enough and you can get more out of it, it doesn't have to be missionary, though its kinda find some way to bleed elements into the video, fun for experiments though
cant bring myself to monetize but I could probably be persuaded to make other specific motion modules on other videos if you have coin :3
good luck!
Animatediff motionmodel, fine tuned on a single video as a proof of concept. The idea is no different than lora strength. from steps 100-1000 are about the same as their 0.1-1 strength for lora. I'll upload the couple highest because below a certain point you also lose the motion. prompting helps but the finetune wont let you stray too far, but you can do doggystyle as well, I haven't experimented that much.
I'm trying to do one that isn't sex, or maybe other sex, I could use the support :)
NOT TESTED ON EXTENTSION, but I imagine it would work fine.
THIS IS VERY MUCH A TRIAL RUN, I plan to iterate and improve on the design, but I thought it would be cool to share the potential of the next possible stage of animation.
OTHER PLANNED EXPERIMENTS: (updated 8/7)
Train on 1.5 modeltesting seems to suggest almost no real difference despite the 2 being so different in normal use.different weights: I tried some different frame-rates today, needs more testing...
more videos for more variation then just one video: ruined the result, further testing is required.
Description
Not nearly as much movement as 1, but I think this one is best, can keep style and gets the motion, yes you can prompt other things but the motion will always kinda be the same.
FAQ
Comments (13)
Tutorial? :)
using animatediff, drop this safetensor into the models/motion-module/, change in your config to the filename for when you run inference. I'm using animatediff-cli for more depth and other improvements.
@CubeyAI I can't select options other than "mm_sd_v14.ckpt", "mm_sd_v15.ckpt" in AnimeateDiff's options.
The "missionarymotion.safetensors" you provided should go in this option, right?
@Smallmi I did not test this with the extension, you could be sneaky and try renaming it to replace them, I do apologize as I don't have a work around for the webui-extension at this time
awesome :) can u do pronebone as well some time pls ?
you might even be able to prompt it on this one and get lucky
how to use?
This is pretty amazing! Esepcially your 1.0 str one. Could you do a tutorial on how you made the fine tune module? I would love to fine tune my own as well.
sure, my main struggle is currently that my understanding of finetuning is still being researched, the issue is; sometimes it works, othertimes it doesn't... but basically https://github.com/tumurzakov/AnimateDiff/tree/main has the resources to do the finetune, the truth is it involves some understanding in using virtual environments. but if you can get through it and you can run the command much like making loras sd_scripts with stable diffusion.
This is really a great model. But I'm having some problems with this model. Thanks for your attention!
1. What "length, width, and frame" do you use for training and inference? In my inference, the output video may be white in the last few frames. I think the reason is that my "frame" is larger than yours.
2. The last few frames repeat the motion. For example, the output video contains 16 frames, and the last 6 frames are very similar to the first 10 frames.
1. this was trained on a 512x768 video with 16 or 24 frames (I don't quite remember anymore)
2. this was an earlier experiment, the main struggle of animatediff right now is that motion is hard to control, so I don't have a good answer, but I believe better solutions are coming very soon as there is alot of groundbreaking going on.
@CubeyAI Thanks for your answer. I will continue to pay attention to developments in this research area.
Truly incredible model, can generate quite stable fuck animated images so impressive.
I wonder if you would make other motion modules? Your work looks so great :)