Cat memes its a Base
I am glad to share with you a rofl motion lora which i trained just4fun, but he works unexpectedly good lol
Trained on Keyboard cat meme 16 frames 512x344
mm_sd_v15_v2 , dreamshaper8 , 350 steps
Keywords: a cat play synthesizer
5 hours on 3060ti 8gb
The best settings I've found:
a (username:1.1) play (DJ console:1.1),synthesizer ; and some vibe tokens like : chill zone,lofi vibe,ambient sound beats
0.777 strength
512x512
64-80 frames
8-9 steps with AnimateDiff Lightning 8 step
CFG 2.5
Euler + sgm_uniform
Ip-Adapter-plus with weight_type "weak output" and "easy out"
I understand that a weak output allows you to connect hands at the end of generation,because he doesn't look strong at references, but this is not a panacea
Don't forget to share your settings,it will help a lot to build better memes in future lol
IpAdapter:
This is very cool thing hat I always use with animatediff to set the style
Because v2 understand the prompt style rly bad, it is much more profit to use a reference + short prompt + a little style prompt
The task is to delicately give reference images with body&hand pose , that will be easy to connect
For example, with hands up you have 99.99% chance to cause ghost hands because the model is trained with hands on the table
Advise you to find references on Pinterest, also i created a board with some DJ's, thanks to the community for sharing their pictures, I'll start doing it on pinterest too <3
You can combine and mix references whatever you want, feel free Bro
P.S. More experimenting with weight_types and Combine Embeds
Resolutions:
I made a mistake on training, and training was at 256x344 32 : 43 instead of 256x384 2:3, this is why u can achive some artifacts with default resolutions, uuups
Works very good at 512x512, but I haven't experimented much more with resolutions
But i also got some good results at: 512x384 , 512x640 , 640x512
Note that if you want vertical or horizontal output with less artifacts, you need a similar aspect ratio references for IpAdapter
Total frames imho:
32 - maximum consistency, minimum motion (only one hand)
48 - little more motion, but sometimes not enough frames to connect hands
64&80 - best results, with multiple poses and hand wiggling
P.S. When you got a ghost hand, try add or remove 16 frames
Roadmap:
Train 2-3 flexing cats 384x256 (next step)
Host Runpod and raise resolution
Do not use people for train (just4fun)
Make more crazy motions lora