amiiari

video generation test 3!

Added 2025-10-26 20:45:03 +0000 UTC

hi again! spent... the entire night (until 9am) messing around with wan2.2 and i've gotten some preliminary results! the other model, hunyuan, was the one for the akira videos!

sadly most of my time was trying to build the proper sageattention and triton optimization packages compatible with blackwell but... nothing... i will try to get those! makes things 2x faster omg need

but yes! videos with a better model! i have a lot more control over things, and can still make them as long as i want, these are just 3 seconds because they're tests! (and generate slower than i'd like to), so yes! let me know what you think!

i just looked around my content and saw stuff i thought would work decently, however the motion lines and hearts that i usually have clash heavily with the model so actual video generations would not have that haha

and! you may see a significant difference in quality for the videos, starting at v (12), the difference between them are some are guided with motion models, and some are not, you can probably tell haha, was just testing capabilities!

i can't upload videos directly! i would upload gifs like i did with akira but .gifs degrade the quality a lot + theres a lot of videos to convert! so... mega, also in my discord server!

thank you all so much!!

the mega!