The generated feet look straight up Lovecraftian and it's a total train wreck.
#3
by skzz - opened
The generated feet look straight up Lovecraftian and it's a total train wreck.
steep memory usage. example doesn't use any memory clean up and the KV cache expands +4 GB during inference.. then adds the VAE on top of that. You can 'del pkv_c / del pkv_u' in the t2i_pipeline and use bitsnbyes too.. in the app add gc.collect();torch.cuda.empty_cache() to generate more than once (i have 24GB vram)
The generated feet look straight up Lovecraftian and it's a total train wreck.
im not trying to add insult to injury but i actually laughed reading "its a total train wreck"... idk why i laughed but i did lmfao made me stop downloading immediately