This is a INT8 (row-wise) quantization made with "convert_to_quant" script with --int8 --scaling_mode row --anima --low-memory flags. The purpose of this is to squeeze as much vram as possible on low-end GPUs (6gigs) like mine, at least i noticed significant speed up during loading and generating images, so i want to share this with people.
8-16 steps
CFG 1-2
Sampler - Euler/ER_SDE
Scheduler - Simple/Beta57
You can also use int8 qwen3_06b text encoder provided by reakaakasky
https://civarchive.com/models/2760490/anima-int8?modelVersionId=3109193
to squeeze more memory.
AnimaTurbo 1.0 is identical to
https://civarchive.com/models/2458426/anima?modelVersionId=3108589
AnimaYume 1.0 have "Expressive" Turbo Lora by kai
Thanks circlestone_labs for amazing Anima model, duongve13112002 for great AnimaYume finetune, kai for excellent revision of turbo lora.
Description
FAQ
Comments (10)
is this normal int8 or int8_convrot ?
normal (row wise) int8 - because convrot is sadly slower on my pc.
_ because convrot is sadly slower on my pc.
May I ask which gpu are you using and what's the group size of convrot?
@reakaakasky RTX 2060 (yeah i know 6gigs it's nothing), don't know about group size, i just converted checkpoint through https://github.com/silveroxides/convert_to_quant
with this flags
--int8 \ --scaling_mode row \ --anima \ --low-memory
How step you recommanded? cfg1 and 8 step? same as turbo lora?
i recomand cfg3, 12step