This mix is a strange combination of turbo and style loras. Its goal is to quickly generate images with VERY limited resources. I don't really understand what I'm doing, since i'm experimenting and constantly testing things.
The models in INT8 (row_wise) for those who can't upgrade CUDA for convrot acceleration.
I hope you will find this mix useful.
Versions
T1 - Imagine throwing a bunch of LoRAs into a blender and mixing together several checkpoints—that’s exactly what this is. I don’t even remember exactly what’s in there.
T2 - So i saved list of loras that i like and mixed weights, based on Anima 2.9b (sadly this model is slower due to extended blocks and knowledge).
T3 - I discovered SVD merging not just lora stacking and saving model and picked the same loras as in T2, and the AnimaYume 1.5 came out so it's based on this version.
T4 - I decided to start from scratch, reviewed loras, ditched old ones and merged with SVD with higher rank, i don't know what else i can do, so this version might be final if i don't find something interesting again.
Usage
Minimal steps is 6, but 8-12 obviously better.
CFG-1 - as for all Turbo loras and models.
Sampler - Euler (too chaotic for my taste) or ER_SDE (stable).
Sheduler - Simple or Beta57
I haven't tested high resolutions because I'm used to generating at 1024 and prefer inpainting or Face Detailer over upscaling and Hires.fix.
Description
Based on Anima 2.9B
FAQ
Comments (2)
The new T2 (T1 works fine) doesn't work for me. Is it convrot? I haven't updated comfyui to that yet. I just get static images. I've tried things that worked in past models; sampling discrete vpred true, sd3 model, and auraflow sampling nodes. No dice. Tried also with no loras, still static. Used Euler and beta scheduler
You need to update Comfy to the last commit yes, also on huggingface page of Anima 2.9B, you might want to check author's page there is a patch of Load Diffusion Model node - you need to place it in custom nodes, should've mention that in the description (aah etto bleh).
It's not convrot - regular row_wise int8 for older gpus, convrot is about 10 seconds slower on my setup.