This LoRA tries to create an "oily anime" style, characteristics of artists like Yohan/Freestyle, J@ck, Shexyo and so on.
The version so far is in the style of these three in specific, with a bit more weight given to Yohan.
The goal is to refine as I find more interesting pictures and artists in this style. To kind of create a "perfect" type of "oily anime" look.
LORA DIRECTORY + COMMON PROMPTS HERE:
Google Sheets link;
PLEASE SUPPORT THE ARTIST/GAME/ANIME! Do not monetize your generations or use the model to do anything that competes with the original author/works (or that gets the model taken down). Use this for casual, personal, small-scale use. Be reasonable!
(Model will be taken down immediately if requested)
Description
FAQ
Comments (4)
Have you tried reducing the file size? 900MB is overkill for a style lora. One way to achieve this is by lowering the Network Dimension/Rank value in whatever lora training tool you're using. I assume you have it set to sth like 128, when 32 is probably enough, without losing details.
Very fair point. This is still an early version. I always start with 128 to capture as much detail as possible and then slowly taper down as I ensure the quality remains. I usually find the ideal rank at 96. I have near unlimited storage so it doesn't make a lot of difference for me, but I assume it does for many others. Tks for the feedback.
@rwwills All good, as long as you know what you're doing. Here's another angle to consider, though: raw storage is not the only thing influenced by lora size. Every lora you add for generation also adds to the necessary vram. In practice that means that a workflow with say a style lora, a character lora, a pose lora, a separate body feature lora and maybe a slider lora of some sorts can easily run you out of memory on medium vram setups (~12gb).
@xnapx Also very true. I mostly use the CivitAI trainer for everything, so I stopped caring about these, but on a personal computer these would be very relevant. Tks for the feedback