Anima have been here for 2 months by now, but I've only now got access to A4000 gpus ,thank you mr astromahdi from gpu.garden
Please support the original artist: yoshi_houndbell
The style is pretty special, you can prompt "outside border" and a color background tag like "yellow background" to create the window-ish border containing a background type image that you'll see often in my preview images for the LoRA.
(V2): Uh, so I went a bit crazy trying to find a way to improve the forgetting issue and I think i found at least one way to do that on the trainer I'm using, this config could be better for style accuracy as well (although I'm still feel it's prone to overfitting a bit,please feel free to lower weight if needed).
Yapping:
But uh, so the things is, Anima learn pretty well and is capable, but it has a unimaginably bad problem that may ruin its long term success (well beside the bad license that is): the model will begin to catastrophically forget past knowledge (styles, characters, basic concepts like tanks - breast size, background types,...) the more you train it, this applies to both lora training and finetuning. From what i've understand, this is due to the dev (Tdrussel) decision to use a small ahh adapter for text encoder to attach the Qwen LLM to T5 tokens , and every anime knowledge the model has learned is crammed into that tiny adapter. So the model is being fried and tortured beyond belief, and cant handle new knowledge much longer. This is proved by removing the adapter from the model, in which the model revert back to its base knowledge (looking like medieval ahh realistic gens on danbooru tag prompting).
.....
So yeah, very fun, yap over.
Suggested Tags: holding sword, black gloves, parted lips, black hair, very long hair, long hair, black pants, black jacket, blue eyes, medium breasts
Disclaimer
I respect the artists who honed their crafts for years to achieve these styles among all else, if any artists whom LoRA i've made from shown displeasure/objection to their arts being used as AI training material or if they request a takedown notice on Danbooru , I'll delete the LoRA with no further notice.
Description
FAQ
Comments (6)
So jackknife for anime is also in the plan?
oh wow, didnt think anyone would want it, but i still have the OG dataset so sure! From my experience genning on Anima Preview 2 I think most style in the 120-200 image data on danbooru range work well (even to the point of slight overfitting), and jacknife style was one of them so i didn't thought too much of rebaking for anima.
@RicemanT Yes, Jackknife's performance in v1 is already impressive (judged by facial features and color tone), but relying on the default model seems to have a smoother texture and cannot borrow design features. Moreover, it cannot guarantee the intensity of participation during mixing (compared to using Lora's intensity). If you have limited computing power and don't mind keeping the dataset with tsg, I can also train one later . I actually can't bear to let my GPU sleep at night.
@RicemanT By the way... it seems that advanced algorithms such as Locon and loha are not very compatible with models trained on a large scale. Jacknife_Locon'eps100 can work normally on chenkinNoob020, but it has completely lost control on 050 (unreleased)
@ibara0608 hi, just want to notified that jacknife was my first lora and dataset, which means it was incredibly badly maintained and captioned, so right now I'm revising it. And Anima Preview 3 is also rumored to be released pretty soon, so I'll be waiting on that to release and then start training.
@RicemanT I have conducted some tests on V3. When jacknife is used alone, the low-saturation solid colors of its art style can also be reflected, rather than only the similarity in facial features and hairstyles as in previous versions.
Similarly, I have distilled the LoRA for shiej007, as previous versions failed to capture even his rough brushstrokes, though this has been greatly improved in V3. However, the intensity of his lighting, shadows and their variations is not as strong when controlled by the LoRA.
(Yoshi is actually also on my to-do list, though it may be quite far down and I'm very happy to find an improved version when I wake up.:>)