The paper of LyCORIS project is accepted by ICLR 2024!!!!!!!!
"Navigating Text-To-Image Customization:From LyCORIS Fine-Tuning to Model Evaluation"
My first ever paper, as an undergraduate student.
Arxiv here:
The one of best SDXL anime base model: Kohaku-XL delta is released!!!
It is trained on 3.6M images with LoKr algorithm. And it can generate various style and characters
I captioned 7.45M danbooru images with Florence-2 (from msft)
currently only "<MORE_DETAILED_CAPTION>" has been done and "<DETAILED_CAPTION>" is running.
"<OD>" and other targets are planned but low priority.
So the cool thing is here
Basically a special use case of PaintsUndo
I will use this to make sketch-lineart-image pair for reproducing some kind of "copainter"
DanTagGen-beta is released!!
Pretrained on 5.3M entries with 10epoch.
Same 400m param arch
Better performance!!!
Check the Hugginface repo or HF space:
Thx
@Gradio
for zero-GPU resources UwU
LyCORIS 3.0.0 is finally released
over 200 commits
brand new module/functional/parametrize API, more efficient implementations, better testing and brand new banner
The most unsafe part of AI is not the AI themselves, but their users, and "bad Anit-AI tools"
The orig img is from "human artist".
With some "not generative model" process.
I can fake it as "AI generated image" easily.
Will this tool protect "human artists"? or attack them?
Contrast-based downscale is now implemented into Python UwU.
Better color choice when downscale (for pixelart-like style).
Also I add some example of using my outline expansion method to improve other existed Pixelization Method. (Even NN-based method can get improvement)
Just published a metadata database for Danbooru. (Use the official dumps of posts metadata, not by crawlers)
You can now use this database for tagging/analysis or whatever you want.
Model card also provide a link to the dataset which have all the img.
Just found I already spend 5000 3090hour on kohaku xl and DTG project
With my own GPU only.
(All the broken run and dataset processing are not counted, only training time for published model)
I think this is quite impressive
FLUX Schnell with TITPOP-200M-dev (the NL+tag prompt gen I'm developing )
I already released the weight but I will not release the introduction until I finish all the things.
(I may provide a close source space for who is interesting to use it)
Next version of kohaku xl is training
Not much difference for well learned concept or style, but have improved detail, feeling, composition and have more data and newer data
Also, natural language caption are used, not only tags.
it's so fucking shit when ppl using the result from 128 x H100 to say "Lora-like have worse detail"
ok, than show me how to use 4x3090 to do native finetune and get same result as 128 x H100?
or give me 128xH100 and I will show you "lora-like" can get almostly same result
The one of best SDXL anime base model: Kohaku-XL delta is released!!!
It is trained on 3.6M images with LoKr algorithm. And it can generate various style and characters