This is academical model which ws trained with LoKr-method, for replacement of full finetuning. It was trained with 8x3090 for 1.5 months, with frequent pause-and-resume.
The model is capable to generate higher-resolution images, unlike other models, up to 1536x1536 resolution.
You can test img2img directly for example, to examine whether model 'knows' and 'detects' what to denoise and what to make correctly.
Starting from this image, using prompt masterpiece to see the effect-
You can see this quite interesting result, showcasing what model tried during img2img process. Basically, it has to 'detect' then 'denoise', so it means better detection capability is in this model.
When model gets better, it gets more ability to do better img2img / corruption handling.
Here is the result with the txt2img with prompt:
1girl,[artist:WANKE|artist:free_style |[artist:ningen_mame]|ciloranko],tokoyami towa,devil,sensitive,dark theme,glowing eyes,silhouette,sword,
Or, higer resolution native generation without any hi-res fixes ehre.
What would be next?
This was purely academical model, which is not intended to be related with any commercial projects, (including myself!).
However, the stronger model is now being prepared for release.
I believe sustainable training is required for everyone's future, for open source communities too.
Please wait for our announcement very soon <3