Tasai | a modern fantasy art checkpoint

This checkpoint, just like it's little brother (lora) can do both normal art and sketches fairly well. The biggest improvement from the lora is that this checkpoint is trained on a lot more data from more artists (with similar artstyles) and as a result it has better quality and prompt adherence.

Please note that a big chunk of the dataset used for training includes humanoids with colored skin (blue, red, green, grey etc), also a lot of the included artists simply use unnatural skin tones for humans (kind of pale or grey-ish). So while the model is pretty good at doing colored skin gens, it also means that it doesn't really generate "normal" looking skin most of the time. This is not really an issue for me, rather an art style and I am not planning on "fixing" it since that would fundamentally change the art style.

What I do plan on improving though with the next version is watermarks, for this version I didn't really remove any watermarks, since the dataset is pretty big and it would take a lot of time and effort, if people end up liking this model I will work on removing all watermarks and re-training. For now as long as you have the "watermark, signature, web_address, patreon_username" trigger words in the negative prompt you should be fine 80% of the time.

Recommendations:

cfg: 7

scheduler: Karras (but anything works honestly)

sample: dpmpp_sde (pretty much anything works like above)

steps: 25-40 (I use 40 most of the time but 30 performs pretty much the same)

If you're gonna use the same setup in comfyui, make sure to use the custom sampling nodes because karras/dpmpp combo produces horrible results with the ksampler node.

negative prompt:

bad quality, worst quality, worst detail, sketch, censored, watermark, signature, web_address, patreon_username, extra_digits, extra_fingers

positive prompt:

masterpiece, best quality, amazing quality, absurdres,

ofc if you like different settings, prompts etc that is fine, these are just my general recommendations.

Using anything other than danbooru prompting will give you bad results though, the model is trained on danbooru prompts!