SDXL-Deephentai-mix

It's best to mix sentences and words.

It's great if you have FreeU, but it doesn't matter if you don't.

Versions 1 ~ 3 are merge models, and starting with version 4, training models.

Note that if you were able to get away with something like "???, ????, from above" before, it's better to be more detailed like "???, ????, from above, ???????, ?????, ?????, on back", negative"more than 2girl, more than 2boy" now.

And if possible, write it in sentences, which is a hallmark of SDXL.

If you look at the 2nd and 3rd photos I uploaded,

Even though I wrote the prompt as a sentence without any word tags, the

You can see that we have a good depiction of the situation.

These examples are images generated by txt2img without any modifications.

version. 1

hentaiMixXLRoadTo_v21+_hassakuXLSfwNsfw_alphaV09-sdXL_v10_x0.5.fp16+ssd-1b-emerge-lora-052000

Note that I'm a beginner in this area, so I don't think I can do much more.

For negatives, I recommend simply "bad ???????, worst quality, low quality, bad quality, sketch, ugly, negativeXL_D, unaestheticXL_Jug6", although these are not proven optimizations.

version. 2

I've included both 0.9 and 1.0 versions in the minus because the artwork has a strong sketchy feel, and to further remove the style of the existing SDXL photorealistic model.

"(hentaiMixXLRoadTo_v21+_hassakuXLSfwNsfw_alphaV09-sdXL_v0.9_x0.5.fp16)

+

Deephentai-mix1+_kohakuXLBeta_beta7Pro-sdXL_v10_x0.2 + ssd-1b-emerge-lora-052000"

For negatives, we recommend "(bad ???????:1.1), worst quality, low quality, bad quality, sketch, ugly, negativeXL_D, unaestheticXL_Jug6".

version. 2-softer

Test results. I think the normal version is better for the general ones, and the softer version is better for the nsfw.

version. 3

I feel like body is less twisted.

However, it is better to put the hand and hands tags separately.

Negative recommends "(bad ???????:1.15), worst quality, low quality, bad quality, sketch, ugly, (negativeXL_D, unaestheticXL_Jug6:1.05)"

version. 4 test

learning more tags related to NSFW.

The version currently uploaded is a test version.

List of learned tags

https://drive.google.com/file/d/131lxFopvoGyFI5DrJA5EaGc9U20G8Ama/view?usp=sharing

version. 4

It's not perfect yet, but I've definitely seen ???? tags corrected.

So I trained with the same data but increased the training resolution.

Planning for version.5

I expect this goal to take some time.

The goal is to increase tag diversity.