Style of the artist: muchin
Update to v2!
V2 is an updated, rebuilt dataset.
99% of the dataset has been reworked, clarity has been improved, cropping is less aggressive, and it's updated. Artifacts from censoring have been almost completely removed, and the actual censorship in the dataset is no more than a few percent.
The goals remain the same: reducing the impact of blurry backgrounds (theoretically), reducing the impact of tanned skin, and stabilizing the color range and overexposure. The stylistic approach remains the same: removing all extraneous characters, including symbols, and excluding some motion lines.
The style is more modern, an attempt to streamline the old style (so I know which of the several approaches to style is the right one, lol). The approach with three quality categorizations remains, focusing on subjectively high-quality work. The lower the clarity, relevance, or quality, as well as subjective aspects, the lower the category. And the impact of each work in that category.
The training was conducted at 1536. The model has excellent detail, although there are some minor stability issues.
It's quite possible that some improvements will be made, especially in terms of signatures, but there will be no further fundamental alterations to the dataset.
All generations in the examples were done at 1536 x 1536 resolution, which is recommended for better detail.
Trigger word: @0tx1
Recommended weight: 0.7-0.9 (0.8)
Recommended words:
veiny belly
There are issues with the background. it is generally fine, but by default it's blurry background and photo background. Use these words in the negative. Also, if your skin seems too tan, try adding tan to the negative. The default negative that I would recommend use is:
worst quality, low quality, score_1, score_2, score_3, artist name, logo, adversarial noise, jpeg artifacts, censored, blurry background, tan, photo background
I will soon release new sdxl ver.
I don't know what's wrong with TensorArt, previous versions are in some shadowban or some kind of error, you might remember me from these models: