I just added a TCD model. This is a model created to produce quality pictures close to v2.0 but with only 9 steps, I chose TCD instead of lightning because it is a more realistic model.

I wrote the wrong model name in the sample photo when I left it as v1.0

Some notes when using the TCD model:

Sampling method: Euler a (You can try other sampling methods, but most of them are unusable)

Sampling steps: 9 (If it is lower, it will lack important details, but if it is higher, redundant details will appear)

CFG Scale: 1,5 (Lower will not have enough accuracy or sharpness or vivid colors in the image, and anything higher will cause spots and color errors to appear)

Negative prompt: (This doesn't have to be the same as mine, perhaps yours might be even better.)

(worst quality:1.5), (low quality:1.5), (normal quality:1.5), lowres, bad anatomy, bad hands, multiple eyebrow, (cropped), extra limb, missing limbs, deformed hands, long neck, long body, (bad hands), signature, username, artist name, conjoined fingers, deformed fingers, ugly eyes, imperfect eyes, skewed eyes, unnatural face, unnatural body, error, painting by bad-artist

Please note, the sample images above are indeed from the model, but they have been used with specific settings, and added extensions on automatic1111 such as hires fix, freeU, ADetailer. If you want to get the closest results, please stick to the information obtained from the sample photos, or follow the instructions at the bottom of the article.


This second version has made some minor adjustments to the NSFW 3D-style data, the most significant changes in this version lie in the data of different art styles, along with the textual dataset within the model. I want it to become more comprehensive and expansive, but it will still focus on NSFW 3D unreal. When compared, you will see it produces images with higher flexibility, more diversity, smoother body lines and facial contours, more user-friendly with some manga and anime keywords, along with other art styles, the hands have also improved compared to v1.0. However, please note that although there is a considerable amount of data, it is still a fine-tuning model, so there will be a lot of data not yet added. If you enter a prompt that falls into the results not included in the data I trained into the model, it will revert to the original data of the base model and the results might not look good.

I created a youtube channel specializing in posting songs made from AI. If anyone is interested, you can try visiting:

https://youtube.com/@duchaiten?si=o3GZZ3aODRafPKUH

If you like my content, support me at:

http://paypal.me/duchaiten

Patreon in here:
https://linktr.ee/Duc_Haiten

If you want to use my models to create unlimited images at a small cost without having to install anything, mage.space will be a good choice, i have some models that are exclusively available on there.

The priority settings for this model are:

Sampling method - DPM2 Karras.

Leave ADetailer at default and only adjust the inpaint width and inpaint height to 1024 since you are using the sdxl model.

For Hires.fix, use nmkdSiaxCX_200k or ESRGAN_4x, it will result in blurrier outcomes compared to other hires fix models but it helps retain small details such as skin, hair strands, water droplets, and muscle lines...etc without distortion.

You could also use the hires fix model SwinIR 4x if you want it sharper, but some small details will be distorted and hard to preserve. Also note that the Denoising strength of hires fix should not exceed 0.2, otherwise, some errors in the images may occur.

As for FreeU, I have switched it to be suitable for SDXL and kept the default parameters.

Negative prompt:

(worst quality:1.5), (low quality:1.5), (normal quality:1.5), lowres, bad anatomy, bad hands, multiple eyebrow, (cropped), extra limb, missing limbs, deformed hands, long neck, long body, (bad hands), signature, username, artist name, conjoined fingers, deformed fingers, ugly eyes, imperfect eyes, skewed eyes, unnatural face, unnatural body, error, painting by bad-artist