Up-and-coming modelers join the fray! Over 300 commercial Stable-Diffusion illustrations!

This model aims for beauty rather than realism. It merges proportions and styles from various models, including custom trained ones, and applies fine-tuning. It can generate NSFW content.

To get the best results, use negative prompts about quality issues and artistic styles. One good prompt is enough, but you can add more like “bad hands” if needed. Positive prompts about quality like “4k” and “masterpiece” can help too. Adding “(NSFW)” tends to make images more beautiful.

Good settings are 1-2 steps clip guidance, DPM++ 2M Karras sampler, 30-50 steps (usually 40), high-res fix enabled, 0.25-0.5 denoising strength (usually 0.32), upscaler 4x_Sourcetex-v2_136000_G.

Image sizes of 512×512 to 768×512 tend to work best. CFG scale around 7-11, and Adetailer can reduce blurriness.

There’s still much to explore with this model. Please experiment and enjoy!

This model is capable of generating both NSFW and SFW images, including stunning landscapes and portraits.

Example prompt:
‘hyper realistic gopro action photo of a beautiful 20yo Dutch girl with small breasts (looking at camera:1.3), windy, wearing old trashy worn torn out clothes, in a themepark, analog style, masterpiece, exposed breasts’

Negative prompt:
‘3d max, grotesque, desaturated’

Special thanks to Hoblin for this negative prompt!

Settings: Euler a, steps between 25 to 35, CFG7

• v5.0 update: Training used less than $30 and about 1000 civitai credits. Maintains 768/832 resolution generation in 9-12 steps with specific samplers. Used deformation, worst quality, oversaturation, etc. for negative training. Merge weight -0.3.
• v4.6 update: Trained on “ugly” datasets and merged with “add differences”. Improved anatomy at 768968 and 968768 resolutions. Over saturation issues.
• v3.7 update: 10x faster than SDXL. Used low lr 4e-7 and adafactor. After 4 epochs, learned from more realistic models. Focused on portrait and landscape generation. Optimal resolution 768*960.

• This introduces an anime model checkpoint that combines favorite checkpoints chosen from CIVITAI.
• All checkpoints use only the “CreativeML Open RAIL-M” license.

This model is the result of multiple merges with various models, mainly hardcore adult content models. It also uses my Lora and other Lora models. Lora works well with this model.

Basic prompts I use for quality improvements (others can be used):

best quality, masterpiece, realistic, hdr, detailed, intricate, (perfect embarrassed face:1.3)

Negative elements:

EasyNegative, by bad-artist, bad-hands-5, (low quality, worst quality, bad anatomy:1.5), (monochrome, grayscale, sketch:1.2),(signature, text, logo mark:1.5), soft focus.

I merged some failed LoRAs and found the result to be pretty decent, so I’m releasing it publicly.

Since it often draws eyes oddly, it’s best to specify details about eyes in prompts to get better results.

“The Night Princess”
For the best infels in Stable Diffusion WebGUI, follow these:

Prompt:
highres, high quality, 1girl, uehara meiko, small breasts, black choker, pink dress, juliet sleeves, outdoors, sky

Negative prompt:
(worst quality, low quality:1.45), blurry, monochrome, sketch, lineart, emphasis lines, (anime screencap:0.1), (interlocked fingers:1.315555), (mutant hands:1.455555), deformed body, bad anatomy, bad proportions, anatomical nonsense, what, signature

Settings:
–xformers –no-half-vae –opt-sdp-no-mem-attention

More examples can be found on my Pixiv.

This model integrates checkpoints I used for illustration creation. It uses Textual Inversion without LoRA to generate beautiful thin-lined anime-style girls. It combines various LoRAs for anime-style and NSFW illustration creation.

Integrated models: gy动漫010 v1, seizaMix v2, FiaMix++ Dynamic v4.0, YesMix v1.6 (original)

Commercially usable. Follow merge sources’ policies.

Recommended settings:

Sampler: DPM++ 2M Karras

Steps: Around 20

CFG scale: 7 – 8.5

High-res fix: Generally not needed. If used – R-ESRGAN 4x+ Anime6B, 20 steps, 0.4 noise reduction, 1.5x upscale

Recommended Textual Inversion: easynegative, bad-artist, badhandsv4, bad-picture75v, deep negative 1.75t