|
--- |
|
license: openrail |
|
--- |
|
SulphMix is my first attempt at creating a Stable Diffusion mix, therefore there might be a couple of problems with hands and other defects, i will hopefully fix those in upcoming updates. |
|
As this is mostly a prototype, i haven't really recorded all my recipes, however it is based on a couple of checkpoints such as: |
|
1. Store Bought Gyoza https://huggingface.co/Jemnite/GyoZanMix |
|
2. Akkaimix https://civitai.com/models/78434 |
|
3. Noosphere |
|
|
|
SulphMix2 incorporates a few other models such as: SummerSolstice, EtherBluMix, Lunar Radiance (night) |
|
SulphMix noosphere mix (again): 0.75(SulphMix2) + 0.25(noosphere_v3): https://civitai.com/models/36538?modelVersionId=107675 |
|
|
|
Recommended settings are as usual 512x768, 768x512 or 512x512, but i have gotten good results even with 1024x512 and 1120x512, with hiresfix and adetailer on. |
|
I mostly use the VAE vae-ft-mse-840000-ema-pruned but it should work with any. |
|
|
|
Samplers: DPM++ 2M Karras Steps:30, DPM++ SDE Karras Steps:30, Restart Steps:36 |
|
Positive prompts I like to use are: detailed background, best quality, masterpiece. But check out which ones you like, most likely the ones you already use work here too! |
|
Negative prompts: It is recommended to use negative embeds such as Easynegative and EasynegativeV2, but see which ones work best for you. |
|
Sometimes, the model will give holes in the ears therefore it is recommended to use "earrings" in negative prompt to get nicer ears. |
|
|
|
Loras i recommend and like very much are: |
|
|
|
GL or gigalighting loras by Segole, found in this repo: https://huggingface.co/segole2003 |
|
Style Buff by Ocean3: https://civitai.com/models/143557 |
|
|
|
|
|
Recipes and changes to the models: |
|
|
|
Work in progress, however for detail, eyes and backgrounds i went with Gyoza. |
|
For a more cuter look i applied AkkaiMix and other similar models i found. |
|
Therefore, this model works best at 1girl+ and not 1boy+, howver you can generate men if you are more precise in your prompt. Tho it can be harder than other models. |
|
|
|
Preview images: |
|
|
|
 |
