Attention: You need a VAE to use this model. I recommend you try this one out.
This model involves dreamlike photoreal, so here is the license that you must abide by. Also check out the license in the chillout mix page regarding commercial use, we have chillout in this model.
This is Mars Mix, closely related to Ares Mix, but allowing for a hit in realism in exchange for some wilder poses / artistic possibilities and bigger breasts. The model uses the same photorealistic core as Ares Mix, but uses a rebuilt anime core with a much more aggressive block merge that allows for a lot of the posing and texture information from the anime model to come through, while keeping the photorealistic skin textures and faces. The result is a photorealistic model that allows for a more exaggerated anatomy, understands the booru tag prompting style really well and is compatible with CLIP skip 2, while still mostly keeping the hardcore capabilities of Ares Mix. If you're a member of the anime side of the community and would like to learn the ropes of photorealistic, I consider this model a good place to start.
All example images were made using CLIP skip 1, ENSD 31337. The parameters are included in the images. I recommend you always have child or infant in your negative prompt. Since Analog Diffusion is making an indirect appearance in the model list (by way of Chillout Mix), you may sometimes notice some facial blur or haze. This can be remedied by including blur and/or haze in the negative prompt.
Longer explanation for model merging enthusiasts:
This model uses the same anatomy core as Ares Mix, you can read about how to build it from that model's page. The anime core has been rebuilt using Chillout Mix instead of Basil Mix for the block merge, the complete removal of gape60 and anything v4.5, and updating grapefruit to version 3.1. Chillout is merged directly into grapefruit using the Abyss Orange Mix coefficients, then the result gets 15% merge with RPG v4. This forms the new anime core
The anatomy core is block merged with the anime core using the following Coefficients (anatomy is model A, anime is model B, base alpha set to 1.0):
"0.05199847612695355,0.05722134125067434,0.06354625877794251,0.07135480548979826,0.08122523963562354,0.09407671474206218,0.25,0.4,0.4,0.4,0.4,0.8,0.7,0.8,0.4,0.4,0.4,0.4,0.25,0.09407671474206218,0.08122523963562354,0.07135480548979826,0.06354625877794251,0.05722134125067434,0.05199847612695355".
Description
Replaced grapefruit with Herb-mix.
Sometimes, v0.1 would produce weird anatomy with big hands, big head and miss important parts of the prompt. The only way I could fix it was by trying to minimize the effect of anything v3 on the model, This is one way to do it. New version is more responsive to prompts, has better anatomy and produces better hands.
FAQ
Comments (14)
Why my generations are washed out?
Probably you are missing a VAE. I can't help much more without knowing your setup, but the model description has a link to where you can get a VAE.
@rocp I just used civitai extension on sd webui and downloaded directly.
Same here, it's the VAE, I use the XY plot to assign, otherwise, I think it should be merged into the model, though I have the vae in the folder but it does not apply unless I assign it
@Imanent Really? how do i assign that?
@Imanent you have to select the VAE in settings if you are using auto1111. Settings -> stable diffusion -> forget the name, but something about a VAE. Don't leave in automatic, it doesn't work.
@rocp it doesnt work even if enable it manually
Anyone have some prompts to produce a person in the process of taking off their top?
Try shirtlift Lora
Hey Rocp,
Thanks a lot for sharing your workflow for the model merging. Also, I am not sure I completely understand : your list of alphas is of length 25, also I thought the Unet only have about 11 "main" layers.
Could you please tell me what your list of alphas corresponds to ? Does it takes into account also the text encoder ?
The UNET has 25 blocks, 12 (0 to 11) IN blocks, one bottleneck block, 12 OUT blocks.
The list I posted can be copy-pasted directly into the Merge-block-weighted extension for auto1111 to get the correct weights, but the alphas are in natural order, first the outermost in layer all the way to innermost, then bottleneck, the innermost out layer all the way to outermost.
As for the text encoder, that is addressed in base alpha. For Mars, I used a base alpha of 1, meaning the encoder is taken from the anime core.
Can i inpaint with this model?
You can, yes.
for increased quality, you may want to turn it into an inpainting model. Here's instructions:
https://www.reddit.com/r/StableDiffusion/comments/zyi24j/how_to_turn_any_model_into_an_inpainting_model/