Attention: You need to get your own VAE to use this model to the fullest. While it does work without a VAE, it works much better with one. I recommend you try this one out.
Hey everyone. After GrapeLikeDreamFruit hit, I started missing having a more general purpose model for the mundane kind of pictures - nude photographs on different backgrounds and some light hardcore capabilities. This model here is my response to that need. It handles the female nude superbly, and while it's less of an artistic model than GrapeLike, it's still quite capable in that regard. It's quite good at hardcore, even if that is just a secondary goal for this model, and can be prompted for a variety of acts. Model has a good response to Danbooru tags.
This model involves dreamlike photoreal, so here is the license that you must abide by.
See examples for some ideas.
Deeper explanation:
This model merge was done following a similar phylosophy to GrapeLike: a Block merge between a realistic anatomy + skin core and a posing anime core. The block merge is done in such a way as to emphasize the anime core in the center of the Unet, while rapidly decaying back to the realistic core at the edges. This has the effect of bringing a lot of posing and composition ideas from hentai models inside the photorealistic core we have available without touching textures and photorealism. Full recipe follows:
Anatomy core:
izumi, F222, dreamlike photoreal, realistic vision 1.2, sxd, all at the same intensity. ie, the merge chain was:
(((izumi + F222 0.5) + dreamlike photoreal 0.33) + realistic vision 0.25) + sxd 0.2
Anime core:
Anything v4.5 merged with Basil Mix using the same block merge coefficients used in mixing Abyss Orange Mix, then merged 40% with grapefruit, the result was merged 30% with gape60 and finally 15% with RPG v4.
Bringing both together:
The block merge for both was done using a formula. I kept the bottleneck Model A was anatomy, model B was anime. I kept the center lalyer at 0.7, as well as base alpha, then followed 0.8/(n**1.1)as a merge rule, with n being distance from the center. Full numbers were "0.05199847612695355,0.05722134125067434,0.06354625877794251,0.07135480548979826,0.08122523963562354,0.09407671474206218,0.11146117361039158,0.13621438760332552,0.17411011265922482,0.23892225595753655,0.373213196614723,0.8,0.7,0.8,0.373213196614723,0.23892225595753655,0.17411011265922482,0.13621438760332552,0.11146117361039158,0.09407671474206218,0.08122523963562354,0.07135480548979826,0.06354625877794251,0.05722134125067434,0.05199847612695355".
Description
This version includes a completely redone anime posing core. Exact details about how to reproduce that merge can be found in the page for that version.
Things that are different:
Model no longer trends towards asian women. This happens because Basil Mix is no longer a part of the mix.
Breasts are smaller by default. You can still prompt for
gigantic hanging titsif you want to, and you will get them.Noise offset is baked into the model for low key lighting. Sources were iCoMix and Cine Diffusion.
Things that are better:
Model is more consistent and more responsive to prompts. It's also less biased towards NSFW (while still biased towards it). Unprompted penetration is a thing of the past.
More capable in fantasy and horror situations, better backgrounds.
Anatomy, especially genitals, is much improved. Hardcore capabilities are also much better.
Easier posing, better emotional response from subjects. Also more varied faces, though not improved by much.
Things that are worse:
More difficult to get standard people. The first version was very good in getting non-super-model women. It now trends a bit more towards Hollywood pretty. Regular looking women require prompting.
NSFW stuff and nudity are not necessarily default in this version, so if you want someone nude, you will likely have to prompt for it.
Feet. I'm sorry to all feet people, but you probably want to prompt for socks / shoes.
Merge Details:
Obtained by fusing the photoreal anatomy core (unmodified) with the anime core using the Merge Block Weighted extension for Auto1111 with the following weights:
"0.05199847612695355,0.05722134125067434,0.06354625877794251,0.07135480548979826,0.08122523963562354,0.09407671474206218,0.15,0.25,0.4,0.6,0.6,0.8,1.0,0.8,0.6,0.6,0.4,0.25,0.15,0.09407671474206218,0.08122523963562354,0.07135480548979826,0.06354625877794251,0.05722134125067434,0.05199847612695355"
base alpha set to 1.0
Full list of models used:
Anatomy:
Izumi
F222
Dreamlike Photoreal
Realistic Vision 1.2
SXD
Anime:
cine diffusion (yes, it's in this block)
iCoMix v3
Counterfeit v2.5
gape60
As always, there is a special license attached to dreamlike derivatives. You can find a link to that in the model card.