Shidoshi Mix

Why hello again!

Edit: Shidoshi mix 6.0

I've been away for a while and noticed a lot of people were having issues with washed out and grey haze on v5.0. I noticed it too so try this one and let me know how it works! I think it's better? I dunno... I've burned the VAE in on this one.

Cheers!

-Shidoshi

Hi!

Edit: Shidoshi5.0

Same info as below. This time i built up from version 3.5 and got rid of the hentai factor. Back to realism. People seemed to like 3.5 the most. As with before, this one has no VAE burned in.

This model can produce photorealistic pictures with a basic prompt and no negatives. See my examples. I use it with low "quality" tag strengths (i.e "worst quality, low quality:1.2" and avoid the term "masterpiece" in the positive. "Masterpiece" seems to make it not listen to your prompt for some reason. I dunno. Try it out.

Post your samples (I like it) and give me feedback! Thanks.

Edit: Shidoshi4.1-pruned - My latest mix with more hentai infusion! Blended some abyssorangemix2 and some hassakuHentaiModel_v12 in. It's still kind of realistic, but breasts tend to be massive now. As with my previous mixes, generic terms like "Woman" tend to lean towards an Asian look.

It also does nudes, genitals, sex, and cum very well and also works pretty good with Loras.

Always use the 'hires fix' or 'second pass', otherwise it looks bad.

NO VAE or LORAS BURNED IN. You need--> https://civitai.com/models/76118/vae-ft-mse-840000-ema-pruned

Oh yeah, Face Restore sucks. Just use ADetailer if you need it.

Recomended negs: (worst quality, low quality, normal quality:2), cartoon, painting, illustration,

Recomended pos prompt (I usually dont use them): highres, extremely detailed, 8k photo, best quality, intricate details, realistic, detailed face and eyes,

Samplers: Any of the DPM++. I prefer the DPM++2M SDE Karras on the second pass. Need to apply more steps though. Euler A looks kinda crappy on this one.

Plus if a face/teeth/hands look like ass, use ADetailer.

Also use the add_detail Lora if you need. It doesnt always need it.

Oh yeah, Face Restore sucks. Just use ADetailer if you need it.

Edit: v3.5 is one of my newer mixes. There are no burned in Loras or burned in VAE. See below for the recommended VAE. Only pruned models are going to uploaded unless someone wants the full float version. Nobody seems to care about the full float models.

Hi! I was aiming for realism with this model mix. Works well with both "Mature Themes" and "Not Mature Themes". Someone asked if I could share. So I did.

Comparative Grids found in V1 previews. No embeddings were used in the samples... So extra fingers and such maybe be present.

Added pruned (FP16 and no ema versions as requested).

My favourite is the base 2.0 and the bake-in that's not ++.

______________

Note:

----VAE NOT BAKED IN, as i will sometime switch it up to play around---- vae-ft-mse-840000-ema-pruned is used most often. https://civitai.com/models/76118/vae-ft-mse-840000-ema-pruned

______________

There are 4 Versions here. The original Shidoshi Mix and Shidoshi Mix 2. Shidoshi Mix was my original "final". It almost has a flash photography like effect and brighter contrast. Shidoshi Mix 2 was mixing a few more models in and ended up with a little darker feel to it. The other Shidoshi Mix 2's are with baked in loras for detail. I used Detail Tweaker LoRA (细节调整LoRA) https://civitai.com/models/58390/detail-tweaker-lora-lora and Hipoly 3D Model LoRA https://civitai.com/models/8730/hipoly-3d-model-lora for the "bake-ins". The "++" model is with the Loras ratios at almost at 1. Looks darker and the eyes look kind of small and funny at times and look as if they got salt water in their eyes. Hah.

The base "Woman" and "Man" look like a mish-mash of every race on the planet leaning towards an Asian look. I find that using something like "Irish Woman" or "English Man" will give white people output. Also using "Asian" in the neg along with "Style Asian Less" embedding created by the great Zovya https://civitai.com/user/Zovya/models , also helps. This model seems to excel at generating Asian people. These models also love generating nudes.

Model Info:

These are a mix of several models I've merged (using "trained difference" and the new elemental merge adjustments) over the last couple of months or so.

The models used:

-A-Zovya Photoreal (best model and embeddings!) https://civitai.com/user/Zovya/models

-omnidiffusion - ?? Vanished ??

-chillout mix https://civitai.com/models/6424?modelVersionId=11745

-uberrealistic porn merge https://civitai.com/models/2661?modelVersionId=15640

-Pornmaster https://civitai.com/models/80473?modelVersionId=107794

-And more... I just can't remember which ones I used and I didn't save the metadata into most of the mixed models as I had no intention of sharing.

Personal opinion crap:

Best res: 768x576 (4:3 ratio looks nice) and it can do 1024x768 quite well too. Always use the Hires Fix though (0.2-0.4 denoise?).

I use the DPM++ 2M SDE Karras sampler with higher steps (35-50). A personal preference.

CFG scale: 6-9.5... Gets a little over saturated at 10+. Dynamic thresholding works well though.

If it wont listen to your prompts, flip the Clip Skip to 2 and try it out.

Most Loras, embeddings, and LyCORIS I've tried work well. Can't comment on hypernetworks, as I don't use them.

Irises and pupils get funky at times. Use ADetailer for correction. Usually the Hires Fixes it.

Samples uploaded are without negative embeddings.

But always use the:

NEG PROMPT: (worst quality:1.8), (low quality:1.8), (normal quality:1.8), else it looks like poop.

I usually use the "Add_detail" Lora for the extra kick of realism.

Enjoy.

Please don't use this for commercial use. I didn't train the original checkpoints and therefore I shouldn't have a say on what can be done with this model.

_________________________________

Important note: Love, support, and like the ORIGINAL (OG) checkpoint trainers that make this possible. People like me just remix the originals. It's kinda like the people who put out the crap remixed music coming out these days, mumbles some crap vocals over top, and then takes credit for being some wonderful artist . If it wasnt for the original checkpoint artists, we wouldn't have remixes.

_________________________________

And no, I did not make an inpainting version.