sexual situations

physical violence

disturbing

male nudity

hanging

hate symbols

nazi party

revealing clothes

weapon violence

female swimwear or underwear

male swimwear or underwear

partial nudity

white supremacy

adult toys

graphic male nudity

illustrated explicit nudity

nudity

graphic violence or gore

graphic female nudity

pg-13

corpses

wide hips

convenient censoring

peeing

oral invitation

emaciated bodies

exposed female nipple

blowjob

female nudity

sexual activity

sexual intent

undressed

male underwear

female swimwear

genitals

female underwear

thick thighs

breasts out

strapless leotard

vore

breast out

one breast out

huge breasts

gigantic breasts

huge butt

covered nipples

hair over breasts

no panties

sitting on face

anal

dildo riding

downblouse

oral

porn

futanari

hentai

nude

lingerie

sexy

nsfw

suggestive

child on child

self injury

extremist

hate speech

diapers

urine

incest

scat

latex clothing

swimwear

bukkake

fellatio

cumshot

implied fellatio

eat_cum

cumdrip

cum in pussy

cum on face

after fellatio

cum on hair

cum on body

cum on tongue

cum on hands

cum in mouth

triple fellatio

autofellatio

fucked silly

cum on pussy

pov fellatio

Ponydiffusion is an excellent model for 2d content, but it seems rather inconsistent with 3d. This model is designed to more consistently produce photorealistic 3d images of a variety of subjects. Currently, the beta version still produces a more CGI effect as I do not believe I have enough sample images, but hopefully future versions will be more realistic. I would recommend checking the description of each version to see what it does and what its drawbacks are for the time being for more detailed info.

So after looking over my model and the training data of some other LORAs, I determined that the issue causing the overblown look was that I had set the batch size for my training far too high as well as over trained the model. I also removed some images from my dataset that contributed to the overexposure issue and added some replacements. I am proud to say that this version produces much better looking faces as well as overall more realistic images. However, as the example images show, some character or expressions may mess with the realism factor, leading to some funny looking half-anime, half realistic results. I have tested the model on multiple different samplers at different CFG levels and DPM++ 2M SDE Karras at 16 steps with a CFG of four and Euler Ancestral at 30 steps with a CFG of around 7 seem to work best. For full body shots you may still need to use ADetailer to fix face, as if often messes up the eyes depending on the character. To trigger the model, simply use either "source_photo" or "source photo" and then your prompt, like you would with any other pony source. The model seems to work at a strength of around 1.0, but you can lower it a bit if you want to use it with other loras. Ive also tested it on multiple checkpoints, from base pony to more realistic ones, and it gives pretty decent results across all of them, and is quite capable of generating stuff like animals and vehicles.

Release V3

While I am calling this release 2.1, theres a fairly significant number of changes to it. For one, ive shortened the trigger tag to "source_photo" for ease of use (though I could probably just drop it to just "photo" in the future), and used Booru tags instead of sentence captions for the images. Its still capapable of processing sentence style prompts, obviously, though it does seem to like booru tags when specifying backgrounds. This version also doesnt need an excess of negatives like the prior versions, and seems to function best with a CFG scale at 4-5 and the DPM++ 2M Karras sampler at around 15-17 steps. After some additonal testing, I have determined that a CFG around 7 with a Euler A Automatic sampler at 30 steps also yields very good results, though it takes a bit longer on lower end hardware. Depending on the character, you may need to add more negative and positive tags to enforce realism and discourage it from doing a 2d or CGI effect. I also recommend putting "plastic,plastic skin,overexposure,blurry" in the negatives for most prompts unless you like the shiny skin effect that most realistic AI models seem to give people.

Release V2.1

So after much testing, I have finally produced what I can confidentally say is a relatively stable version with realisism exceeding that of V1.0 using civitai instead of my potato of Graphics Card. Thanks to Ai_photos for providing his training parameters for more realistic character generation, as it was also helpful for improving this model. This version doesnt require as many negative keywords or sources, but negative embeddings certainly still help with the quality.

Release V2

Experimental release version trained at 512X512 steps. Still requires a lot of additonal keywords in both positive and negatives to produce realistic results, but is decently consistent in making a realistic style. Also doesnt seem to suffer from the overexposure issue of previous versions.

Beta Release  V2.0

Renamed to beta due to sometimes producing lackluster results (at least in terms of convincing realism compared to v1.0. Trained on 75 images at a 1024X1024 via civitai resolution this time. Still trends towards the realism of the previous model but does seem to have the shiny skin issue the previous beta versions had, which makes me wonder if its due to using a higher batch size. Using negatives does help with the issue to some extent.

Beta Release V1.5

I am happy enough with the results of this version to call it a release version, though it does have some kinks in it still. I found that reducing the number of source images from about 200 to 50 seems to have massively increased the quality without reducing flexibility any. Since it was only trained at 256X256 it might have some blurryness issues, but using an upscaler should help with that. It also seems to occassionally produce an oil painting style, so I recommend using negatives such as "2d,cgi,source_pony,source_anime,source_cartoon,source_western,source_furry,shiny,plastic,glossy,2d,cgi,traditional media" to prevent it from doing 2d or cgi style art as much. Just the keyword "source_photorealistic" should work in positive prompts, though adding "photrealistic,photo" and similar triggers can also help.

Release v1.0

Photorealistic Pony XL

Type	LoRA
Stats	373 2,179 23
Reviews	Positive (27)
Published	Apr 16, 2024
Base Model	Pony
Training	Steps: 750 Epochs: 5
Usage Tips	Clip Skip: 1 Strength: 1
Trigger Words	source_photorealistic
Hash	AutoV2 5EB7798E29