still no creation_service -> write a ticket to civitai pls :*
sry for spamming, new one is different and whoooo
THX ALL 15000 that downlaods all my REAL-Pony versions :)
-> here, seems my models not that bad (thx to Lord_Varos)
https://civarchive.com/articles/5942/lets-test-some-pony-checkpoints
for individual faces read my article "faceswap"
(all models are merges at least of 6 models, thx to jugger, leosam(still deleted), Pony Diffusion, GODDESS of Realism and some more)
newest not allways best for you (oc each has its weaknesses but try), most of my pics are without a lora and adetailer or hiresfix, most other models are blending with those stuff
-> choose yourself ... higher version more real but less pony_style (but hardcore)
Try my fully real NSFW-SDXL model
https://civarchive.com/models/270312/sevenof9photoartfantasy?modelVersionId=304720
to get back to cartoon/anime use:
https://civarchive.com/models/395523?modelVersionId=675712
with weight ~0.8 and (anime, cartoon, manhua in positiv | 3D, photo in negativ)
Versions: (suggestion, BEST, smart_evo, allyouwant and keephandsoff)
THEBEST - Base is my Allyouwant with merge a bit SDXL and Pony again, big changes are body-size (eg young girl in a kitchen vs woman), hands, faces ... a bit anime inside (force it) and all XXX stuff -> forget all other PONY-REAL models ;)
Smart_Evolution_contrast(last) - bit more contrast and more details, more sharp background, CFG 10 and right prompt is anime_style, try "instragram selfie" many other pony dont work with that ;) ->Schedule type SGM Uniform
Smart_Evolution - (CFG up to 10, Clip-2), more curvy, more naughty, handle complex prompts, resolution up to 1200pix, hardocre, for anime use NoctFlatStyle
nextgen - (low CFG, if high than steps up to 100, Clip-2), sharp, cozy, different than last, handle complex prompts, resolution up to 1400pix, hardcore back, no anime style
iTgoesON - (CFG 3-10, steps up to 60, Clip-2), sharp, contrast, call background, play with it, new base, bit less hardcore, resolution up to 1400pix
Allyouwant - (CFG 2-5, steps up to 60, Clip-2), bit brighter and more complex environment, not better in some cases. (Hardcore still insde), no_score! , most easy and lora compatible
for anime/cartoon: ad in positiv "manhua, anime" and in neg "realism, realistic, photo"
Keep_Hands_off - (CFG 2-5, steps up to 60), NO Pony-style "try loras", but all ALL is such as real like a Photo ... OMgush ^^ (Hardcore still insde), no_score!
for anime/cartoon: use https://civarchive.com/models/395523?modelVersionId=479629 and in negPrompt zPDXLrl style OR (without any lora)
ad in positiv "manhua, anime" and in neg "realism, realistic, photo"
Real_add - (CFG 3-7, steps up to 60), added more Pony-Sytle and realistic athmosphere (sorry for the fast updates, but i like the new), only change the words anime and photograph from pos to neg prompt to change style, sometimes better cartoon_style without score, sometime zPDXL work sometimes dont -> version final_face still the realest and more easy futa!
final_faces - (CFG 3-4, steps up to 60), added a face lora, now more than 10 different kind of cultures (african, japanese, brazillian ...), also a bit more neutral light, for close-up portraits you can leave any words to make it look like real, call background
final (still most complex prompts) - it follows prompts as much as you want, rise up steps until 60, CFG is good at 3-4, great complex and detailed, so no more anime - because most is gone, you can try to leave prompts that forces good photos, (you should call backgrounds if you want).
(also hardcore possible, score not absolutely necessary)
UPDATED(7/24) inpaint-model(for A1111) - push up CFG to 7 or more
only for following models:
so negativ and postiv prompt are important! but it is not 100%pony and not 100%realisitic!
Real:
Positive: source_realistic, photograph, (score_9, score_8_up, score_7_up, score_6_up)
Negative: sketch, drawing, cartoon, signature, greyscale, monochrome, anime,
Anime:
Positive: source_anime, (score_9, score_8_up, score_7_up, score_6_up)
Negative: photograph, realistic, 3d, render
i suggest an ebedding: zPDXLrl https://civarchive.com/models/332646?modelVersionId=372656 and you can leave that score stuff
at least 40 steps (CFG 3 to 7)
V88b - better background, more listen, more interesting, use my LORA-CFG-Slider to get more real images on very low CFG use it with (-0.4) https://civarchive.com/models/283589?modelVersionId=319151 ; maybe i merge it in next - "last" version ;)
V77j - mostly better hands and better small faces - background more consistent and interesting, only simple nipples -> try ((dark nipples, (puffy nipples:1.2) ,(big areolas:1.2)):0.8) (lactation:1) ; but sorry a bit blurry all in all, bit pony
V9 - bit more easy to provide NSFW, very sharp, less good small faces, also pony style
be aware that pony is a strong NSFW model!
(i only say: horse, futanari, gigantic, (size difference:1.4), (gigantic penis), egg, teats, knot penis, canine ...) you can also create good backgrounds and landscapes
So be private with hardcore stuff !!! i will delete all animal and cp !
Description
newbase, much more background, more interesting,
bit grainy, small part of pony(style) is left, but works in all NSFW, i suggest zPDXL for better images
->
https://civitai.com/models/332646/pony-pdxl-negative-embeddings
and a different CFG Slider
->
https://civitai.com/models/283589/cfgslider
Lora weight ~ -0.4 (CFG~2)
FAQ
Comments (22)
Im getting a washed out picture. Check out the pic below. My VAE is on automatic and not forced, and when I dont use the score tag at all my picture is completely fried.
I've found good success using SDLX.vae. give it a shot
maybe its your lora you use in addition or your negativ-prompt.
sometimes it is a bit pale, so you must force anime - or real, like in my post ;)
and to weight somewhat with 2 is unusual with real-models "(1990s supermodel:2), (wearing barbarian clothing:2)"
MAKE SURE TO USE CLIP SKIP 2
and please more than 30-40steps
there's a new category filter called pony. this is a pony model. you should recategorize :)
Also getting awful looking results using SDXL vae.
Only way to get anything recognisable is to crank the CFG and sampling right up, but the results are mediocre at best. Certainly not 'real'.
It's like getting blood from a stone.
For anyone that is having problems.
Positive: realistic photo, (score_9, score_8_up, score_7_up, score_6_up)
Negative: sketch, drawing, cartoon, signature, teeth, greyscale, monochrome, anime,
Sampler: dpmpp_3m_sde
CFG: 8 (this can be adjusted)
Steps: at least 40
I also don't recommend going any lower that 768x768 on any Pony model.
After the first generation is produced, run it through a Latent Upscale with the same settings with an upscale factor of 2x and 2 steps: followed by a Face detailer
Optional:
1) After your latent upscale, run it through UltimateSDupscale @ x1 with a realistic 1.5/XL model and a very low de-noise (<0.10), then proceed to run it through the face detailer.
2) Cold Oil Style PonyXL Lora @ 0.5 model weight and 0.5 clip weight.
3) You can run the first generation through a controlnet, and have a higher de-noise for the latent upscale for more refined details.
Yes this process isn't as easy as typing a prompt and having the AI do it's magic, but its the closest you're going to get trying to do realism with Pony at the moment. Works with NSFW, but I don't post NSFW here so no examples for that category unfortunately.
All of the images I've posted for this model were just quick generations with intentional film grain/blur on them.
thx for long explanation...
maybe my all in all model v1 is for you ;)
@sevenof9247 This looks like it could be nice! I'll probably give it a try if I ever get tired of Pony. I'm obsessed with Pony lol.
Don't forget you can use a refiner as well. There's an aniemagine checkpoint that works well with pony models.
what face detailer does everyone use for Comfy? The facedetailer that i found seems to slap a face onto a model that isn't at all congruent with the original image. If they are frowning or eyes are closed or wearing makeup etc it just paints a completely normal and unmatched face.
@Up_punchington in automatic1111(adetailer) you can write an own prompt for this part
Your post is more informative than several guides I`ve seen, thank you. I`m looking forward to know your content.
Please add more examples in your galleries. 6.46 GB is a sizeable amount of data and I would like have more idea what I am getting into before hours of download (slow internet). Thanks for your hard work though!
you can doo a lot stuff whty oyu whana see ?
@sevenof9247 the examples are fine, I think 99% of people here have a solid concrete idea of model capabilities. Your earlier model version issues have seemed to dissipate too.
"More anime or more realism", unfortunately, if you mix up the two, your model no longer produce quality realism.
This model name is deceiving, you want this model to do it all, but it ends up doing nothing. Make up your mind of what you want, realism or anime, don't mix the two up.
Otherwise rename this a semirealism model.
no one knows exactly what is coming out, all these mixes are 1month old ...
search for "pony real" you can write all users your oppinion :D
if you don't know how to prompt properly, you should keep quiet, otherwise you'll know what to do to get the realism properly.
Can you give a bit more details of the 77j version compared to v9, is it more anime, more realistic, stronger NSFW tag following?
mostly better hands and small faces - background more consistent and interesting


