GoR_v6ILXL+VAE
SETTINGS:
DPM++ SDE Karras, DPM++ 2S a Uniform
EULER A Karras
clip skip 2
Steps +/-30
GFC+/-7
Well, I committed this version. Take into consideration that trying to make this model realistic may have messed with the base illustrous a bit. The effects may be surprising ;)
SDE can sometimes burn through but for that Euler A is finally useful.
You can also test clipskip 1, but for me it is too “smooth”
PS: And one more thing, if you see that I used your original prompt, please link the post in the comment under the image. Sometimes I'm surfing the civitai gallery and testing the prompts on my models and then just can't figure out what's from where
-GoR_PONY_v5ilustros_H+VAE!
This is experiment, it is Pony but with a liitle of Illustrious, lets say it is beta test
SETTINGS:
DPM++ SDE Exponential
clip skip 2
CFG: 4,5 to 6,5
Use negative prompt not only pony ones
Check my pics for settings
Please upload your work, it will allow me to improve the model. I myself don't have much time lately to do thorough tests
-GODDESS of Realism - GoR_PONY_v4photo+VAE!b
Sampling method: DPM++ SDE Karras
Sampling steps: 25 up+
CFG Scale: 6.5
Width: 832
Height: 1216
Clip skip 2
For distance shot recomended ADetailer for better faces
Others:
I don't quite know if it makes a difference but it seems to me that it improves the quality of the images
Clip skip SDXL on
Emphasis mode No norm
Enable T5
The models used initially before my own training to create base GoR Pony were definitely Pony Realism and probably CyberRealistic Pony, DucHaiten-Pony-Real, Zonkey - Realism, DAMN! and in v1 Everclear PNY. I don't remember all exactly because I had probably all the realistic PONY models on disk. Thank you to all the creators.
GoR pony v3
The basic idea is to create the most versatile model, from nice views to NSFW and from robots and monsters to fury. V3 derives from v2art but managed to get crisp colors like in beta or 1. While the quality of textures is 100x better. I noticed that this version likes tags from sd1.5 such as HD, cinematic etc.... Also, the negative long shuttles have a very big impact but be careful.
Beware of too big values especially for tags related to lighting (1.3), (1.5) and too much (((()))) etc.
The model has a lot to be perfect, I'm still not satisfied with the face and hands in the distance shots. If anyone finds some lore or embedding that improves this please share in the comments.
Enjoy
V2art_fix+VAE!
All DPM++ works good,
for scifi/art 2m SDE and 2m SDE Heun
for human realism best DPM++ 2S a and DPM++ 3M SDE, DPM++ SDE Karras
CFG scale 8-9
Steps 30up
VAE baked in
clipskip 1
wuler and euler a not recomended
Edit : The size of the model is a consequence of using the "checkpoint model mixer." I checked with other settings and it just comes out that way. As you can see from the added images it works correctly
Edit 2: LORAs trained on Pony Diffusion and Pony Realism works good. On Everclear not.
I turned on "Lora/Networks: use old method" in a1111 settings
GoR_PONY_V2art:
I added some custom lora to change mood of images
It is very sensitive model so don't go to high with ((()))) and (1.4)
All DPM++ works good, I prefer 2M karras
GFC from 5 to 7
Steps 30up
VAE standard sdxl
clipskip 1
If you see others problems sher it so I can fix them in next version
tested in automatic1111
GoR PONY_v1
GFC works up to 10
DPM++ 2M karras = fast
DPM++ SDE karras = quality
3MSDE, Euler, Euler a = works fine to
Steps= 30-45
VAE standard sdxl
Be careful with adding too many tags like: photorealistic, photography, bokeh etc.... Sometimes their excess causes errors
GoR PONY_beta
DPM++ 2M SDE karras-30steps for simple image and 40-50 for more complicated,
for more info check posted images dataset and compare of all samplers
GFC: 2 for realistic up to 7 to get more vibrant
Pony "scores" and tags for positive and negative
SD:1.5 description below
This is a mix of very many models from this site. The main goal was to create more artistic photographs in different styles.
Prompt;
photo type > angle > subject >subject details > background > quality tags (not necessarily) > style (not necessarily)
samplers
best with DPM++ SDE Karras
steps 25+ gfc 7
hires fix as you like ( I prefer foolhardy)
Clip Skip 1
It dislikes "full body shot" use full shot and others
Negative prompt
dangerdonkey, kkw-new-neg-v1.6 ,(low water quality:1.45), low particles efects, ( 3d render, anime, draw, cartoon, grotesque),( over saturated:1.1), jpeg artifacts, out of frame, (((watermark, signature))), lowres, (low quality skin:1.1), , error, compressed, 000, dolly skin, (plastic doll), barbie, (lolicon, loli:1.2), childlike, 000, bad anatomy, bad hands, mutated hand, missing fingers, fused fingers, extra fingers, extra arms, long neck, amputation, (poor quality feet:1.0), wrong number of feet, blending objects, extra digit, fewer digits, cropped, out of frame, error, extra limbs, mutation, bad proportions, cropped head, malformed limbs, mutated hands, mutated feet, missing limbs, cropped, deformed, ugly, mutilated, disfigured, face cut, head cut, legs cut, poorly drawn face, animals, nsfw,
Description
Better: faces, skin textures, not so overburn, more photrealistic
FAQ
Comments (32)
two prompts that can help generate better "full body shots":
1. "pov" as it generates a direction
2. "ideal" as it forces better faces
but those are only working on the sd version of this checkpoint
Wow! This is great!
nice checkpoint ;)
Extremely good checkpoint! Very good with ControlNet too! Please develop this further!
in all my tests, your new pony 1.0 model actually looks less realistic than your pony beta model... the beta model makes outputs that look much more like real photos, more detailed skin, while the 1.0 model looks too "perfect", more anime style, too smooth skin etc
Yes, i know. Next fix
@Oppkllll nice, so you will soon make a v1.1 with a fix?
@OrangeJuiceAlien fix is redy, but I want to train it a little
This checkpoint is missing training with Danbooru tags.
Dood you overbaked it, dont you see? Its distorted and output is fuzzy, instant delete, no model should require over 20 septs nowadays, and even with 30 its still fuzzy on SDE++ tahts twice the time .You can see on lots of examples that this is not ready for release
Specifically which version?
@Oppkllll beta ! Im testing against other stuff like authuman and your merges are totally opverbaked, maybe merge with something crisp to save it
@gogorangers890 I agree with you, I didn't think this version would gain any popularity. I miss some kind of tutorial for MBW under SDXL/Pony. In 1.5 I knew where what was, and here blindly
@Oppkllll I like taht your model has vignette cause most realistic ponies have this flat look, i hope someone wil merge some vintage photo style so its not so sterile from a lab
the more fuzzy output of the beta is good. it looks realistic. people on reddit say the beta of this model is the currently best realistic pony model: https://www.reddit.com/r/StableDiffusion/comments/1e6pxo5/i_created_a_improved_comparison_chart_of_now_20/
Man, how unperfect this model is, it is stunning !
I would really like that model because it creates very good backgrounds.
But I'm always having a white/gray veil over the images. Even when I use your examples.
I'm using Automatic1111. Any ideas what could cause this? The same settings work fine with other models.
show something
but probably wrong VAE, use standard sdxl vae
@Oppkllll I'm using the standard vae. I will look into it. Ty.
I switched to the fp16fix VAE. Looks better now. The standard VAE is not working for me. Don't know why but I'm still not an expert.
Hey, could you add preferred VAE next time alongside? I actually like the results without any VAE, but it creates glitches. All and all it says that VAE creates some bias for the model (everything becomes darker and less cinematic)
works extremely well with loras trained on ponyrealism 2.1
@mongot what are your top 3 favorite Loras that you would recommend that go well with this model? Would you mind sharing the Civitai links?
Also, what is the best way to find out which Loras are trained on PonyRealism 2.1?
@pqnisher i use my own loras pretty much exclusively, so i cant really help you there. i've uploaded two loras trained on ponyrealism and i just added one trained on this checkpoint. but i'll likely upload more in the coming days.
no shade, unironically: what's the point of these real pony models? dream booth style 1girl portraits? why use pony then? seems they all just destroy flexibility and anything "pony" about them. is this one different? are any..?
Technically there is no way to make a PONY model perfect realistic so as not to “damage” it a bit. And why? Because not everyone likes the style of anime and such mixes can still do more than sdxl models in the nsfw/wild category. In photorealism and art we have great sdxl models to work
@Oppkllll so people just like this kinda barbie look? i'm really only asking cause it's tough to gauge what you can do with models beyond what's 'super popular' in the gallery.. i got downvoted to hell but i'm just asking a (imo) pertinent question lol
edit: for clarity (if anyone here cares), the NSFW/people factor in Pony isn't my jam, i like finding models to supplement animation backgrounds, objects, environments and i kinda like what i've gotten from some Pony models.. to each their own eh
Petition to enable onsite generation
Hey, this version is a little broken but Beta (Pony) works very well. We are awaiting a fix for 1.0. It's a very good if not the best realistic model! I would know, I have more than a dozen =)
Nice model, but does not follow prompt as good as PonyRealism V2.1 based on few tests
I have Stable Diffusion GUI 1.11.0, but I can't install it. I run 'Convert Models' and I can click "convert", but always I see: 'Failed to convert model'.
Please, help?



