Sign In

GODDESS of Realism

2.4k
32.5k
602k
981
Verified:
SafeTensor
Type
Checkpoint Trained
Stats
5,243
Reviews
Published
Jul 12, 2024
Base Model
Pony
Hash
AutoV2
41F3B63DC1
default creator card background decoration
Oppkllll's Avatar
Oppkllll

Please, post your works. I can't do enough tests myself. Thanks to your work, it will be easier for me to improve the model

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,