Yet Another Pony Merge!
Merge of Everclear PNY and Pony Realism. Originally had no intent to upload, but it seems to work better with my Fursuit LoRA, so I thought I would upload for the benefit of those who generate using Civitai's services.
VERSION 2.0
Are You Ready for a Massive Improvement?
Despite my personal enjoyment of the original model, people did have some valid criticisms, namely the loss of important details over Pony Realism.
To try to remedy this, this new version was weighted block-by-block over the course of several hours with a number of intentions in mind:
Important details should be retained without having the "deepfried" look that Realism often has,
Prompt adherence should be of maximum priority,
Realistic, believable aesthetics should be as high as is possible,
Colors should be vivid without being oversaturated, and
Lora compatibility should be as high as possible
I consider myself rather humble, but must say that I am extremely pleased with the results!
2 versions of this updated model are included: v2a is the base re-weight, whilst v2b is v2a merged 1:1 with v1.
v2a produces better results in 99% of cases, but in the rare case of things like prompt incompatibility of differences, v2b will be an improvement over v1.
Model weights are as follows:
Model 1: Everclear PNY v3
Model 2: Pony Realism v21
Time Embed: 0.15
Label Embed: 0.5
Input 0: 0.1
Input 1: 0.2
Input 2: 0.5
Input 3: 0.05
Input 4: 0.15
Input 5: 0.9
Input 6: 0.2
Input 7: 0.2
Input 8: 0.85
Middle 0: 0.5
Middle 1: 0.75
Middle 2: 0.6
Output 0: 0.85
Output 1: 0.55
Output 2: 0.9
Output 3: 0.85
Output 4: 0.75
Output 5: 0.95
Output 6: 0.6
Output 7: 1
Output 8: 0.65
Out: 0.95
CLIP: 0.18
Happy generating!
VERSION 1.0
Reasoning behind the merge was that I prefer the prompt adherence and image quality of Realism, but prefer the realistic look of Everclear, and so decided to merge them such that the model weights were primarily Everclear, and the CLIP was primarily Pony Realism:
Model Weights:
Everclear : Realism - 0.85 : 0.15
CLIP Weights
Everclear : Realism - 0.15 : 0.85
In my personal opinion this seems to result in the best of both worlds, but your results and tastes may vary.
Description
Are You Ready for a Massive Improvement?
Despite my personal enjoyment of the original model, people did have some valid criticisms, namely the loss of important details over Pony Realism.
To try to remedy this, this new version was weighted block-by-block over the course of several hours with a number of intentions in mind:
Important details should be retained without having the "deepfried" look that Realism often has,
Prompt adherence should be of maximum priority,
Realistic, believable aesthetics should be as high as is possible,
Colors should be vivid without being oversaturated, and
Lora compatibility should be as high as possible
I consider myself rather humble, but must say that I am extremely pleased with the results!
2 versions of this updated model are included: v2a is the base re-weight, whilst v2b is v2a merged 1:1 with v1.
v2a produces better results in 99% of cases, but in the rare case of things like prompt incompatibility of differences, v2b will be an improvement over v1.
Model weights are as follows:
Model 1: Everclear PNY v3
Model 2: Pony Realism v21
Time Embed: 0.15
Label Embed: 0.5
Input 0: 0.1
Input 1: 0.2
Input 2: 0.5
Input 3: 0.05
Input 4: 0.15
Input 5: 0.9
Input 6: 0.2
Input 7: 0.2
Input 8: 0.85
Middle 0: 0.5
Middle 1: 0.75
Middle 2: 0.6
Output 0: 0.85
Output 1: 0.55
Output 2: 0.9
Output 3: 0.85
Output 4: 0.75
Output 5: 0.95
Output 6: 0.6
Output 7: 1
Output 8: 0.65
Out: 0.95
CLIP: 0.18
FAQ
Comments (3)
For me, what I always look for first and notice A LOT of pony merges have that try to go realistic, is the eyes; There's always that dark circular spot under or around or on the lower eye lids. It's there too in these pics, almost like they're a bit tired. I think it has to do with how pony used.. well pony images and the eyes on them are very prevalent in the model (even with source pony in negs) Just my two cents.
Not something I ever really noticed, but now that you've pointed it out I don't think I'll be able stop noticing it lol.
Eyes weren't something I was especially focused on during weighting, though I have noticed that the eyes appear to be better quality than the 2 models it was trained on, and what you have mentioned seems less prominent in the comparison images I made. Would be hard to fix if it's hard-coded into both models, but I'll keep an eye out for anything that seems to remedy this for a potential future release.
Very nice model thank you for sharing