This is my first attempt at a checkpoint merge. I used:
Absolute Reality by Lykon (https://civarchive.com/models/81458/absolutereality)
Realistic Fantasy by JCSimon (https://civarchive.com/models/123777/realistic-fantasy)
NooshereREAL by scumerz (https://civarchive.com/models/199878/noospherereal)
I checked and none of them explicitly forbids the release of merges. But if I made a mistake, if you do not want me to do what I did, please contact me at once. I do NOT want to violate / disregard anybody's rights.
When creating pictures, I found that I always came back to one of the three Checkpoints I used here. So more out of curiosity than anything else I merged them, interested in what would come out of it. I found the result not too shabby and was even asked about the checkpoint I used concerning some of my pictures.
So ... be warned that this is something thrown together by an amateur who has no idea what he is doing ... but if you find you enjoy it and can use it, be my guest. Please give the original creators at least a like.
I have no idea concerning pruned / unpruned or f16/f32 ... so that information might be wrong.
Other settings:
Sampler: DPM++ 2M Karras
Steps: 30
HiRes fix enabled, HiRes Steps 15, Denoising at .35, Upscale 1.5
CFG Scale 8
ADetailer is used for faces
I have https://huggingface.co/stabilityai/sd-vae-ft-mse-original/tree/main but I'm not sure it is needed.
Description
FAQ
Comments (3)
This is good model. Both speed and quality are good.
With same seed and prompt, output - exact same image as from Absolute Reality. Maybe if try more prompts could find different outputs, but with few attempts is seems just copy of AR.
To reply to you about fp16 fp32
Not pruned = no compression on model (no quality loss)
Pruned no ema = cut values near zero (limited quality loss)
Pruned ema = cut values following exponential moving average (high risque of quality loss)
compression fp16 = quality divided by 2 - bad precision quality on small images/medium image, like bad face on full body in 512x768
fp32 = no compression, best precision quality on small image like face on full body in 512x768
Basically, the general quality is dependent from the training. The fp16/fp32 come when to render image.
fp16 or also fp8 are good to save vram but loss quality in rendering. Use fp32 when you want no compromise with quality.
In render stage, more you have pixels in image and more is the precision more the image is quality. rendering fp32 model on large image is the best.
If you can, give a fp32 not pruned model for quality and a fp16 not pruned to save vram.
a fp32 model size is around 4gb, and fp16 around 2gb. Your actual model is in fp16 precision and has the size of fp32. You can reduce it to 2gb without quality loss.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.