Type | |
Stats | 7,656 |
Reviews | (1,205) |
Published | Jan 2, 2024 |
Base Model | |
Usage Tips | Clip Skip: 2 |
Hash | AutoV2 FC8F4E23FB |
This merge is created with the primary goal of generating female characters with slightly exaggerated proportions. As such, it may not be suitable or flexible enough for all scenarios and use cases.
I've included samples of various character and pose LoRAs to demonstrate how I use them/how they work with the model. I did not do much cherrypicking for the sample images, only discarding those with notably bad hands or anatomy.
The quality of your generations may be heavily improved by using textual embeddings in your negative prompt. The embeddings I use are listed below and in the recommended resources.
There are two styles that I tend to use with default quality tags and embeddings. I use a different sampler for either, but I recommend experimenting.
Style 1 with Euler a:
Less detail, gives a smoother, more colorful look.
Positive - (masterpiece, best quality, absurdres, 4k, aesthetic, detailed, intricate, perfect lighting),{prompt}
Negative - (worst quality, normal quality, low quality, greyscale, monochrome),easynegative, FastNegativeV2, negative_hand-neg,{prompt}
Style 2 with DPM++ 2M SDE Karras:
More detail, gives a sketchier look.
Positive - (masterpiece,best quality,absurdres,beautiful,aesthetic,detailed),{prompt}
Negative - (worst quality, monochrome), (EasyNegative, negative_hand-neg, aid210),signature,artist name,watermark,{prompt}
Images generated using Berry's Mix VAE and 4X AnimeSharp Upscaler at Clip Skip 2
My typical settings for images:
Sampler - DPM++ 2M SDE Karras/Euler a (see above)
Sampling Steps - 35 to 40
HiRex Fix - Upscaled by 2 at 15 to 20 steps, 0.35 Denoising Strength
Dimensions - 512x512, 512x768, or 768x512
CFG Scale - 5.5 to 7
One pass of ADetailer HIGHLY RECOMMENDED
I typically specify eye color/expressions/facial features here (blue eyes, grin, smug, mole, freckles etc).
This specific ADetailer model is made to detect different types of faces, so you may need to crank the detection confidence up or it'll start finding faces in clouds, trees, or whatever.
See images for examples of how I prompt. I prefer short, simple, tag-based prompts, but there is probably a lot of room for flexibility/improvement here.
As this is a pretty horny merge, adding nipples,covered nipples in the negative can help if you aren't going for that kind of thing.
Merging weights and process are available in the 'About this version' section per version.
All credit goes to the original model authors, please check them out.
SCH Excelsior v3 - Original author is gone, this is a reupload by another user