Originally posted to HuggingFace by ProGamerGov
The embeddings in this repository were trained for the 768px Stable Diffusion v2.0 model. The embeddings should work on any model that uses SD v2.0 as a base.
Currently the kc32-v4-5000.pt & kc16-v4-5000.pt embeddings seem to perform the best.
After downloading the embedding, change the filename to whatever you want to use a the trigger word. For example, renamed the file to "knollingcase.pt" in order to use "knollingcase" as the trigger word.
Example prompt words are available on the HuggingFace repo's ReadMe, and these embeddings should work on any model that uses v2.0 as a base!
Knollingcase v1
The v1 embeddings were trained for 4000 iterations with a batch size of 2, a text dropout of 10%, & 16 vectors using Automatic1111's WebUI. A total of 69 training images with high quality captions were used.
Knollingcase v2
The v2 embeddings were trained for 5000 iterations with a batch size of 4 and a text dropout of 10%, & 16 vectors using Automatic1111's WebUI. A total of 78 training images with high quality captions were used.
Knollingcase v3
The v3 embeddings were trained for 4000-6250 iterations with a batch size of 4 and a text dropout of 10%, & 16 vectors using Automatic1111's WebUI. A total of 86 training images with high quality captions were used.
Knollingcase v4
The v4 embeddings were trained for 4000-6250 iterations with a batch size of 4 and a text dropout of 10%, using Automatic1111's WebUI. A total of 116 training images with high quality captions were used.
Usage
To use the embeddings, download and then rename the files to whatever trigger word you want to use. They were trained with kc8, kc16, kc32, but any trigger word should work.
The knollingcase style is considered to be a concept inside a sleek (sometimes scifi) display case with transparent walls, and a minimalistic background.
Suggested prompts:
<concept>, micro-details, photorealism, photorealistic, <kc-vx-iter> photorealistic <concept>, very detailed, scifi case, <kc-vx-iter> <concept>, very detailed, scifi transparent case, <kc-vx-iter>
Suggested negative prompts:
blurry, toy, cartoon, animated, underwater, photoshop
Suggested samplers:
DPM++ SDE Karras (used for the example images) or DPM++ 2S a Karras