Attempts to imitate Ilya Kuvshinov's style. This is a Dreambooth model based on Anything v3. Best when used in combination with this textual inversion:
https://civitai.com/models/202/kuvshinov
The example images with the embedding upload look amazing, but that's because they're actually the training images and not the generated output. The examples here are generated output from the model.
Not affiliated with Ilya Kuvishinov.
This actually took about 12,200 training steps, despite the name. My disc filled up just after step 8,000, so I resumed the training for another 4,200 steps.
Many example images were generated with this prompt (prompts for each image are included in metadata):
- masterpiece, best quality, 1girl, perfect eyes, perfect face, kuvshinov
- Negative prompt: ugly, greyscale, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, signature, watermark, username,
- Steps: 20, Sampler: LMS Karras, CFG scale: 12, Seed: 3878555472, Size: 512x768, Model hash: f2894629, Batch size: 32, Batch pos: 1, Clip skip: 2
I also used the textual inversion embedding linked above: https://civitai.com/models/202/kuvshinov
Note that I didn't spend a whole lot of time experimenting with prompts, so I'd be interested to see what other people come up with.
Files
kuvshinovStyle_v1.ckpt
kuvshinovStyle_v1_trainingData.zip
kuvshinovStyle_v1.safetensors