@X's comment on 'DreamBooth Marseys :marseyastronaut: (added colab instructions)'

DreamBooth Marseys :marseyastronaut:

Colab:

Open automatic1111's colab: https://colab.research.google.com/drive/1Iy-xW9t1-OQWhb0hNxueGij8phCyluOh
Replace this cell:

#@title Normal 1.4 model
# get a token from https://huggingface.co/settings/tokens
user_token = "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" #@param {type:"string"}
user_header = f"\"Authorization: Bearer {user_token}\""
!wget --header={user_header} https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt -O model.ckpt

With:

!apt install megatools
!megadl "https://mega.co.nz/#!FxclSKoL!EFSM4nLlXMuOvBLkoZNmtOH4Y8oycjrU7h2Hn6mKl1k"

Run (it'll take a few minutes to download)
Refer to Marsey as a sks cat in your prompt. a photo of a sks cat as a gardener etc

Download the weights: https://models.rdra.ma/

Development on Stable Diffusion has been happening at a wild pace, and there's already written a working implementation of DreamBooth for it. Whereas Textual Inversion refines your prompt, DreamBooth is able to finetune the model itself to teach it new concepts. I did one run with 87 Marseys and another with just 3 upscaled ones. So far the upscaled run (u1000.ckpt) is looking best.

See the Textual Inversion posts for a comparison with that. DreamBooth is far better at keeping her colors consistent and producing recognizable images. Decent results were something like 1 in 50 before, now they're nearly every image.