Meet Marsey! An adorable cat from a Telegram sticker pack. I've been trying to get SD to generate more of this character, and wanted to share my results for anyone else working on a specific 2D style.
Comparisons
a photo of a spaceman Marsey in outer space
Textual Inversion / DreamBooth
a photo of Marsey as a lifeguard
Textual Inversion / DreamBooth
a photo of Marsey as a scientist
Textual Inversion / DreamBooth
a photo of Marsey as a gardener
Textual Inversion / DreamBooth
What I've noticed:
Textual inversion:
Excels at style transfer. "elephant in the style of Marsey"
May benefit from more images. My run with 74 images performed better than the one with 3
Best results (both in terms of style transfer and character preservation) at
25,000 steps
DreamBooth:
Far, far better for my use case. The character is more editable and the composition improves. It doesn't match the art style quite as well, though.
3 images worked better than 72
works extremely well with cross-attention prompt2prompt (the "img2img alternative test" script in automatic1111's UI)
1,000 steps (30min on an A6000) is sufficient for good results
Worth mentioning - it's usable with deforum for animations
Combining the two doesn't seem to work, unfortunately. The next step might be either to directly finetune the network itself and apply one of these techniques afterwards, or possibly training the classifier.
Jump in the discussion.
No email address required.
So wait, dreambooth takes 30gb of VRAM to run, right - but does it spit out embeddings that you can use with Stable Diffusion, like Textual Inversion does? I hope someone rents a GPU and makes a big database website of popular characters and shit, especially if you could fetch that data from a stable diffusion client. That would be extremely useful.
Exciting times for AI, very nice marsey results btw![:marseystars2: :marseystars2:](/e/marseystars2.webp)
(this is my 1000th comment!!!!!
)
Jump in the discussion.
No email address required.
It gives you an entirely new 2gb model, so sadly it's pretty heavyweight. It might be possible to train multiple objects into one model in the future, though. I'm expecting all this to keep changing rapidly for a while
Jump in the discussion.
No email address required.
More options
Context
More options
Context