Type | |
Stats | 429 |
Reviews | (45) |
Published | Aug 7, 2023 |
Base Model | |
Training | Steps: 43,600 Epochs: 4 |
Usage Tips | Clip Skip: 1 |
Trigger Words | Barbarella BattleAction FlightSuit FurryTail IceForestExplorer LabyrinthExplorer Mathmos Nightclub SpaceSuit Labyrinth background +9 more |
Hash | AutoV2 9BFAD8A560 |
SD v1.0
This is trained from some images sets of Barbarella and her outfits hastily thrown together on a whim. I intended to go through each training image adding a custom txt files, but then I decided to try training a test version with minimal tags (Barbarella, woman, outfit) while I was at work just to see what would happen. I had low expectations that it would even make a coherent image, but was pleasantly surprised.
It can be a bit limited with poses and some outfits had fewer training images to work with than others, but its fun to see what pops out.
Warning: if you don't specify an outfit, it tends to prefer none at all. The nightclub does as well and I missed removing or cropping one of the training images for that one that had a small text bubble in the corner so it like to put text and other odd things in the corners with that outfit.
A simple prompt with just her name and an outfit such as "Barbarella, FlightSuit" will often try to create an appropriate background inspired by the movie, but you can also specify your own background if you want something else.
If the face doesn't get a good likeness, try inpainting a better one at around 768x768 resolution.
You can also try dressing your own characters in some of her outfits with varying degrees of success.
I might continue to clean it up and improve the training set and train a better version in the future. I'm open to any suggestions that might improve usability.
I have also posted a Jane Fonda (c1960s) LoRA for more general purpose images other than the Barbarella character.
SDXL v1.0
I have added a lot of additional images to the training set and added some additional secondary concepts to help direct the images a bit better. It seems to have a bit of trouble creating coherent images with the spacesuit, which is to be expected given she's spinning around in non-standard positions while removing it in the credits. It also has trouble with nightclub outfit and tends to want to make it symmetrical instead of lopsided. If you are getting a mixture of the IceForestExplorer and the FurryTail together when trying to generate the FurryTail (might only happen when trying to use "IceForest background"), try putting IceForestExplorer in the negative prompt. You can also do something like "icy lake background" instead of "IceForest background" to get an okay substitute background. If you are not getting the tail, try adding "fluffy tail" to the positive prompt. The face is hit or miss. You can also put her in totally different backgrounds and scenarios outside of the movie sets.