It's everyone's favorite android-thing, Haydee!
This is a lora trained on 136 hand picked images of Haydee from... Haydee. There are no 3d/sfm images included, it's exclusively digital art. It handles her features pretty well, and can do most angles you can think of and usually gets the mechanical arms and head right. It was trained on danbooru-style tags, so use those when trying to prompt - I've included the tags in a .txt file in the "Training Images" download below.
It works pretty well with most ckpts, but I recommend Anything 4.5, I used it for most of the example images.
NOTES:
-The helmet and the thing around her neck is virtually unremovable, even if prompting nudity. It also seems impervious to trying to make it a different color, although the leotard she wears is easier to change colors.
-It can do a "male" Haydee pretty easily, too. See last example image.
- It can kind of do the handgun from the game, however the results can vary.
Example prompt: (See example images, most images still have their metadata. Just drag the images onto the PNG info tab and import it.)
Thanks, and have fun!
Description
FAQ
Comments (9)
YES! Was hoping someone would make this, can you make Cortana and EDI from mass effect.
We need more ROBO gals :D
How do you change the color of the leotard or arms ?
If you add something like "((pink leotard, pink skin, pink mask))" to the prompt, and "white skin, white leotard" to negative, you should get results. You also would probably have to lower the weight of the lora a bit, maybe to 0.8 or so.
@ZoochMan Indeed, I never thinked about adding those to Negative prompt, now, at least 5/8 images have the desired color. Excellent Lora by the way ! Can I ask you how do you achieve such good training ? Like the helmet being unremovable ? Even with almost not details added, a clear haydee pic is given.
@Shadw It's surprisingly easy once you get the intial workflow down. I just download all character images I can find online from any art website I can think of with a tool called Grabber, then I manually sort through all of the images (usually in the hundreds). I look for images where the character is the main part of the image, and where their important details are (outfit/face/etc), and try to mainly train on digital art. Then I just train with the LoRa easy training script with mostly default settings. This model came out quite well mostly because her design is fairly simple: Helmet, Robo-arms, Leotard.
The hardest part is probably datasetting, which is sorting through those images, but after that it's quite easy to make these.
What model did you use for those 2.5D examples? I tried AnythingV4.5 but I didn't get same images.
If you imported everything directly it shouldn't work directly. You'll need to fiddle with the resolution, and I might have used a seperate lora to achieve a certain style. If there is a specific image you are trying to emulate, comment on it and I can look at the metadata myself and tell you if that's the case.
Works brilliantly, thank you!!!!
Awesome! Thanks a lot!