Sign In

Jay Garrick (The Golden Age Flash)

7
63
28
1
Verified:
SafeTensor
Type
LoRA
Stats
44
28
Reviews
Published
May 23, 2024
Base Model
SD 1.5
Training
Epochs: 10
Usage Tips
Clip Skip: 1
Strength: 1
Trigger Words
GA_Flash_JG
Hash
AutoV2
8965BCD7C6
default creator card background decoration
91signs's Avatar
91signs

This model is intended to depict the Golden Age version of the Flash, Jay Garrick. I used a few of images from the CW Flash tv show, some screenshots from various DC animated productions, a bunch of comic book images, and a bit of cosplay. I've mainly tested it on photorealistic checkpoints, but I'll be curious to see how it holds up to other styles.

Trigger is GA_Flash_JG

Start at Strength 1.0, unless you're mixing it with other LORAs.

Trained in base 1.5, with Booru tagging, using the hollowstrawberry colab notebooks.

This is still a work in progress, and can be a little unstable. I'm uploading two versions. Neither one is inherently better than the other.

--Version 1 tends to favor the John Wesley Shipp/CW costume in closeups rather than the classic comic look with the big lightning bolt, especially if it generates a picture of an older Flash. I used images of both modern stories and those set in the 40s and 50s, so it can generate both young and old versions, with proper prompting. It struggles with the helmet--getting the little wings on the helmet tends to require either trial and error or inpainting.

--Version 2 used a bigger sample set of comics images, helping it stick closer to the comics look...when it works. For some reason many of the epochs in this run totally lost the plot and started generating female athletes, or Grant Gustin from the CW Flash (despite there being no images of either in the dataset). This final epoch seemed to pull itself together, but it can still be a little unstable. ADetailer is an absolute must for faces, likely due to my reliance on low-resolution Golden Age comics scans with inconsistent faces and anatomy.

Expect a Version 3 shortly--I'm currently working on a better dataset to cure the weaknesses of the last two versions, but I thought while I work on that I'd put these into the wild to see how they work for other people.

I plan to work on a mix of comic book and other nerd culture characters, with occasional real people from time to time for variety. Let me know if you have any advice/requests/etc.

While I ask to be credited for any use of these LORAs, all intellectual property rights to the underlying artwork and character remain with DC Comics and its affilliates and licensees. Please only use these in a manner consistent with the principles of fair use and with the CivitAI TOS.