Type | |
Stats | 659 |
Reviews | (67) |
Published | Apr 1, 2023 |
Base Model | |
Trigger Words | t800 endoskeleton cyborg wounded heavily wounded robot |
Hash | AutoV2 03E8AE2E45 |
Trained with about 240 images on stock SD 1.5. I get the best results from about 0.8 weight. Always use the t800 tag and add more tags onto it. Add the 'cyborg' tag for terminators that still look like humans. Add the 'robot' and 'endoskeleton' tags for terminators that just look like metal robot skeletons. Add 'wounded' or 'heavily wounded' to show more battle damage on the cyborgs.
There are some non-arnold terminators in the training data but not many, so most of your cyborgs will skew towards looking like ol' arnie. I haven't had a ton of luck with inpainting, but txt2img and img2img seem to work fairly well.
For the training data, I edited, cropped, and resized each one by hand because most of the automated stuff was doing a terrible job of focusing on the right spot. txt2image will still sometimes spit out endoskeletons with the head cropped out of view though. I also manually added tags to each image because all of the automated AIs are astonishingly bad at tagging t800s. There are slightly more robot endoskeleton images in the training data than cyborg images, but I tried to keep it somewhat balanced.
This is the first model I've uploaded. If you like it, let me know. If you don't, constructive feedback helps. I spent a ton of time on this, so If you merge it into something else please include a mention about my contribution to the effort.
Example prompt: john doe as a heavily damaged t800 cyborg
Go light on the negative prompts. Note that a lot of common negative prompts might directly contradict this model. Negative prompts like "ugly" or "scarred" may completely negate the damaged look normally seen on a t800.