. ((cars are on fire)), intric.jpeg)
.jpeg)
![01260-3260352924-Zod person (( Riding a horse down the street)) endprompt] _You're not the only one, I don't know why I can't see you, but everyo.png](https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/fac02f38-5b85-43d3-a4be-e0a284de14f5/width=450/01260-3260352924-Zod person (( Riding a horse down the street)) endprompt] _You're not the only one, I don't know why I can't see you, but everyo.jpeg)
. (fire), intricate cinematography, directed.jpeg)
. (fire), intricate cinematography, di.jpeg)
. ((cars are on fire)), intric.jpeg)
) that is harassing him. Film di.jpeg)
Type | |
Stats | 186 |
Reviews | (1) |
Published | May 24, 2023 |
Base Model | |
Training | Steps: 6,000 |
Trigger Words | Zod Person |
Hash | AutoV2 E342DB9D38 |
LoRA of General Zod from the 1980 film "Superman II". Trained with Dreambooth using 50 images of actor Terence Stamp portraying Zod in the 1980s movie. The training was done locally using the runpod/local Dreambooth repo by Joe Pena. Resulting CKPT was extracted to LoRA using the A1111 supermerger extension.
Put "Zod Person" in your prompt to trigger it.
For best results use "Hires fix" in the .1 to .3 denoising range (unless using the latent upscalers which you should use higher).
For best results use VAE: vae-ft-mse-840000-ema-pruned.ckpt
For optimal results use Controlnet w/ Openpose/Canny/Reference/Shuffle/etc and reduce the "ending control step" to 50ish or less to guide the image in the right direction.
Trained on SD v1.5 so it should pretty well on any model trained on that as a base. It worked pretty nicely for me using RealisticVision v2, however for best likeness results use the pure SD V1.5 model.
*Note: Many of the sample images were generated with Controlnet - Due to that you may not be able to recreate many of the samples since I use imput images of random film stills (in 16x9ish format) along w/ Openpose/Canny/Shuffle or Reference for ~50 to 60% to direct the output. This limits "double people" from appearing in wider shots, and also prevents the character from being way off in the distance which often doesn't hold the likeness that well. Also: I renamed the LoRA file before uploading it - the name in the metadata on the samples is the same LoRA model.
Thanks to anyone who shares cool/interesting generations they generate w/ this model. It's fun to see and somewhat motivating to try to put some other LoRA/Checkpoints together.
Oh, and now Zod needs machines to fly - Re: samples - :)