Here is my Swamp Monster model for the Civitai Bounty Contest.
The model works well with a weight around 1. The trigger word is 'Gill_man'.
The model is based on the legendary Swamp Creature Gill Man from the 1954 film 'Creature from the Black Lagoon'.
In murky depths, where shadows play,
Gill Man lurks, by night and day.
Swamp creature with scales of green,
A mysterious legend, rarely seen.
In the moon's soft, eerie glow,
He swims where dark waters flow.
Description
FAQ
Comments (11)
I'm seriously impressed. Rarely do I find a LoRa that works so perfectly and adheres to any prompt I come up with. Really good job! As someone interested in making loras themselves, may I ask what your dataset looked like to get this so good?
Thank you ^^ The dataset consists of screenshots from the movie 'Creature from the Black Lagoon'. I got the images from Google. Then upscaled them using Topaz AI scaler, created tags with WD14, and manually edited a bit with Bebooru dataset manager. Then, I trained it with CivitAI Trainer. I can't recall the settings right now, but they were close to the presets.
@denrakeiw Wow, really? It came out pretty dang good for screengrabs - albeit jazzed up screengrabs. I threw a few random prompts at it to see if it was one of those LoRas that just keeps giving me the same thing - but nope - I had my own Creature from the Black Lagoon in a phone booth, sitting on a sofa, doing whatever I want! Really awesome. I'm looking in to how to do a lora on Civitai - never have made one anywhere before. I have no idea what resolution/size source images should be and I assumed they had to be super high quality or something - and hundreds of them, but I'm starting to learn that that doesn't appear to be the case. Thanks so much for the info! And even more for the amazing LoRa! There's not nearly enough 'classic' type loras out there, so I'm glad to see one.
@Picky2011670 You're right, they weren't just screengrabs, there's also an action figure of the creature, whose pictures I also used. And a few drawings of the creature. Due to the bucketing function, you can go for a fairly high resolution and also different aspect ratios. The number of images is tricky, I always like to use quite a lot of images. But I've also trained pretty good LoRas with only 10 images.
@denrakeiw Oh cool, so you can use mixed image types (screen grab, figure, drawing) - maybe that's what helps it be more versatile. I'd heard 50-100 images is a good number to shoot for for an xl lora. Maybe a Frankenstein one would be good. So few classics!
@Picky2011670 Yes, for a style Lora, usually more is better. However, for a character Lora, it's more about the quality of individual images. Here, only take the best pictures. It's better to choose just 10 good images rather than 200 mediocre ones.
@denrakeiw Point taken, thank you! Quality over quantity. How'd you learn all this, anyway?
@Picky2011670 I wrote a book about stable diffusion and model training. During research, one delves deeply into the subject matter. Otherwise, it's a lot of experimenting. https://www.amazon.de/Diffusion-umfassender-Leitfaden-Bildgenerierung-Modelltraining/dp/B0CL3FJVRC/ref=sr_1_1?crid=JTHIUY9E4FVV&keywords=dennis+sch%C3%B6neberg&qid=1702189944&sprefix=dennis+sch%2Caps%2C86&sr=8-1
@denrakeiw Oh, didn't know you were THAT much of an expert! Is it only in German right now? I'd go for it I spoke or read German! Pretty impressive if you wrote a whole book on it!
@Picky2011670 Well, to be honest, ChatGPT wrote most of it, but I told ChatGPT what to write ^^
@denrakeiw Ha, fair enough. Wish it was in English so I could read it! Now I'll just have to go and learn German! Nein!