Training Setting/parameters
First off thank you for sharing your files they are fantastic? By any chance could you share the parameters you used to create VikingPunk? Such as The vector per token, Learning rate, Which Prompt template file, and what not. Also how many images were in your data set? The last custom models I created with dreambooth we're pretty good I was quite happy with what I've been able to create so far, although I'm still tweaking things, HOWEVER When I used your VikingPunk embedding with my created model, The results were amazing. The details were fantastic, More than what I get just using the standard V2.1 model or my model alone.
I also used a combination of midjourney images with other images when training my dreambooth model. So not sure what is happening but something with my model and you're embedding really amplifies it. So I just want to play around with it more, if you could share your setting that would be so helpful. Thank you for your time.
Hey there,
For Vikingpunk I used 6 vectors and 53 images using deterministic sampling method. I dont remember the learning rate for this one exactly but it was around 0.004, slightly lower than default. The images were an mix of midjourney and my own trained 1.5 model, slightly edited to fix mistakes or adjust the color scheme.
The prompt template file is usually a "[filewords], [name], + style/lighting descriptors", with around 5-10 prompts to add some variation. All images have a corrected CLIP interrogater description as well.
Whilst the settings are important, 75% of getting a nice result is still getting a good dataset to train on. So make sure your focus is on that.