1 MB iA3 Network
*UPDATE*: v6 is trained with Debiased Estimation Loss on the dev branch of kohya_ss sd-scripts. This makes image structure more coherent and have better color. If you train a model with Debiased Estimation Loss, you shouldn't need to use Min-SNR gamma or Color Aug as it does both better.
Prompt format: very dark focused flash photo of <subject> at night, wide angle, black background
I recommend using DPM++ 3M Exponential/Karras with 40-100 steps. Use the refiner at 0.8 for photorealism.
Description
Trained with more images in the dataset and removed images that did not fit the concept.
Trained with 0.1 input perturbation noise https://github.com/kohya-ss/sd-scripts/pull/798
FAQ
Comments (3)
I've used other flash photography Loras, and honestly "flash photo" seems to be enough of a trigger. Personally, I don't combine these kinds of Loras, so I don't see where SD would get confused. And "dark" or "night" can just be optional in the prompt somewhere, but seeing as this Lora is very specific in its intent - and a really useful one - anybody who invokes it should likely expect that it will generate a flash photo in a dark place. Why wouldn't it?
I wrote a bit about why I chose the phrase in this post, see the bottom of the page for the brief explanation. (https://civitai.com/posts/684622)
I think I chose it specifically because I didn't want any contamination from the character "The Flash," which might be a concern if it's just "flash photo".
is not work on apple m1 processor.
error "Trying to convert BFloat16 to the MPS backend but it does not have support for that dtype."
please make model with f32 or f16 dtype