How an idea became a project in a flash ...
Flash photography is not necessarily the most beautiful way to depict people or objects, but it is the most honest. Ruthless, nostalgic, dirty and cold. The brief, blunt moment in the flow of time, captured by a harsh flash. Now available for Stable Diffusion.
This LoRA is the converted version of an idea to make images not "beautiful", but ... different. It is only used to make people or things visible in a dark environment.
The correct application for it is therefore clearly in the images of:
night situations (streets, alleys, parks, ...)
dark rooms (discos, clubs, cellars, apartments without electricity, ...)
generally dark places (forest, caves, castles? ... )
A bright sunny beach rather not (although that might even be possible under some circumstances). :)
The LoRA was created with SD v1.5, but I recommend the use of other photorealistic 1.5 checkpoints. First of all Analog Madness v3.
The style brings partly a little analog colors and a certain blurriness with it. This is quite intentional. The majority of the images were trained with Polaroids and the like. Not every one of them was pin sharp.
It was trained without references to people, animals or objects. It should therefore be quite flexible. However, it may be that a little more prompting is needed for the one or the other checkpoint.
The model has a slight tendency to represent asian ethnicities. With some minor prompting this can be counteracted.
This project will continue to evolve and everyone can contribute. Anyone who has flash images and would like to provide them, please feel free to contact me.
And now ... get flashed!
Update:
For the perfect subject in the flashlight, check out Sibience's Body Horror Creatures!
Description
More flash! Version 1.5a is now the "proper" first release of Flashlight Photography LoRA.
Changes made:
Handedited some images to better quality
added 50 images to a total of 110 images used for training
FAQ
Comments (19)
Great stuff, You must have very good training set because hands are very good - this lora can help models to generate body parts better (besides flash effect). I prefer v1 because effect seems more natural but great work anyway!
Thanks for your comment. Actually the LoRA should not have any influence on the anatomy. I think it's the checkpoint, it's really good and my most used one with this LoRA.
@woisek From my experiences LoRA contains various informations not only those that they were intented to learn, for example about images compisition
@krodek502 OK, that's why it's great when many users are using and testing this. To get mor information about it. 😀 Thanks for getting back.
Hello. Tried couple of times and it works really well. So I would like to know if you allow me to use your LORA in my SD generation service (competely free to use)?
Hi! I'm glad you like it! 🎉 If it's a free service, it's OK to use it. Thanks!
@woisek thank you for your kindness. Of course it's free and I have no plan to introduce any other paid plans or other such things.
@smugo Great! I wish you luck with it. 👍
This is such a cool lora, but its giving all my people tattoos even though I have tattoos in the negative prompt
Hi! Glad you like it! 🎉 I'm not sure about the tattoos, I don't think that any person in the training set had one. Also, Maybe you tell me your prompt and checkpoint you use, so I can test it.
My models don't understand why there is so much light in the image and adds lamps everywhere. The best results are with "street" or smth like this prompts.
Thanks for your feedback. Could you provide what checkpoint you used and maybe the prompt where this behavior occurs with? I had never experienced this what you describe.
Nice occupation of the frozen embedded token 'flashlight' as trigger word. Bet that works wonders for it to work with other LoRA or with embedding.
"A flashlight (US), or torch (CE) is a portable hand-held electric lamp."
Thanks for your comment. Yes, I didn't think much in the beginning. When I can create v2.0 it will have a different spelled trigger word.
@woisek try using the tags for the dataset images.
flash photography, night flash photography,night portrait,bounce flash,fill flash,fill light
Then to increase pose control specify how much of subject is visible using
full-length portrait, three-quarter portrait,bust portrait,headshot portrait
And for what part of subject you see
front view,side view,rear view, three-quarter view,low-angle view,high-angle view,dutch angle, panoramic view
Google them to see explanation of each. Much better than from below, from above,from behind and similar since they do not specify that it is the view and can just as well be specifying something approaching from that. They are speficy movement more than composition
This is amazing, but for some reason it causes the scene in my image to tilt to the right a lot of the time. Any idea why?
Thanks for the comment. I never experienced a tilting towards any direction. All used training images were straight. Did you make sure it's not caused by some word(s) in your prompt or maybe a LoRA you use? 🤔