Amateur Selfies is a LoRA trained on ~1000 casual self-portrait photos, designed to capture the look, feel, and imperfections of real-world selfies. It emphasizes natural framing, handheld angles, everyday environments, and the aesthetics of self shot phone photographs.
Recommended settings after more experimentation
LoRA strength: 0.9 - 1.0. I accidentally generated the 'redux' showcase images with 0.9 and they're fine.
FP8 or BF16 UNET. GGUF quants cause checkboard artifacts, do not recommend! FP8 can actually 'enhance' the amateur look
50 steps
CFG 3.0
euler/beta
<= 2.0 MP
Workflow is included in the sample images.
I have identified a couple of issues with training captions and masks:
captions were natural language formatted in paragraphs. The newline paragraph separators however are interpreted by SimpleTuner as delimiter for individual captions.
masks weren't applied due to a derp in the dataset config
I'm working on training v2.0 now using the same image set, but with new captions and fixed masks (face removal). Also appending some quality captions which should improve the sharpness of outputs when not prompting specifically for poor quality / compression artifacts.
Description
It's not perfect, but is most cases produces a better result than the other ones I have tested.
966 images. 7 Epochs, batch size 8
FAQ
Comments (10)
this is good but why does everyone create these from 1990s images? It's possible to capture the aesthetic but have high quality images, it's 2025!
Probably because there is an abundance of training data captured using older tech..
I'm trying to partially address the default of blurry / poor quality outputs with the captioning of the dataset for v2.
For a potential v3, I might see if I can do a two-step training process, and fine-tune it with some high quality images
Here is a test output for the in-progress v2.0: https://civitai.com/posts/22990803
@psychologicau kindly do not listen to these suck a55 suggestion because newer pics or dataset will be likely 100% AI distorted. all nowadays cameras has AI processing inside , if a someone bypass this it will goes to next AI distortion from reality which is called filters. using high quality "real" images is great, but using fake AI touched is not needed (we can fake AI touch the generated images later by editing lol) , at this point that 1990s or even 2000s mobile phones picture is more handy.
@amazingbeauty there is some valid criticism in there though - the output quality from v1 was generally bad and I could somewhat fix it with better training captions. v2 can be prompted for the same stuff by specifically asking for blurry / jpeg artifacts.
I think what I said ^ about potential v3 is unnecessary, see the examples where I stacked my other photography LoRA with this one, much 'higher quality' outputs which are closer to what a newer camera could produce
This is more like mid 2000s Myspace time. People weren't doing much selfies online in the 90s really.
@psychologicau okay sorry , i just want good lora for people including me 'till i be able to train loras ..:('
@amazingbeauty try this LoRA in combination with my other one, like I did in this image / post: https://civitai.com/images/104096315
Because the 90s and early 00s were a time before bullshit snapchat filters and fake influencers ;)
Because they look real.