This model has been trained on photographs of young adults taken in Tokyo during the 1980s. The photos were taken on 35mm film using consumer cameras. This results in a model that creates candid compositions, often with bright flash lighting. The distinctive tones of 35mm film and tungsten lighting are recreated, along with the bold fashions of young Eighties Japan.
Description
FAQ
Comments (2)
Just wanted to say thanks for all the off-the-beaten-track stuff you post here! I'd be curious to know if there's anything different you've had to do/learned to do in the training process, since you're using such unusual materials. Like, do you still crop all your images down to 512x512?
Thank you very much for your comment, and I'm glad you've enjoyed my work so far. I've writing an article about my training method which I hope to have live here on Civitai in the next few days.
The most important thing I've found is image selection: I'd rather have 5 images that absolutely exemplify what I'm trying to capture than 25 images that only sort of exhibit it. I train 10-15 epochs and save a checkpoint for each, then test each checkpoint with the same prompt to see which is the best representation.
I scale the images so that the shortest side is 512px. I use the batch process function in IrfanView to do this. I use Kohya's toolkit for training LoRAs, and fortunately this can handle images in multiple sizes within a training set, so there's no need to crop the source images to be square.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.