Three Versions the difference between is mainly dataset.
Ver 1: 50 photos US Department of Commerce and Labor, different people working nice high res and quality old photos.
Ver 2: previous 50 + 50 generated in style by Grok with focus on light and contrast.
Ver 3: previous 100 + 100 variety from panoramas to nudes and objects.
Check them all i really like to bump strength to 1.5 and gen in 2048 resolution, a more excessive prompt can help with the damaged look.
Description
Version 3, added 100 more photos, this time large variety landscapes, people, nudes and objects. I'm not saying its better then previous versions but it is different for sure. I like to bump strength to 1.5 and generate in 2048 resolution.
FAQ
Comments (2)
i love that color effect in old photos, even in movies they are hand-painted
It is just great. Besides its strong performance in rendering 1900s-style photos, I used it off-label at a negative strength to prevent contrast burn from another high-strength LoRA. It is remarkable how well it achieves this effect without sacrificing silhouette stability.
Other LoRAs I tried for this purpose usually messed up the composition at -0.42 strength because fading contrast lines made it hard for the model to determine shapes in the noise. I suppose the large amount of dark canvas tones in the training data drives this effect—silhouettes aren't just defined by fine outlines, but by the overall frame fill. If you are serious with diffusion generated pictures, this lora is a must have. Thx for sharing.