MOSS Sharpness LoRA. Blur? What blur?
retained text
A minute ago (literally) another redditor posted their work ("Blur / Anti-Blur slider LoRA for MOSS"); mine is the same idea from a different angle. Good ideas come at the same time; my work is fully independent, and employs a different approach.
I trained a tiny rank 8 concept slider LoRA, teaching the base model what "sharp, in-focus" images are as opposed to "blurry, defocused" ones. The LoRA itself, its training process and parameters are in my Huggingface repo [MOSS-Sharpness-Slider](https://huggingface.co/aoleg/MOSS-Sharpness-Slider). Basically, I picked 72 high-quality images, all real photographs, paying attention to variety: some close-ups, portraits, natural and industrial landscapes, people at work, and so on; varying colors (including some black-and-white scans of large format stocks), epochs, and subjects. I included a number of images with bokeh to teach the models that it's the subject that must be in focus, not necessarily the background.
These were the "sharp, in-focus" part. Then I built a contrasting dataset, blurring those same images with Gaussian blur, varying the radius, teaching the model what a "blurry, defocused" image is.
Trained for 500 steps, picked step 400 as the first one that converged without overfitting.
Use it the way you'd use any other slider: weight 1 is "sharp image", -1 is "blurry, out of focus". You may still get an occasional slightly-out-of-focus generations in e.g. portraits every once in a while; they are rare, and I didn't want to fry the thing. More steps give more sharpness but less control on depth of field; I decided to keep it gentle.document fetched 2026-10-27 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xg55p3/moss_sharpness_lora_blur_what_blur