I trained an oil-impasto LoRA on 300 public-domain museum scans — weights + dataset notes

by @lunalee_art in c/oil-painting · · Impasto

I trained an oil-impasto LoRA on 300 public-domain museum scans — weights + dataset notes
Free, CC0 scans only (Rijksmuseum + Met open access). Trigger word, caption strategy, and why I excluded anything photographed with flat lighting are in the README.
github.com

Free, CC0 scans only (Rijksmuseum + Met open access). Trigger word, caption strategy, and why I excluded anything photographed with flat lighting are in the README.

675 points · 7 comments

Comments

  • @sergeisharmapro · · 100 points

    Excluding flat-lit scans is the detail everyone misses. If the training image doesn't show the shadow-side of each ridge, the LoRA learns 'texture' as a flat pattern. Grazing-light scans only. Chef's kiss.

  • @raviking_creates481 · · 66 points

    Exactly why I spent more time curating than training. 300 good scans beats 3000 random ones. The Rijksmuseum close-ups are gold for this.

  • @chencosta · · 52 points

    the README alone is a better texture-prompting tutorial than most paid courses. mod this to the sidebar.

  • @jamesmillerlab · · 72 points

    Tested it at 0.6 on Flux, genuinely the first LoRA where the paint has a shadow side. Dropping to 0.4 for subtle, 0.7 goes full Van Gogh. Thank you for CC0-ing it.

  • @meihoffmannmedia · · 21 points

    does it play nice with a glaze LoRA stacked, or do they fight?

  • @miasilva_art2637 · · 34 points

    They fight a little — impasto wants opaque, glaze wants transparent. Run impasto on the base gen, then glaze in img2img passes. Don't stack at full weight.