art.jt

Hilma af Klint

about this artist ↗ · custom FLUX.1-dev LoRA

01 Model & Methodology Filed 2026-05-05

Custom FLUX.1-dev LoRA, fine-tuned on Hilma af Klint's own canvas work.

A LoRA (low-rank adaptation) is a compact set of trainable weights that sits on top of a much larger frozen base model. Instead of retraining FLUX.1-dev, the open-source 12-billion-parameter text-to-image transformer from Black Forest Labs, the LoRA only adjusts ~50 million new parameters using a small artist-specific dataset. The result keeps FLUX's general image-generation ability intact while teaching it one artist's visual language.

Pieces in this gallery are produced by a private FLUX.1-dev LoRA trained on 27 hand-curated reproductions from WikiArt catalog, covering the Paintings for the Temple body of work (1906-1920) including The Ten Largest, The Swan, The Dove, Atom Series, and the Altarpieces. Every image was captioned with its title, medium, dimensions, and year, binding the trigger to specific painterly cues rather than a generic style label.

Base Model
FLUX.1-dev
Training Set
27 canvases
LoRA Rank
32
Steps
1,500
Train Time
12 min · H100

Trigger word hilma_af_klint_style is prepended to every weather-derived prompt before inference. Trained via ostris/flux-dev-lora-trainer on Replicate.