Type | |
Stats | 137 20 |
Reviews | (16) |
Published | Jan 11, 2025 |
Base Model | |
Training | Steps: 4,300 Epochs: 100 |
Usage Tips | Strength: 1 |
Trigger Words | The image is a painting from Carl Spitzweg inbetween The painting is done in between a realistic and impressionistic style |
Hash | AutoV2 4D1F67880D |
version 2.0 update:
The paintings had some very different stroke styles, so I sorted them in three groups: the Biedermeier realistic style, the later impressionistic style and those inbetween. I trained each of them separately and then tried to merge the three LoRAs. But that failed, it was not possible to trigger the three styles. So, I had to publish all three LoRAs. The images for the showcases are made with the same prompts, just different keywords for the styles. You can pick which one you like best - it is interesting which difference the strokes make. No surprise there that the v1.0 version is close to the inbetween style, which may suffer from the fp4 precision used for training (visible grid pattern in bright or dark parts of the image).
I had the spontaneous idea to make an image in the style of Carl Spitzweg, a German artist from the 19th century. But Flux didn't know the style even if the creators call themselves "Blackforest Labs". I was halfway through aquiring and tagging the paintings when I found out that someone else already made a LoRA. But I continued anyway because with that one, I got funny results - the buildings very much looked like Spitzweg, but the people didn't.
Spitzweg started painting in the Biedermeier style and later moved on to Impressionism. I could get a dataset of 194 paintings which are a mix of portraits of people while doing activities and landscapes. There is a strong streak of social criticism in his paintings, depicting situations which are part of everyday life but don't line up with the social norms of the time. A lot of yawning monks and soldiers doing everything but standing on watch.
After I had all the paintings together, I realized that they would have looked much brighter and happier as they do now, due to age. And with prices between 30k and 150k Euro, the effort made to restore is not very likely. It also looked like many of the images were taken on analog film and the film also degraded. In short: I restored the images to what they would have looked when new.
On my test runs, it looked like the LoRA with 75 epochs made the images best fitting the style, starting with "The image is a painting from Carl Spitzweg of " and a LoRA-tag <lora:CarlSpitzweg-75:1> at the end. Keep the guidance low (at 1), otherwise it also deviates from the style.