The Lora is trained on 140 MidJourney Images that use my "Personalised Style" + Art Nouveau
The dataset is diverse and tries to capture the true spirit of Art Nouveau seen through my lens, not just portraits in Alphonse Mucha style.
Works without the trigger words ArsMJStyle, Art Nouveau, but gives stronger results with them.
MJ7 VERSION:
The dataset was rebuilt entirely from scratch using MidJourney V7, once again mixed with my personalized style.
While the overall composition is similar to V2, this version expands the range by including everyday objects, architectural details, and jewelry.
Expect improved structure and more natural colors.
V2 Changes:
I've completely rebuilt the dataset, to better reflect the evolution of my MidJourney style and LoRA structuring/labeling I use.
This new version (v2) is based on a curated dataset of 250 images:
100 Portraits
150 Landscapes, Buildings, Animals, and Abstract Paintings
The colors and composition in v2 offer a distinct experience compared to v1, so I encourage you to try both versions and see which one you like the most.
Recommended settings remain the same as v1.
Description
Recommended range is 0.2 - 1 strength .
The LoRA give REALLY strong effect. I would recommend starting low ~0.6 and adjusting according to your preferences. If you don't use other LoRA you will have significat difference in the composition and colors compared to stock Flux even as low as 0.2 .
I would not recommend using it above 1.3 strength.
All images in the Showcase Gallery were created using this workflow and the Flux Q4 Dev model on a machine with 8GB VRAM:
https://civitai.com/models/658639/super-simple-gguf-quantized-flux-lora-workflow
FAQ
Comments (4)
It generates same stained glass frame on every image. Not much use of such Lora it it just outputs same picture.
Not sure what version you are checking, but you can lower the weight/add negative , if you don't like the frame :)
I was thinking about fixing / extending this at some point as I have long abondended the process I used for the dataset creation here for a much better one.
@ArsMachina Late in the train but could you share what this 'better one' is?
@mmarshal Sure.
My dataset gathering process is much more extensive (minimum 150 images when I make LoRA, with greater diversity from the start),
+ I used to cherry-pick from ~400 images to create a dataset, now I usually start with 1000, but the numbers are quite high for some of the LoRAs ( 80s Fantasy 3 is 1000 images dataset , cherry-picked from ~12k )
+ For caption I use 2 sets of taggers one after the other , Florence 2 Large and WD 1.5
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.