@alphonse mucha
First of all, I want to clarify that one of the main purposes of training this LoRA was to automate the captioning process through vibe coding.
There are 171 Alphonse Mucha artworks registered on WikiArt, and this LoRA was trained on all of them randomly. The dataset includes not only Mucha's famous Art Nouveau decorative style, but also other types of works such as self-portraits, sculptures, and oil paintings. Therefore, this LoRA is also intended to generate those less commonly associated aspects of Mucha's work.
For tagging, I used "SmilingWolf/wd-eva02-large-tagger-v3", and for natural language captioning, I used "qwen2.5-vl-7b-instruct-heretic-i1". I combined the outputs of both systems into a single TXT caption file. However, I am not fully satisfied with the accuracy of either method.
The former, in particular, frequently produces incorrect tags, such as miscounting the number of people or hallucinating objects that do not exist in the artwork (for example, tagging "mask" when there is no mask, "weapon" when there is no weapon, "deer" when there is no deer, etc.).
Despite these imperfect captions and inaccurate tags, Krea 2 is still able to generate reasonably good results. Even with such a rough captioning process, the model performs surprisingly well.
However, the recognizable "Art Nouveau decorative elements" tend to appear quite strongly, and those decorations are often generated in circular or ring-like patterns.
Description
Anima v0.0.0(Initial)
After performing automatic tagging, I manually corrected several critical mistakes and trained this LoRA using the aforementioned 171 images.
Similar to the Krea 2 version, there is an issue where using the "art nouveau" tag tends to produce frequent circular or ring-shaped decorative patterns.
Also, this LoRA does not work well with the base model. Instead of producing a convincing Mucha-like style, it mainly results in a more "westernized" painterly appearance caused by style-related tags such as "masterpiece" and "score".
For better results, please use the FT model instead of the base model.
Interestingly, when used with the FT model, the LoRA seems to "remember" the training data and reproduce the learned characteristics much more effectively.
dim16/alpha16, unet 5e-05 (adamw8bit / cosine), 1024res, 171 images / 2000 steps