Artisan XL: Unparalleled Fine Art Generation
Artisan XL is a state-of-the-art text-to-image model that pushes the boundaries of AI-generated fine art. Trained on an extensive dataset of 200,000 high-resolution oil paintings, this model offers unparalleled style fidelity for a wide range of artistic styles throughout human history.
Key Features
Trained for 1200 A100 hours on 8 A100 GPUs
Captioned with Claude Sonnet 3.5 for precise style understanding
Exceptional style fidelity for artists like Picasso, Van Gogh, Rembrandt, Klimt, Repin, Serov, Kandinsky, and many more
Supports all basic SDXL tools and community techniques
Usage Tips
Prompt Engineering: We recommend using long, detailed prompts with weighting. Ideally, generate prompts using Claude Sonnet or Haiku for best results.
Negative Prompts: Use empty or very short negative prompts. Common negative keywords include:
simplified, rough, blurred, crude, imperfections, sketchSettings:
Works well with guidance limiter
Recommended CFG scale: 3.0-5.5
DPM++ 2M with Karras sigmas or UniPC with Karras sigmas
Model Variants
Full Model: Complete SDXL fine-tuned model
LoRA Variant:
Linear and Conv modules
Rank 128
Provides 95% of full model capabilities
Multi Aspect Resolution
This model supports generating images at the following dimensions
Acknowledgements
The development and release of Artisan XL would not have been possible without the invaluable contributions and support from the following individuals and organizations:
kohya-ss: For training scripts.
Limitations
While Artisan XL offers exceptional capabilities in fine art generation, it's important to note:
The model specializes in fine art styles and may not perform as well for other genres.
Results can vary based on prompt quality and settings.
As with all AI models, occasional unexpected outputs may occur.
Contact
For any inquiries, collaborations, or custom training requests, please contact:
We're excited to see what you create with Artisan XL! Happy art-making!
Description
Full finetune
FAQ
Comments (9)
Interesting model! Question: how does your LoRa relate to the basemodel? Is LoRa enough to use the trained knowledge for image generation?
LoRA was extracted from base model using SVD approximation, and the output is almost exactly the same, so it should be enough!
https://x.com/GrigoryEvko/status/1833938118749065472
@ArtisanLabs Oh, I see = LoRa is also very much great for testing
is there a dataset of artist names used in training cause i dont know alot of classical era painters except for van gogh and rembrandt. this should be a cool way to discover artists i didnt know about, thanks
Unfortunately, we can't share a dataset, but you can use various online resources such as https://artsandculture.google.com/ , https://gallerix.org/ and https://www.wikiart.org/ to learn more about different painters and art in general!
Van Gogh was Post-Impressionist, not classical era.
Rembrandt was Baroque, not classical.
What is generically known as classical style is really lots of smaller eras; only one of them is technically known as classical. (I'm 100% sure this is true in music, and 99% sure regarding painting.)
A starting point might be Renaissance era style.
It would be interesting to find out if this LoRA is a mashup of multiple eras, covering 500 years, or sticks faithfully to/within one.
@davy272 usually it all depends on prompt, model itself learned the concepts and styles, but sometimes with suboptimal prompt you end up with only 90% style match.
Also this is full finetune, not a LoRA training (LoRA was extracted from the full checkpoint), so much higher training capacity.
a wildcard would be helpful
Awesome art🤩