modello*

Modello* is an artist tool that I created and it can be downloaded from the Apple App Store. It creates generative images based on any image that a user provides to it.

The result is rendered to look like an oil painting in my style and these can be tuned to closely match the source image or be allowed to go off on a wild tangent. This was made possible by training a diffusion model on hundreds of high-res scans of my own artwork.

Concept

The name comes from an italian preparatory study called a modello, usually done at a small scale as a proposal for a larger painting.

The asterisk * is because it’s like a small study, but not really, it’s inspired by the input, but rendered by the algorithm. The resulting images often have unexpected combinations and plenty of errors.

I wanted it to work this way, to just come up with ideas for starting a painting, not a finished work. Existing image generators focus on polish, and are frustrating to use because they often try to imitate other artists. By training on my own paintings I found I could get it to create things that I might actually like to paint.

Privacy & Power Usage

Does this app upload a users image to train AI? No.

Modello* runs entirely on the device, all images are stored only on the iPhone or iPad where it’s installed. Read more in the privacy policy.

The app never logs into a data center or the ‘cloud’ for any reason. It runs entirely off-line as it does when running my installation Typologies. Power comes from the battery of the device.

While programming and training the model, I did use cloud computing power for the processes that required massive computation. That was a one-time use of data center computing during development and is not ongoing.

Source Data

The source material is my own artwork, not a scraped image dataset, and the app does not use artist names as prompts.

That does not mean modello* is outside the larger history of generative AI. The app still depends on open-source diffusion tools, model architectures, and licensing frameworks developed by others. The app’s visual direction and training images are mine, while the underlying diffusion technology comes from a broader open-source AI ecosystem.

The app includes or builds on the following model components:

• Apple Core ML tools for running Stable Diffusion on device
• Stable Diffusion 1.5
• DreamShaper 8 base model
• Three custom LoRA models trained from my artwork: style, motif, and texture

License information for the model components and related open-source software is included in the app under Notes → Licenses.

Download link: modello*