A couple of friends and I are A Fret of Nothing, a guitar, mandolin, and mandola trio. We get together a few times a month to play mostly classical music. This year we signed up for open mic at the Classical Mandolin Society of America’s annual convention, which was in Kansas City.
We performed Antônio Carlos Jobim’s “Canta Mais” and Astor Piazzolla’s “Oblivion,” both kindly provided to us and arranged by Robert Margo.
This year the Classical Mandolin Society of America had its annual convention at Kalamazoo, Michigan. Some members of the Minnesota Mandolin Orchestra performed at the Saturday open-mic session. Michael Tognetti generously recorded and mastered the following videos of us playing John Goodin’s “Wedding March Suite” and Peter Ostroushko’s “Red Lodge Reel.”
For a few years now, I’ve been fascinated by programmatic art. For ten years or so, there’s been a stream of projects that do something along the lines of applying an artist-inspired filter (usually Van-Gogh) to existing photos to stylize them. Last year, the DALL-E project went a step further by generating images from textual descriptions and a library of images.
I was stunned at both variations on existing images and images synthesized from descriptions. This summer, generative art got a little more accessible to the masses when Stable Diffusion was released, using similar technology, but publicly available models (at least for now).
I think the business model leverages that you can make interesting images from the public models, but that most businesses will want to tune the image text-to-image models to suit their intended message. The models require an enormous amount of energy to train — the default one has an 11 metric-ton carbon footprint — and the model propagates cultural biases present in the underlying data. Both aspects seem significant enough to motivate an entrepreneur to hire out specialists.
Stable Diffusion is a little tricky to install on a M1 Mac — you have to checkout a particular branch of stable-diffusion, currently `apple-silicon-mps-support`, and hope that nightly PyTorch builds are stable. You also have to register with HuggingFace to download an image model, which seem to be updated frequently at this point.
There’s currently no GPU support for Apple silicon for Stable Diffusion, so you need to be patient, but here are some samples of the Stable Diffusion imagination…. that might sound a bit like “stable genius” and artistically sometimes the images resemble another product shoveled from a stable…. but again, samples of
