Anthropic / algorithmic-art
Can Claude make generative art you can reproduce?
Yes, and the seed is the part that matters.
What to remember
- Seeded randomness is the feature. It turns a lucky output into a repeatable one.
- Parameters are more useful than a finished image, because they let you explore.
- It is a system generator, not an illustrator.
- Original systems are safer than imitating a named living artist's signature style.
01What it changes in practice
Asking a model for art usually produces one result you cannot vary. Asking for a parametric system produces a space you can move around in, which is what artists actually want.
Because the output is code, it is also editable, versionable and cheap to rerun at another size, which matters when the piece has to work as a background at several resolutions.
02Where it disappoints
It is not the tool for figurative work. If you need a specific illustration of a specific thing, an image model is the right instrument.
Aesthetic judgement remains yours. The skill will happily generate a hundred variations of something mediocre.
Install and first run
- Install it from anthropics/skills under Apache 2.0.
- Ask for a system rather than a picture, for example a flow field with two parameters you can vary.
- Record the seed of anything you like, because without it the piece is not reproducible.
Questions people actually ask
- What does seeded randomness mean here?
- The same seed produces the same artwork, so a variation you like can be regenerated and adjusted rather than lost.
- Can it copy a specific artist's style?
- Building original systems is the better use, and it avoids copying the signature style of a living artist.
- Is the output editable?
- Yes, it is code, so it can be versioned, tuned and rerun at any size.
Sources
Machine readable versions of this page