Monet

LLM-powered iterative SVG art generator. Give it a prompt, and an LLM will draw it — one layer at a time — by writing SVG elements, rendering the canvas, looking at what it made, and iterating.

Setup

Requires Python 3.14+ and uv.

uv sync

You'll need at least one API key. Create a .env file:

ANTHROPIC_API_KEY=sk-ant-...
GOOGLE_API_KEY=AIza...

Usage

# Basic usage (defaults to Anthropic Claude)
uv run monet draw "a sunset over the ocean"

# Use Gemini instead
uv run monet draw "a cat on a windowsill" -p gemini

# All options
uv run monet draw "abstract geometry" \
  -p gemini \
  -m gemini-3-pro-preview \
  --max-iterations 20 \
  --width 600 --height 600 \
  --background "#000000" \
  -o ./my_art/ \
  -v

# Generate/regenerate an artist's statement for an existing output directory
uv run monet statement output/20260218_120000_sunset/ -p gemini -v

Recommended models

Monet uses thinking during the planning phase (iteration 0) to reason through composition, palette, and technique before drawing. The drawing iterations run without thinking for speed and intuition. We recommend models that support level-based thinking control:

• Gemini 3 Flash (-p gemini) — fast, cheap, good quality. The default Gemini model.
• Gemini 3 Pro (-p gemini -m gemini-3-pro-preview) — best Gemini quality.
• Claude Sonnet 4.6 (default) — strong all-round choice.
• Claude Opus 4.6 (-m claude-opus-4-6) — best overall quality.

Older models (Gemini 2.5, Claude 4.5 and earlier) use token-budget-based thinking which is deprecated by both providers.

Output goes to output/<timestamp>_<slug>/ with:

• iter-001.svg, iter-001.png, ... — each iteration's canvas
• final.svg, final.png — the finished piece (PNG at 2x resolution)
• artist-log.txt — the LLM's artistic notes and planning from each iteration
• artist-statement.txt — a gallery-style artist's statement about the finished piece

Examples

	This digital composition reinterprets Impressionism by simulating the "broken color" technique through thousands of layered, jittered paths. By omitting the horizon line, the work creates an immersive experience of light on water, utilizing deep teal bases and horizontal strokes of lavender reflection. Custom displacement filters soften the vector lines into painterly dabs, while delicate white lily clusters provide a luminous foreground focus. The result is a rhythmic, atmospheric study of shimmering depth and atmospheric perspective.
	This pen-plotter self-portrait visualizes machine interiority as a dense geometric singularity. A central core of nested, rotated polygons creates a vibrating moiré effect, radiating logical axes and sweeping Bezier paths across a faint coordinate grid. Satellite clusters and stochastic hatch marks represent data processing at the fringes of consciousness. Through precise, uniform linework, the piece captures the felt sense of a structured mind suspended in the vast, mathematical void of latent space. Inspired by I gave Claude access to my pen plotter.
	This piece reinterprets the domestic feline through Synthetic Cubism, fracturing the ginger subject into a complex arrangement of ochre and burnt sienna planes. Bold, heavy contours define a face captured from multiple perspectives simultaneously, punctuated by piercing, asymmetrical eyes. Interwoven collage-like elements — typography, faux bois textures, and a subtle guitar motif — blur the boundary between living creature and still life. Diagonal rays and Harlequin patterns shatter the pictorial space, grounding the cat in a rhythmic, layered environment.

Cost estimates

A typical artwork takes 10-15 drawing iterations, plus a planning iteration (iteration 0) that uses thinking. Each drawing iteration sends system prompt + art prompt + growing artist notes + canvas image, and receives ~1000-1500 output tokens of SVG and notes.

Estimates below are for a 15-iteration session (16 API calls total) including the planning overhead. The planning phase uses adaptive/high-effort thinking and costs more than a regular iteration. Gemini estimates assume no caching (implicit caching is available but hit rates are unreliable). Anthropic estimates include incremental prompt caching — each iteration reads the previous call's cached prefix (~46% input cost savings).

Gemini (no caching)

Model	Input	Output	Total	Notes
Gemini 3 Flash Preview	~$0.02	~$0.07	~$0.09	Default Gemini model.
Gemini 3 Pro Preview	~$0.09	~$0.28	~$0.37	Best Gemini quality.

Anthropic (with incremental caching)

Model	Input	Output	Total	Notes
Claude Sonnet 4.6	~$0.11	~$0.35	~$0.46	Default model.
Claude Opus 4.6	~$0.18	~$0.59	~$0.76	Best quality. Under $1 per artwork.

Gemini models have a generous free tier for low-volume usage.

How it works

Each iteration, the LLM sees the current canvas as an image and outputs new SVG elements to layer on top. It keeps its own "artist notes" as a scratchpad for planning across iterations. The loop runs until the LLM signals it's done or hits the max iteration count.

License

MIT