atelier

Let your LLM paint.

LLMs can write, reason and code, but when you ask one for a picture it hands the job to an image model. atelier gives the LLM itself the brush: it writes a short Python script, renders it, looks at the result and fixes it, until the image is right. No image model is ever called: no diffusion, no image API, no key, no GPU. Every pixel comes from code the LLM wrote, computed on the CPU.

That covers any style: photoreal 3D, oil and watercolour painting, anime-style painted landscapes, comics and engravings, people, landscapes, posters and typography, pixel art, generative abstracts, and new styles the agent works out itself from a reference, and the same scripts can loop into GIFs. Under the hood: noise fields, vector shapes, a brush engine, effects, an SDF path tracer, terrain erosion, a physical sky, posable figures, and tools that let the LLM inspect its own renders.

Website: https://atelier.wexop.eu · every image below was made by atelier from a script in examples/.

Install

pip install git+https://github.com/Wexop/atelier

Python ≥ 3.11. Dependencies: numpy, pillow, opencv-python-headless, scipy, skia-python.

Install the skill (for AI agents)

The skill is a folder: SKILL.md, a compact cheat sheet that teaches an agent the whole API and the render → look → fix loop, plus style guides the agent reads only when it needs them:

guide	for
anime	painted anime landscapes: cel-shaded clouds, trees and mountains, rim light, god rays, sparkles
oil	oil painting from a colour study, watercolour washes
photoreal	path-traced still life and products, terrain landscapes, sea and sky
comic	ligne claire and toon from 3D, engraving, ink filters
flat	flat vector, posters, typography, people, silhouette scenes
pixel	pixel art with palettes and dithering
abstract	generative abstracts from warped fields
planning	process for complex pictures: written plan + focus, grey thumbnails scored side by side, build passes, hierarchy rules (example: planned.py)
animation	motion: the script reads T, keyframes and easing, seamless loops, GIF / WebP / MP4, motion check sheet (examples: lava.py, daynight.py)
new-style	any look not listed: study it, map it to the toolbox, paint it, then save it as a new guide

If you are an AI agent: run pip install git+https://github.com/Wexop/atelier, then read https://raw.githubusercontent.com/Wexop/atelier/main/SKILL.md and follow it (it links to the style guides).

Claude Code plugin (recommended, gets updates):

/plugin marketplace add Wexop/atelier
/plugin install atelier@atelier

Claude Code (personal skill, available in every project):

git clone --depth 1 https://github.com/Wexop/atelier /tmp/atelier && mkdir -p ~/.claude/skills && rm -rf ~/.claude/skills/atelier && cp -r /tmp/atelier/.claude/skills/atelier ~/.claude/skills/ && rm -rf /tmp/atelier

Windows PowerShell:

$t = "$env:TEMP\atelier"; git clone --depth 1 https://github.com/Wexop/atelier $t; New-Item -ItemType Directory -Force "$HOME\.claude\skills" | Out-Null; Remove-Item -Recurse -Force "$HOME\.claude\skills\atelier" -ErrorAction SilentlyContinue; Copy-Item -Recurse "$t\.claude\skills\atelier" "$HOME\.claude\skills\"; Remove-Item -Recurse -Force $t

For a single project, copy the folder to .claude/skills/atelier/ instead (this repo already has it). Without the styles/ folder the skill still works: it fetches the guides from GitHub when it needs one.

Claude.ai / Claude desktop: zip the .claude/skills/atelier folder (SKILL.md + styles/) and upload it in Settings → Capabilities → Skills. The skill installs the package itself with the pip install line above.

Updating: nothing updates automatically. Package: pip install -U git+https://github.com/Wexop/atelier (the version is bumped on every package change). Skill: re-run the git clone line above, or re-upload the zip on Claude.ai.

Quick start

# night.py
from atelier import *

c = Canvas(1600, 1000)
c.fill(gradient([(0, "#0b1026"), (0.45, "#3b2a5a"), (0.68, "#f4a259")]))
c.layer(blend="screen").fill("white", mask=stars(70, seed=3) * top(0.5, soft=0.2))
c.circle(1180, 250, 44, fill="#fff4d6")
c.layer().fill_below(ridged(2.2, seed=1).ridge_line(y=0.74, amp=0.15), fbm(6).ramp(["#0c0c1a", "#1f1c34"]))
c.apply(bloom(0.8, 12), grain(0.03), vignette(0.3))
c.save("night.png")

atelier render night.py --preview          # fast 512 px look -> night_prev.jpg
atelier inspect night.py                   # contact sheet: value study, clipping, auto zooms on defects, metrics
atelier render night.py --crop 600,500,300,200   # full-res zoom on a region
atelier render night.py --grid --boxes --mark 880,230  # coordinate grid, layer bounds, target mark
atelier critique night.py                  # ranked problems + fixes: contrast, balance, tangents, placeholders, style mix, floating bases, near-miss seams, ambient light, grounding, hierarchy
atelier sheet a.py b.py c.py               # thumbnails / variants side by side with value studies and scores
atelier render night.py                    # final render
atelier animate night.py --frames 24 --check   # GIF loop (T = 0..1 in the script) + motion check sheet

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗