make-art — a working AI-image engine

The engine is Nano Banana Pro = Gemini 3 Pro Image (gemini-3-pro-image-preview) via the Gemini REST API. A zero-dependency, stdlib-only wrapper lives at scripts/gen_art.py. It reads GEMINI_API_KEY from the environment, or from ./.env (override the path with GEMINI_ENV_FILE). Native output up to 4K, ~35s per image, ~$0.24 per 4K image. Billed, no free tier — a busy week of plates can run into three figures, so put a budget cap on the Google Cloud project. Use 2K for drafts and keep 4K for finals.

Delivery: a folder per job (e.g. art/<slug>/, finals numbered by rank, a README with prompts and cost). Keep generated art out of a personal photo library unless asked.

Two hard gotchas

1. Network must be allowed. If your agent runs shell commands in a sandbox that blocks DNS/network, the API call fails; run gen_art.py with network access.
2. Bake text in HTML/PIL, not the prompt. The model renders text unreliably. Generate a text-free image, then overlay type yourself (HTML->screenshot, or PIL) so it stays crisp. Always add "no text, no lettering" to prompts meant to carry an overlay.

Usage

python3 scripts/gen_art.py --prompt "..." --out plate.png --aspect 3:2 --size 4K
python3 scripts/gen_art.py --prompt-file p.txt --out plate.png --aspect 9:16 --size 4K
python3 scripts/gen_art.py --image base.png --prompt "..." --out out.png   # anchor/edit around a base image (repeatable --image)

• --aspect: 1:1 2:3 3:2 3:4 4:3 4:5 5:4 9:16 16:9 21:9. Phone wallpaper -> 9:16; magazine spread -> 3:2; single page -> 3:4.
• --size: 1K / 2K / 4K (uppercase). 4K for print/wallpaper; 2K to save cost on drafts.
• Returns a JPEG or PNG (check the printed mime). For wallpaper, cover-crop to the exact device size afterward.

When the base image is a paste, not a file

--image needs a path, but an image pasted into a Claude Code conversation is not on disk: it lives only as base64 in the session transcript. Extract it:

import json, base64, pathlib
p = pathlib.Path.home() / ".claude/projects/<slug>/<sessionId>.jsonl"   # slug = cwd with / -> -
for line in p.open():
    for blk in (json.loads(line).get("message") or {}).get("content") or []:
        if isinstance(blk, dict) and blk.get("type") == "image":
            src = blk["source"]
            pathlib.Path("base.jpg").write_bytes(base64.b64decode(src["data"]))

Look at the extracted file before spending a 4K gen on it. See claude-usage-mining for the transcript layout.

Prompt craft

• Name the medium + palette up front and keep it consistent across a set. Two styles that work well: "35mm film photograph, visible film grain, muted sun-faded palette of ochre and dusty blue, soft natural light, analog imperfection, contemplative, vertical." and, for print plates, bone paper, near-black ink, a single hot-orange accent, risograph texture, flat full-bleed.

• For overlay-friendly backgrounds: ask for deep negative space where the text will sit, dark/low-contrast zones, and a single light source / light-leak for interest.

• Vary a pool by shifting time-of-day, weather, motif (sea, dunes, fog, rain, night stars, double-exposure, abstract light field) while holding the palette.

• Full-bleed warning: the model likes to add a light "paper/riso frame" border (especially when you say "poster", "risograph", or "paper texture"). For edge-to-edge art, omit those words or trim the border afterward. Don't trim on a visual hunch — probe it. Measure the run inward from all four edges. A real frame is uniform on all four sides; wildly asymmetric runs (166px bottom on one plate, 720px right on another) are art — cloud, snow, sky — legitimately hitting the bleed, and cropping it destroys the composition. Probe luminance, not whiteness — the frame is not always pale. A near-white test is blind to the dark "slide mount" border the model also likes. Threshold on 0.299R+0.587G+0.114B at both ends (<42 and >238), sample three lines per edge and take the median so a patch of legitimate art-black doesn't fool one scanline, then apply the uniform-vs-asymmetric test:

  plate-A      DARK run  L132  R1264  T0   B848   asymmetric -> art (crushed shade)
  plate-B      DARK run  L97   R83    T84  B86    UNIFORM    -> real frame, crop it
  

  Crop the max plus clearance for corner rounding (150px worked), then restore the aspect ratio. TAIL negatives like "no border, no frame, no film strip, no sprocket holes, no white margin" do not reliably prevent the dark mount; the model seems to read those as being about pale borders only.

"One small accent" needs a named carrier

A STEM saying "one small safety-orange accent" can hold on four plates and fail on the fifth when the subject contains large repeated shapes: the engine paints every one of them orange. Name the carrier and forbid the rest: "the only orange in the whole image is a small stripe on the side of each robot; chutes, deck and openings are slate blue and warm grey." When a subject has big repeated geometry, "small accent" is not a constraint.

An edit pass cannot RESIZE a repeated object — rewrite the SUBJECT and roll

Asked via --image to "keep everything, only make every unit tall", the engine kept the flat units and ADDED tall towers beside them. Edit passes are reliable for recolour and remove; they are not reliable for changing the proportions of many instances or inserting a stage into a sequence. Put the corrected geometry into the SUBJECT and regenerate.

See also

image-riff — the TASK method on this engine: turning a reference image into a spread of original plates (invariant STEM / varying SUBJECT / invariant TAIL, vantages, honest ranking, destination crops). brand-direction-plates — several visual languages across one brand.