anon-image

anon-image

Generate and edit images through key-less public endpoints, then compose the results into sheets and comparisons. Pure Python stdlib + Pillow; no SDK, no key, no account, no browser.

When to use

The user wants…	Do this
"make me an image of X"	imgen.py gen "X" -o out.jpg
a specific aspect ratio / wallpaper / phone shot	add --ratio 16:9 or --ratio 9:16
"change this image to Y" / restyle / relight / repaint	imgen.py edit in.jpg "Y" -o out.jpg
N prompt variants at once	imgen.py batch prompts.txt --outdir shots/ --parallel 4
a labelled grid of candidates for review	`imgen.py sheet out.jpg a.jpg b.jpg --labels "A
before/after for a diff	imgen.py compare out.jpg before.jpg after.jpg

Quick start

imgen.py gen "a glass lighthouse on a basalt cliff at dusk, cinematic" -o lighthouse.jpg --ratio 16:9
imgen.py edit lighthouse.jpg "make it a bright sunny afternoon" -o lighthouse_day.jpg
imgen.py compare cmp.jpg lighthouse.jpg lighthouse_day.jpg
imgen.py sheet sheet.jpg a.jpg b.jpg c.jpg --labels "one|two|three" --cols 3 --title "candidates"

All commands print saved <path> (<size>, <time>) and exit 0. Add --json to gen/edit for a machine-readable line: {"ok":true,"out":…,"bytes":…}.

Prompting rules that actually matter

• Write one scene, concrete nouns, an explicit medium and light: "a brass compass on weathered wood, macro, side light" beats "beautiful compass".
• For edits, say what to change and what to keep: "repaint as oil, keep the composition and the lighthouse position". Unsaid things drift.
• The model is FLUX-class. It renders text badly; avoid asking for words inside the image, or expect to fix them afterwards.
• --ratio is honoured at generation time (1248 px on the long edge). Do not generate square and crop later unless you must — you lose composition.

Editing: the rules

• Input must be ≤ 3.5 MB. Resize first if not: ffmpeg -i big.jpg -vf scale=1536:-2 small.jpg
• The instruction is applied to the whole frame; it is not a mask/inpaint tool. For "remove this object" phrasing, name the object and what replaces its background: "remove the parked car, extend the cobblestone street".
• Edits preserve dimensions. Two edits of the same edit compound drift — prefer one well-written instruction over three small ones.

Interfaces

gen · batch · edit · sheet · compare · probe

probe returns 0 if the generation endpoint is alive, 2 if it has gone dark — run it before promising the user a picture.

Failure modes and what to do

Symptom	Meaning	Action
HTTP 502/504, or 60 s+ then timeout	upstream is down or overloaded	retry once after ~10 s; if it persists, say so and offer to assemble from existing images
input is 4.2MB … accepts 3.5MB	oversize upload	resize (command is printed for you)
no imageUrl in response	upstream returned a shape we don't know	print the raw JSON, do not silently retry in a loop
Generation takes 10–20 s	normal	do not lower --timeout below 120

Never hammer it: 10 parallel images is tested-safe and productive, more is pointless and rude. Two images per user turn, on average, is where this sits.

Notes for the agent

• Prefer generating 2–3 candidates and showing a contact sheet over generating one and apologising.
• Always save onto a path the user can open; never leave the asset in a temp dir.
• The image has no watermark and no C2PA manifest — but the upstream terms restrict commercial use of the free text-to-image tier. Editing (image-to-image) is the route that stays commercially usable. Check reference/endpoints.md before telling anyone "you can sell this".