MAI-Image

Free image generation via Bing Image Creator — no API key, no billing, works with or without cookies.

An agent skill plus a standalone Python CLI that talks directly to Bing's own image-creator backend (the one behind bing.com/images/create). It runs in two modes:

• fresh — no cookies at all. Anonymous sessions each carry a small quota and are rotated automatically when they run dry.
• cookie — your signed-in bing.com account, using its daily image tokens.

Both produce the same images from the same model. The script is a single file with no dependency on any SDK; curl_cffi is recommended, requests is the fallback.

pip install curl_cffi pillow
python scripts/mai_image.py "a lighthouse on a cliff at dusk" --fresh

---

What it does

Default model	mai-image-2.5-flash — what Bing actually serves today
Images per request	4 (text-to-image), 1 when using --ref
Aspect ratios	1:1, 3:2, 2:3 (DALL-E 3 ids also accept 7:4, 4:7)
Reference image	yes — exactly one per request
Fresh session budget	~5 requests ≈ 20 images, then auto-rotate
Output	1024×1024, 1248×832, or 832×1248 JPG/PNG

Install

As an agent skill — drop the folder into your skills directory:

~/.agents/skills/mai-image/      (or wherever your harness loads skills from)

The SKILL.md frontmatter is what makes it discoverable.

As a plain CLI — copy scripts/mai_image.py anywhere and run it with Python 3.9+:

pip install curl_cffi pillow
python mai_image.py "a paper boat on a puddle" --fresh

Usage

# fresh mode: no cookies needed, rotates sessions when quota runs out
python scripts/mai_image.py "a brass lantern on a wet ledge" --fresh

# cookie mode: uses your account's daily tokens
python scripts/mai_image.py "a brass lantern on a wet ledge" --cookie

# let it ask you
python scripts/mai_image.py "a brass lantern on a wet ledge" --ask

# batch -- each request returns 4 images, so this yields 24
python scripts/mai_image.py "neon ramen shop in the rain" --fresh -n 6 -o renders

# landscape, and just print URLs
python scripts/mai_image.py "wide desert highway at noon" -a 3:2 --url-only

# image-to-image with one reference
python scripts/mai_image.py "restyle this as polished chrome" --fresh -r jug.jpg

# check what the account state looks like
python scripts/mai_image.py --check --cookie

Flag	Meaning
--fresh / --cookie / --ask	auth mode (mutually exclusive; asks if none given)
-m, --model	mai-image-2.5-flash (default), mai-image-2, mai-image-1, dalle3, gpt4o
-a, --aspect	1:1 (default), 3:2, 2:3
-s, --size	size alias, e.g. 1024x1536 (overrides -a)
-n, --batches	number of requests
-o, --out	output folder (default ./output)
-r, --ref	one reference image for image-to-image
-t, --timeout	per-request poll timeout (default 300 s)
--json	machine-readable result on stdout
--url-only	skip downloading
--check	report auth/quota state and exit
--max-rotations	cap on fresh-session rotations per run

Cookies (cookie mode only)

Export cookies from a signed-in bing.com tab. The file must include _U and SUID, otherwise Bing treats you as an anonymous visitor.

Accepted formats:

[{ "name": "_U", "value": "...", "domain": ".bing.com", "path": "/" }]

{ "_U": "...", "SUID": "..." }

Lookup order: --cookies PATH → $MAI_COOKIES_JSON → <skill root>/cookies.json → ./cookies.json. Never commit this file — .gitignore already covers it.

Notes from reverse engineering

These are the findings that shaped the script; they are documented because the endpoint is undocumented and changes without notice.

• mdl is currently ignored. Every value — 0, 1, 4, 7, 10, 12, even a bogus 999 — renders on MAI-Image-2.5-Flash with 4 images. Model choice is server-side. The script still sends the documented id and always prints the name the server reports back.
• One reference image only. The page's file input has no multiple; posting two imageBase64 parts returns a single bcid matching the first image (the second is silently dropped), and two bcid params in submit are rejected outright. A second --ref is refused locally rather than silently ignored.
• A reference is genuinely honored. A terracotta jug with lavender restyled to "polished chrome metal" kept the jug silhouette, the lavender arrangement, and the table; the same prompt without --ref produced an unrelated chrome form.
• Results may arrive inline. When data-status="2", the images are already in the data-results attribute and data-polling-url is empty. Parse attributes before HTML-unescaping, or the embedded JSON breaks.
• Cold sessions refuse. Posting to a session that never fetched the creator page returns requestmissed; a warm-up GET also seeds _SS (needed for uploads).
• IP-level throttling. Rotating many sessions back-to-back starts refusing everything for a minute. Pacing plus a small retry budget handles it.

Endpoints

GET  /images/create/ai-image-generator            warm-up: seeds MUID / _SS

POST /images/create/ai-image-generator
       ?FORM=GENCRE&q=<prompt>&ctype=image&mdl=<id>&ar=<1-5>&rt=<3|4>
       [&bcid=<upload id>&edt=1]
  -> HTML; read <aigc-results-view> data-status / data-request-id /
     data-polling-url / data-results
     (data-status: 0|1 working, 2 done, 3 refused)

POST /images/create/upload?&sid=<SID>
       multipart imageBase64 = <base64 jpeg>    -> {"bcid":"r.B.XXXX","error":null}

GET  /images/create/ai-image-generator/async/results/<requestId>
  -> {"status":0|1} … {"status":2,"records":[{"mediaItems":[{"src":...}]}]}

ar: 1:1→1, 7:4→2, 4:7→3, 3:2→4, 2:3→5. rt: 3 anonymous, 4 with paid tokens.

Disclaimer

Unofficial and reverse-engineered. Not affiliated with Microsoft or Bing. This is for personal and educational use; you are responsible for complying with Bing's terms of service, and the endpoint may change or block you at any time.

License

MIT — see LICENSE.