🎨 ChatGPT Web Gen

Generate images via ChatGPT Web — no API key, just a Plus/Pro/Team account.

---

What is this?

A single-file CLI tool that drives ChatGPT Web through a headless browser to generate images. No API key, no server, no daemon — just your ChatGPT subscription and a terminal.

Features

Capability	Description
No API key	Uses your ChatGPT Plus/Pro/Team subscription directly
Text → Image	Generate from any text prompt
Image → Image	Reference-based generation with --ref
Stealth browser	CloakBrowser + channel override bypasses bot detection
Encrypted session	Cookies encrypted via Fernet (AES-128-CBC + HMAC), key in OS keyring
Zero infra	Single Python file, no servers, no databases

How it works

                    ┌──────────────────────┐
                    │    gen.py (CLI)      │
                    │                      │
  prompt ──────────>│  Playwright browser  │──> chat.openai.com
  --ref ───────────>│  (headless Chromium) │
                    │                      │<── generated image
                    └──────────────────────┘
                               │
                               ├── cookies.enc ── encrypted session
                               └── temp/ ──────── output images

Every invocation starts fresh — no background daemon, no state management. Cold start ~8s.

Quick start

# Install
pip install -r requirements.txt
playwright install chromium

# One-time login (opens visible Chrome)
python gen.py --login

# Text-to-image
python gen.py "a cinematic cyberpunk city at night"
python gen.py "大稻埕碼頭黃昏，夕陽染紅天空，寫實攝影"

# Image-to-image (with reference)
python gen.py "put this product on a beach" --ref product.png
python gen.py "turn this sketch into a realistic oil painting" --ref sketch.png

Project structure

chatgpt-web-gen/
├── gen.py                  # Single-file CLI entry point (~240 lines)
├── requirements.txt        # cloakbrowser + playwright + keyring + cryptography
├── README.md               # English (you are here)
├── README.zh-TW.md         # Traditional Chinese
├── docs/
│   ├── zh-Hant/README.md   # Traditional Chinese (full)
│   └── ja/README.md        # Japanese
├── LICENSE                 # MIT
└── .gitignore

Runtime-only (excluded from Git):

Path	Contents
cookies.enc	Encrypted session cookies
playwright_user_data/	Chrome profile cache (login flow)
temp/	Generated images

Security model

Layer	Mechanism
Cookie encryption	Fernet (AES-128-CBC + HMAC SHA256) via cryptography
Key storage	OS keyring (macOS Keychain / Windows Credential Manager / Linux Secret Service)
Login browser	System Chrome with channel="chrome" to bypass Google bot detection
.gitignore	All session data (cookies.enc, playwright_user_data/) excluded from Git

Attackers need both the encrypted file and OS keyring access to steal a session.

Selector maintenance (when ChatGPT UI changes)

ChatGPT's frontend updates frequently. If generation breaks, update the S dictionary at the top of gen.py:

Key	Purpose	Common selectors
chat_input	Prompt textarea	#prompt-textarea, div[contenteditable="true"]
send	Send button	button[data-testid="send-button"]
generated	Generated image detection	img[alt*="已產生"]
streaming	In-progress indicator	button[data-testid="stop-button"]
file_input	File upload input	input[type="file"]
assistant	Assistant reply (error text)	[data-message-author-role="assistant"]
logged_in	Login status check	[data-testid="earth-icon"], #prompt-textarea

Debug tip: Set headless=False in gen.py to watch the browser in action.

Limitations

• Paid ChatGPT account required (Plus/Pro/Team)
• CLI only — no server, no API, no queue
• Single-threaded — one image at a time (ChatGPT Web constraint)
• Cookie-dependent — session expires every 1-2 months
• UI-fragile — ChatGPT frontend changes can break selectors
• Cold start — ~8s overhead per invocation

Related projects

Project	Description
chatgpt-image-bot 🤖	Telegram Bot version — queue, admin panel, image editing, visitor auth. Same engine, multi-user friendly.

License

MIT — see LICENSE.