Skip to main content

invertornot.py

Path: build/invertornot.py | Language: Python | Lines: ~96

Experimental GPT-4-Vision script for classifying images for dark-mode inversion. Now obsoleted by InvertOrNot.com.

Overview​

invertornot.py was a prototype solution to one of the hardest problems in dark-mode implementations: deciding whether images should be inverted/negated or just dimmed. A bright white diagram may look fine inverted, but inverting a grayscale photograph of a human face looks terrible. No simple heuristic works—it's truly an "I know it when I see it" visual classification task.

The script used GPT-4-Vision's multi-image capability: it takes the original image, generates an inverted version using PIL's ImageOps.invert(), sends both to GPT-4o-mini, and asks "Does the inverted image look OK for dark-mode? YES/NO." This zero-shot comparison approach worked reasonably well, as the model could directly see both versions.

Why it's obsolete: The cost was prohibitive (~2¢ per image, or $440 to classify all Wikipedia thumbnails on gwern.net). Mattis Megevand created InvertOrNot.com, a free API using a small finetuned EfficientNet running on CPU, with higher accuracy, user feedback for continuous improvement, and zero per-query cost. The script remains as a reference implementation for similar vision tasks.

Key Functions​

  • encode_image(image_path): Converts image to base64 for API upload
  • invert_image(image_path): Creates negated version using PIL ImageOps
  • API request: Sends both images to GPT-4o-mini with comparison prompt

Command Line Usage​

# Classify a single image
OPENAI_API_KEY="sk-XXX" python invertornot.py

Note: the script currently ignores CLI arguments and uses a hardcoded `image_path` inside the file.
# Output: JSON response with YES/NO judgment

# Example output structure:
# {
#   "choices": [{
#     "message": {
#       "content": "NO"  # or "YES"
#     }
#   }]
# }

Note: The script has a hardcoded test image path on line 59. In practice, you'd replace this with sys.argv[1] or similar to accept arbitrary inputs.

Prompt strategy:

  • Sends original image first, inverted second
  • Asks: "This is the inverted image... Does it look OK? Respond YES/NO."
  • max_tokens: 300 (minimal, as response is just YES/NO)

See Also​