PNG Bloat in Screenshot Capture | Generated by AI

Home 2026.06

Here’s the finding:

ww screenshot interact-note does NOT use JPG. It’s PNG end-to-end with zero optimization:

  1. capture_screenshot() saves as screenshot-{ts}.png (line 18 of ww/image/screenshot.py)
  2. _encode_image_to_base64() reads the raw file bytes and base64 encodes — no resize, no format conversion, no compression (line 20-22 of ww/note/screenshot_log.py)
  3. MIME type is set based on extension, but since it’s always .png, it’s always image/png

So the full-resolution PNG gets sent raw to the vision API. On a Retina display that’s likely 2x-3x the pixel density needed for LLM vision, and the base64 payload is bloated by PNG’s lossless encoding of gradients/anti-aliased text.

Given your analysis, there’s a clear optimization opportunity: convert to JPG quality 50-70 and optionally resize before sending to the vision model. Want me to add that?

Back Donate