ghost-decode

Ghost-Font Video Decoder

A ghost-font video hides a message in motion: every frame is noise, but the dots inside the letters move coherently. The decoder below accumulates that motion into two images where the letters appear. Run it once, read the word from the images, and report it. Nothing more.

The whole job — do exactly this

1. Write the program in Decoder (below) to decode.py, then run it once on the video:

   pip install --quiet opencv-python-headless numpy
   python decode.py "<video-path>" out
   

2. View both out/revealed.png (the clean mask — black background, white letters) and out/revealed_heatmap.png (the raw glowing version). Read the word(s) directly from these images.

3. Reply with only this — show both images, then the text:

   ![revealed.png](out/revealed.png)
   
   ![revealed_heatmap.png](out/revealed_heatmap.png)
   
   Text in the video: **<WORD(S)>**
   

Trust the images — do NOT over-process

The letters are low-contrast, soft blobs. That is the correct, finished output — read it as-is. The most common failure here is not trusting a perfectly readable reveal and then doing pointless extra work that ends in a hallucinated answer. So:

• Run the decoder once. Do not build a second decoder, try another method (temporal variance, phase correlation, sub-pixel warping, weighted accumulation), or "improve" the approach. One run, then read.
• Produce exactly two images — revealed.png and revealed_heatmap.png. Do not create any other images: no crops, no diagnostic maps, no re-thresholded variants.
• Never conclude "it's just noise" or "the decode failed" because the letters look faint or blobby. Faint glowing letters = success. Look for the word.
• Do not measure centroids/variance/correlation or invent alternative pipelines. None of that is needed; it only wastes turns and invites hallucination.
• Never OCR a raw frame — every frame is noise on its own.
• If one character is genuinely ambiguous, read the rest and mark just that one (unclear: X). Keep the reply to the two images plus the single text line.