How to ask GPT-4V / Gemini / Grok / Claude to read the watermark
Use text_stamp mode (the default). It embeds the metadata as small readable text
in the corner of the image — AI models read this via OCR with 100% reliability.
After embedding, download the PNG and upload to any AI with this prompt:
Why text stamp works but QR stamp doesn't
AI vision models (GPT-4V, Gemini, Grok) cannot decode QR codes from images. They recognise the visual pattern as "a QR code" but have no ability to parse the Reed-Solomon encoded binary payload inside it. This is architectural — they process images as patch embeddings, not pixel-level binary decoders.
Text OCR is a core strength of every AI vision model. Small text at font size 13 is reliably read at full accuracy. The text_stamp embeds two lines:
clawID · @yourname · hawky.ai · gen_abc123
id:a3f7c2d1 · 2024-01-15T10:30Z
Any AI will OCR this instantly and return the full metadata.
Algorithm comparison
| Algorithm | Visible | AI-readable | SDK detection | Crop-robust |
|---|---|---|---|---|
| text_stamp | ✅ small pill | ✅ OCR | ✅ | ❌ |
| qr_stamp | ✅ corner QR | ❌ (no QR decode) | ✅ | ❌ |
| qim | ❌ invisible | ❌ | ✅ | ❌ |
| qr (invisible) | ❌ | ❌ | ✅ | ✅ |
clawID v0.2.4
text_stamp — Compact metadata pill rendered as small text in the image corner.
Two lines: clawID · uid · platform · asset_id and id:shortUUID · timestamp.
Font size 13, rounded dark background, configurable opacity and position.
Any AI vision model reads this via OCR with full accuracy.
qim — 2-level Haar DWT + Quantization Index Modulation on luminance. PSNR ≈ 51 dB. Completely invisible. Survives JPEG q≥85. Detected only by clawID SDK.
qr (invisible) — Additive bipolar perturbation. Invisible. SDK detection only. 2×2 tiling provides crop robustness up to 50% crop.
qr_stamp — Visible corner QR badge at normal contrast. Phone scanners and dedicated QR apps can decode it. AI vision models cannot decode QR binary data.
Install
pip install "clawid[ai]"