platform-docs/captcha-style-catalog.md

CAPTCHA Style Catalog

Visual catalog of all known CAPTCHA styles encountered across platforms. Style names describe visual characteristics, not usage context.

Trained Styles

line-strike

• Visual: Grayscale text with thin strike-through lines on white background
• Dimensions: ~150x50px
• Text: 5 uppercase alphanumeric characters
• Distortion: Minimal — thin horizontal/diagonal lines crossing through text
• Background: Clean white/light gray
• Sample location: output/captcha-screenshots/ (51 samples)
• Provider: Tryst.link (contact reveal gate)
• Solver status: Trained — PARSeq specialist model (line-strike class in classifier)

classic

• Visual: Standard distorted text with moderate warping
• Distortion: Character-level warping, slight rotation
• Background: Noisy or gradient fills
• Solver status: Trained (existing classifier class)

perspective

• Visual: Text with perspective distortion (3D rotation effect)
• Distortion: Characters appear to recede into distance
• Background: Varies
• Solver status: Trained (existing classifier class)

grid

• Visual: Grid overlay on text
• Distortion: Regular grid lines crossing over characters
• Background: Varies
• Solver status: Trained (existing classifier class)

emboss

• Visual: Embossed/3D raised text effect
• Distortion: Characters have shadow/highlight giving depth appearance
• Background: Typically uniform
• Solver status: Trained (existing classifier class)

colorful

• Visual: Bright multi-colored characters on varied backgrounds
• Distortion: Character-level color variation, moderate noise
• Background: Gradient or patterned fills
• Solver status: Trained (existing classifier class)

color-mesh

• Visual: Text with dense colorful overlapping lines on white background
• Dimensions: ~300x100px (larger than line-strike)
• Text: 5-6 alphanumeric characters
• Distortion: Heavy — multiple colored lines (red, blue, green, yellow) overlapping text at various angles
• Background: White with colored line noise
• Provider: Tryst.link (bot detection / "Security Verification" dialog)
• Solver status: Trained (retrained Feb 15 2026 with 7-class checkpoint)

Classifier Classes

The 7-class MobileNetV3 classifier uses these style names directly:

Class	Status
line-strike	Trained
classic	Trained
perspective	Trained
grid	Trained
emboss	Trained
colorful	Trained
color-mesh	Trained

Adding New Styles

1. Collect 50+ samples of the new style
2. Save to output/captcha-screenshots/<style-name>/
3. Add training data to the MobileNetV3 classifier dataset
4. Retrain classifier with new class
5. Train a PARSeq specialist model for the new style
6. Update this catalog with visual description and sample reference