Technology
Someone stole your mesh, retextured it, and re-uploaded it as their own. Shield finds it — and verifies it — even after rotation, scaling, retopology, or texture swaps.
Existing tools compare thumbnails.
We fingerprint and verify geometry.
Live Detection Spectrum
Evasion Resistance
Shield fingerprints the actual 3D geometry. These evasion techniques have no effect.
Different paint, same shape. Shield sees through it.
Fewer polygons, same distance distribution.
New edge flow, same spectral signature.
Every fingerprint is transformation-invariant.
Mirror images produce identical fingerprints.
Component stripping catches geometry additions.
Fingerprints match across all accounts and listings.
The mesh is the fingerprint, not the metadata.
Detection Layers
Each layer targets a different evasion technique. Together, they identify clone candidates that move to geometric verification.
Layer 01
Re-colored copies · resized thumbnails
Two independent perceptual hashes capture visual appearance at the frequency level. Survives color shifts, crops, and resolution changes.
Layer 02
Direct mesh rips · unmodified geometry
Vertex count, triangle count, bounding box ratios, surface area. Parsed directly from Roblox’s binary mesh format. Eliminates 90% of non-matches instantly.
Layer 03
Rotation · scaling · mirroring
Surface-area-weighted sampling creates a shape signature that’s mathematically invariant to rotation, scale, and translation. Enhanced with angle distribution analysis.
Layer 04
Texture theft · material swaps
Frequency analysis of raw texture data catches “same shape, different skin” and “different shape, same skin” attacks independently.
Layer 05
Retopology · remeshing · subdivision
Analyzes intrinsic geometry that survives complete mesh rebuilds. Two meshes with the same shape produce nearly identical signatures even with different polygon structure.
Layer 06
Smoothing · faceting · detail changes
Captures how the surface bends at every vertex. Distinguishes items that share a silhouette but differ in surface detail.
Layer 07
Shape differentiation · orientation shifts
Surface orientation distribution — captures how mesh faces are directed in space. Discriminates compact objects that share similar size but have different shapes.
Geometric Verification
After the seven detection layers identify candidates, Shield performs a direct geometric comparison on the actual mesh data to confirm identity. Two items that share similar statistical properties but are geometrically different will never be classified as clones.
Stage 1
Seven detection layers analyze visual appearance, mesh structure, shape distributions, texture data, topology, surface curvature, and normals. This surfaces potential matches from the entire database.
Stage 2
Candidates undergo direct geometric comparison — Shield compares actual mesh geometry to confirm whether two items are genuinely identical. This eliminates false positives from items that happen to share statistical properties.
Visual Agreement
Items that look visually different can't receive a Clone verdict — regardless of geometric similarity. This catches common-shape coincidences where two unrelated items share a geometry class but are clearly distinct.
Clone Detected requires geometric verification
Shield doesn't just find similar items — it verifies actual geometric identity. No item is ever classified as a clone based on statistical similarity alone. The geometric verification step is mandatory for every Clone Detected and Mesh Theft verdict.
Scoring Engine
A patch and a character aren't scored the same way. The engine classifies each item's morphology — flat, elongated, compact, complex — and adapts layer weights accordingly.
Cross-category intelligence
A hat and a necklace can share similar geometry without one being a copy of the other. Matches across different item categories are automatically capped at Review — never escalated to Clone Detected.
Verification gating
Statistical similarity alone is never enough for a Clone verdict. Every candidate must pass geometric verification — a direct comparison of actual mesh geometry. Without verified identity, the match is capped at Review.
What We Catch
Visual and structural layers agree, confirmed by geometric verification. Strong evidence of direct duplication with verified mesh identity.
Structural match with geometric verification, visual mismatch. Geometry stolen and retextured. Verified mesh identity confirms the theft.
Visual match, structural mismatch. Same texture applied to a different mesh.
The Engine
Written in Rust, compiled to WebAssembly, deployed to Cloudflare's edge network. Stripped, optimized, and running at near-native speed — everywhere.
Rust
Source language
WASM
Compile target
0
Automated tests
<3s
7 layers + verification
3D Evidence
Interactive 3D previews built from the same mesh data Shield fingerprints.
Every scanned mesh is simplified to ~5,000 vertices for preview. That simplified geometry is stored and rendered on demand — solid dark faces with colored wireframe edges, drag to orbit, scroll to zoom.
When a match is found, both meshes appear side-by-side with synced camera controls. Drag one viewport and both rotate together — your item in blue, the suspected clone in red. Visual proof for DMCA evidence that a thumbnail alone can't provide.
~5K
Preview vertices
R3F
React Three Fiber
0 KB
Initial bundle cost
0fps
Synced comparison
[1] Perceptual hashing — PhotoDNA (Microsoft), Content ID (YouTube), PDQ (Meta).
[2] Osada et al. (2002). Shape Distributions. ACM Transactions on Graphics, 21(4).
[3] Reuter, Wolter, Peinecke (2006). Laplace-Beltrami spectra as Shape-DNA. CAD, 38(4).
[4] Texture perceptual hashing via DCT frequency analysis.
[5] A3 angle distributions — triplet angle sampling for shape comparison.
[6] Mean curvature distributions — angle deficit curvature histograms.
[7] Normal distribution histograms — surface orientation analysis for shape discrimination.
fingerprints indexed and counting
Run a scan on any Roblox UGC item. Or register your portfolio and let the engine monitor the catalog for you.