Skip to content

Technical documentation

How Similiter compares videos and where it stops.

This guide describes the verifiable behavior of the current engine. It separates byte-identical duplicates, full-length visual matches, and unsupported cases. Thresholds produce candidates for review, not permission to remove files. Use these details to interpret a result correctly and decide when manual verification or a new scan is required.

Progressive pipeline

Comparison starts with inexpensive checks and reaches the costliest verification only for plausible candidates. A sampled prefilter precedes full-file SHA-256 for exact copies; five frames perform visual screening, twenty-four build the fingerprint, and forty-eight verify the result. Incomplete verification stays pending or fails closed instead of becoming an actionable match.

  • Full-file SHA-256 confirms copies with identical bytes.
  • 5 frames for initial visual screening.
  • 24 frames for the complete temporal fingerprint.
  • 48 frames for final visual verification.

How to interpret results

An exact duplicate has cryptographic evidence over file content. “Same video” and “probable match” instead derive from coherent visual fingerprints across the duration and require checking proof frames, path, size, and quality. Confidence helps prioritize review but does not replace the user’s decision or justify automatic removal.

  • Exact duplicate: the complete-file SHA-256 matches.
  • Same video: strong visual coherence across the complete video.
  • Probable match: enough similarity to review, not to remove automatically.

Current technical limits

The pipeline aligns samples at relative positions across the same duration; it does not search for one sequence inside another video. Segments, offsets, edits, speed changes, reverse playback, crops, reframing, and timeline reuse are therefore unreliable. Audio is excluded, and Similiter does not read NLE project graphs, timelines, or dependencies.

  • No segment or subsequence search.
  • No audio comparison.
  • No NLE timeline or project interpretation.
  • No normalization for crops, mirrors, reverse playback, or speed.

Performance and benchmarks

No reproducible large-library benchmark is published, so this documentation makes no time claim for thousands of files or terabytes. A credible benchmark must state version, hardware, corpus, codecs, file count, hours and TB, first scan, cached scan, errors, false positives, and verification criteria. Until then, “built for large libraries” remains intentionally absent.

  • Repeat scans can reuse metadata, hashes, and fingerprints for unchanged files.
  • Speed depends on codecs, duration, storage, and decoding.
  • Performance figures will be published only with a reproducible corpus and method.

Let’s talk

Request information

Similiter technical guide and limits Configure your request. We will reply with availability and a confirmed price.

Only name and email are required.

By sending the request, you confirm that you have read the privacy policy.