Feature request: ML-assisted movie identification with confidence scoring and manual review

[grigkirill 0]

Please add optional ML-assisted movie identification to improve matching for multilingual titles, transliterated filenames and ambiguous releases.

Problem

Incorrect identification can silently attach an unrelated movie’s title, artwork, cast and external IDs to a file. Fixing these matches manually becomes time-consuming in larger libraries.

Examples observed during an audit of my library:

• 1999 - The.Ninth.Gate.1999.BDRip.1080p.-CtrlHD.mkv was identified as The Green Mile, while the filename and runtime were consistent with The Ninth Gate (1999).
• The.Invite.2026.1080p.AMZN.WEB-DL[Rus.Eng]Rutracker.mkv appeared as Invite the Devil to Come and Play in an earlier snapshot. Its identification was subsequently corrected to The Invite (2026).

These are observed library states; a clean-library reproduction of the original mismatches has not yet been established.

Current environment: Emby Server 4.10.1.0, Linux, NAS-hosted media, Russian and English titles.

Proposed functionality

Build on the existing movie naming conventions and Identify workflow with an additional identification stage:

• Generate and rank candidates using multiple signals. Consider the filename, folder name, embedded title, localized and original titles, transliteration, release year and runtime. Handle edition differences and release tags without treating runtime or year as conclusive evidence.
• Use ML to improve candidate ranking. Resolve multilingual and semantic title similarities while keeping results tied to verified entries in enabled metadata providers.
• Support uncertainty. Automatically apply sufficiently strong matches; send ambiguous cases to a “Needs identification review” queue with several candidates and an explanation of the evidence. Leave files unmatched when evidence is insufficient.
• Check consistency between sources. Flag conflicting external IDs and prevent metadata from unrelated works or media types from being combined.
• Respect explicit identification. Preserve manually confirmed matches, locked fields, NFO metadata and explicit provider IDs according to their configured precedence.

Optional content analysis

For difficult cases, offer an additional background step that extracts title cards or credits using OCR. Audio or video fingerprint matching could be explored later where a suitable reference catalogue exists.

Controls and usability

• Enable the feature per library, with configurable automatic matching thresholds.
• Show the original filename, candidates, provider IDs and reasons for each suggestion.
• Provide a preview before applying corrections to existing items.
• Preserve watch history and playback progress when correcting identification.
• Support CPU-only NAS devices through workload limits, caching and processing only new or changed files.
• Make any cloud processing an explicit opt-in, with clear disclosure of the data transmitted.

Acceptance criteria

Evaluate against a labelled set of multilingual and ambiguous movie filenames, compared with the current matching workflow. Report incorrect automatic matches, successful matches and cases requiring review separately.

The main goal is to reduce silent incorrect matches while keeping confirmed identifications stable across rescans.

[Neminem 1965]

Name your files according to naming conventions.

And stop pirating.

[Luke 43418]

Quote

The.Invite.2026.1080p.AMZN.WEB-DL[Rus.Eng]Rutracker.mkv

This case can actually be resolved if we replace dots with spaces within what we send to moviedb search.