AI Video Asset Management: Cataloguing What the Model Made

Why AI Video Asset Management Fails First

AI video asset management is the layer most generative pipelines skip, and it is the layer that fails first. A traditional shoot produces a known quantity of footage: one day, one unit, a few hundred gigabytes, all of it captured under a single contract with a single set of usage terms. A generative pipeline inverts that arithmetic. A single thirty-second spot can burn through two hundred generations across four models, each with its own licence, its own prompt, its own seed and its own resolution ceiling. The footage got cheap. The context around the footage did not.

The failure mode is specific and it always looks the same. Three weeks after delivery the client asks for a fifteen-second cutdown, and nobody can find the source clip for shot four. Somebody has it, probably, in a downloads folder named after a date. The prompt that produced it lives in a chat thread that has since been archived. The model version has been deprecated. Regenerating is not an option, because the seed is gone and the current checkpoint renders the product at a slightly different angle.

This is not a storage problem. Storage is trivially cheap and getting cheaper every quarter. It is a metadata problem, and metadata has to be captured at the moment of generation or it never exists at all. Every day you postpone it, the only person who remembers what a file was for moves further from remembering.

The Five Fields Every Generated Clip Needs

Keep the schema small enough that an operator will actually fill it in under deadline. Five fields cover almost every retrieval question a production ever asks: source type, generation parameters, model and version, reference assets, and rights tier. Anything beyond that belongs in a project document, not on the asset. The model and version field is the one teams skip and later regret, because routing each job to the right engine only pays off if you can prove afterwards which engine produced which shot.

Source type is the one field with a real standard behind it, so use the standard. The IPTC Digital Source Type vocabulary supplies controlled values instead of free text: trainedAlgorithmicMedia for media created algorithmically by a model trained on captured content, compositeSynthetic for a mix in which at least one element is generative, and humanEdits for material a person altered with non-generative tools. The 2025.1 release of the IPTC Photo Metadata Standard went further and added AI Prompt Information, AI System Used and AI System Version Used to its Extension schema. Adopt the vocabulary even if your system has to hold it in a custom field, because controlled values survive a tool migration and free text does not.

Generation parameters means prompt, negative prompt, seed and any strength or guidance values, stored as structured data rather than as a screenshot of a settings panel. Reference assets means the stable identifiers of the stills, plates or audio you fed in, not their filenames. Rights tier means the licence class the output inherits from the plan it was generated on. Five fields, filled at ingest, answer the questions that otherwise cost a producer half a day each.

Naming Conventions That Survive a Deadline

A naming convention is only useful if a tired person can follow it at eleven at night. Keep it to four segments separated by hyphens: project code, shot number, variant index and a short model tag. Resist the urge to encode the prompt, the date or the reviewer's initials, because those belong in metadata where they can be searched rather than in a filename where they can only be misread.

Folder shape matters more than most teams expect, because it is the one structure that survives being zipped and handed to somebody else. Group by shot rather than by day or by model, since a shot is the unit a client gives notes on. Inside each shot folder keep three subfolders and no more: generations, selects and delivered. Anything that has not been promoted to selects within a week is a candidate for deletion, and treating it that way from day one is what stops the library from silently turning into an archive.

Reference frames deserve a top-level folder of their own rather than living inside the shot that happened to use them first, because the reference-first workflow behind character consistency depends on feeding the same reference back in across weeks rather than regenerating a lookalike. Bury a reference three levels down inside a shot folder and the next campaign quietly starts from scratch, which is how a brand's face drifts between flights without anybody making a decision to change it.

Provenance Does Not Survive Your Export

Content Credentials are worth attaching, but do not mistake them for a filing system. The C2PA specification binds a manifest to an asset using a hard binding, defined as one or more cryptographic hashes that uniquely identify the asset or a portion of it, and that binding is designed to match only that asset and explicitly not assets derived from it or renditions produced from it. A rendition, in the specification's terms, is what you get when a non-editorial transformation such as re-encoding or scaling is applied.

Read that again in production terms. Your mastering file carries a valid manifest. The compressed review copy is a rendition, so the hard binding no longer matches it. The graded, trimmed, captioned cut is a derived asset, so it needs a fresh manifest describing what changed. The specification's answer for tracking across those transformations is the soft binding, a fingerprint or invisible watermark computed from the content rather than from the bits, but a soft binding is a detection mechanism, not a catalogue.

The practical conclusion is unglamorous: your own database is the authoritative record, and embedded provenance is a courtesy to whoever receives the file. Write a stable asset identifier into your system at generation time, carry it through every transcode, and let Content Credentials ride along as a signal rather than as the index you actually search.

Rights and Disclosure Are Delivery Blockers

The metadata you fail to capture in week one becomes a delivery blocker in week six. YouTube requires creators to disclose generative AI content that makes a real person appear to say or do something they did not, alters footage of a real event or place, or generates a realistic scene that never occurred, and it labels the upload accordingly. Non-realistic or minor edits do not require disclosure, which means somebody has to make a judgement per asset, and they can only make it if the asset records what it is.

There is a sharper edge to this. YouTube also applies the label automatically to content that carries C2PA metadata, and its own documentation states that a label applied automatically to C2PA-tagged content cannot be adjusted by the creator afterwards. A provenance manifest attached casually during an experiment can therefore decide how a paid flight is presented to viewers months later. That is a decision worth making deliberately at ingest, with the rights tier recorded next to it.

Store the licence class on the asset rather than in the project brief, because whether a generated clip is cleared for commercial use varies by model, by plan and sometimes by month, and a brief is not the document anybody reads during an audit. Two clips sitting side by side in the same sequence can fall under different tiers if they were generated on different accounts, and that difference is completely invisible in the picture.

Retention: Decide What Gets Deleted

An asset library without a deletion policy is not a library, it is a landfill with a search box. Set three tiers and enforce them on a schedule. Delivered masters, approved selects and the reference set are permanent. Generations that lost a comparison are kept for the length of the revision window, typically thirty to ninety days after final delivery, then dropped. Everything produced during exploration before a brief was signed off goes at the end of the week it was made.

The reason to be aggressive is not the storage bill, which is negligible, but retrieval quality. A shot folder holding twelve candidates is browsable in a minute. The same folder holding four hundred candidates is not browsable at all, and the practical effect is that the next producer regenerates instead of searching, which is how a team pays twice for the same shot and gets a slightly different one the second time.

Retention only happens if it is explicitly somebody's job, which is why a producer-led operating model puts the asset library under the producer rather than the editor. Editors optimise for the cut in front of them; the producer is the only role with a standing incentive to protect the campaign after it ships. Give that person a fifteen-minute slot each week to run the deletion pass, and the whole system holds without anybody writing a policy document nobody reads.