Fast Video Cataloger 10.4 Adds Local Text-Based Video Scene Search on Windows

What shipped​

VideoStorm released version 10.4 of Fast Video Cataloger on October 2, 2026. The headline feature is searching by plain-language description. The vendor's example is "a drone shot over a city at night". It says the search returned matches from an 11,766-scene catalog where no clip carried a drone or aerial keyword. That demonstration comes from the company. It is not an independent benchmark, and nothing supplied shows how it would perform on your footage.

Fast Video Cataloger is not new. The vendor describes version 10 as video clip management software for 64-bit Windows, built to handle thousands of video files on a standard PC. The vendor says all the new AI features in version 10 run locally on the user's machine. Version 10 also brought face recognition, built-in speech-to-text and automatic scene tagging. Visual search builds on that AI groundwork.

How the search works​

According to the vendor's documentation, you open the Search window, expand Scene Search and tick Scene search. You then type a description in the "Describe the scene" box. The results window lists the best-matching thumbnails first. The search starts when you pause typing, or immediately when you press Enter.

The search compares your words with the image itself. It knows nothing about the video's title, file name, keywords or notes. A client's name or a place the model can't recognize from the picture won't help. For those, the documentation points you to keywords, actors or face recognition, which you can combine with the visual search.

The vendor's tips for writing a description:

• Use plain, concrete English that says what is in the frame, what it is doing and the setting.
• Describe what is present, not what is missing. "A street without cars" will still rank scenes containing cars, so write "an empty street at dawn" instead.
• If results look off, add a distinguishing detail such as "at night", "close-up" or "from above". If a long description matches nothing well, remove detail instead.

Prerequisites​

This is the part most likely to trip people up after upgrading.

1. Download the model. Visual search needs a separate Visual search model, fetched through Manage AI models. If you already use image tagging or trained tags, only the text half is downloaded.
2. Have an index. From 10.4, each thumbnail gets a compact description recorded as a video is indexed. The setting is Preferences > AI Classification > Visual search index, and it is on by default.
3. Backfill older catalogs. Videos indexed before 10.4, or indexed with the option off, aren't searchable by description. You add them with the AI Scenes pass and its Visual search index option. That pass shows how much of the catalog is covered and can be stopped and resumed. The Search window shows the same coverage figure.

So after you upgrade, "searches the whole catalog" really means "searches the part that has been indexed." If you see a coverage line below 100%, expect gaps. The documentation says scenes outside the index can't be found by description until they are added.

Controlling the results​

• Filters still apply. Folder, date, cast, keywords and notes narrow which scenes get ranked. If a clip you expect is missing, check these filters first.
• Show best caps the list at 200 by default. You can set it anywhere from 50 to 2,000.
• Per video caps how many scenes from one video appear: 1, 2, 3, 5, 10 or All, with 3 as the default. Thumbnails from one long take look alike and score alike, so without a cap they can fill the list.
• Relevance score. The tooltip shows a number from 0 to 1. The vendor says scores are only comparable within one search. A good match to a description scores around 0.15 to 0.30. Two scenes that look alike score 0.8 or higher. A "low" score on a text search is therefore normal.

Find similar scenes​

Right-click any thumbnail, in a video's scene strip or in the results, and choose Find similar scenes. The catalog is re-ranked by visual likeness to that frame, and the starting scene is left out. This is handy for finding other takes from the same location. Per the documentation, this mode compares the stored descriptions and doesn't need the model downloaded at search time, though the index must exist. A text description and a similar-scene pick are alternatives. Choosing one clears the other.

Shared catalogs, REST and AI assistants​

When a catalog is shared through the Fast Video Cataloger server, the work is split between machines:

• The index is built wherever indexing happens, using the model on that machine.
• Any desktop client that searches by description needs the Visual search model installed locally.
• Web-client and AI-assistant searches use the server's copy of the model.
• In 10.4, the server picks up a newly downloaded model without re-running the sharing wizard or restarting the service.

The vendor documents a Model Context Protocol (MCP) connector, fvc-mcp.exe. It talks to the server over the REST API, so the catalog must be shared. Having the desktop app open isn't enough. The documentation says a search_scenes_semantic tool handles description-based search. It needs the catalog's index and the model on the machine sharing the catalog.

Admins should read the security guidance. The documentation says an API key is a long-lived credential, so treat it like a password. It recommends a Viewer key unless the assistant needs to tag or build bins, and says to change the default admin password before the port is reachable from other machines. It also notes that clearing Require Authentication serves the REST API to anyone who can reach the port. ChatGPT, which runs in OpenAI's cloud, can't launch a local program. The vendor says it needs OpenAI's Secure MCP Tunnel instead. That requires a paid plan with developer mode and a separate Platform account.

The local-processing claims describe the product's design. Nobody has independently audited them. Footage and the index staying on your machines is what the vendor says, and you should confirm it for your own deployment.

Trained Tags​

The release announcement says 10.4 also refines Trained Tags, which arrived in 10.3. A trained tag is a custom AI keyword taught from a few example thumbnails. A wrong match can now be right-clicked and marked "Not this". The keyword is removed and the tag stops matching similar frames. The Trained Tags window also flags examples that weaken a tag, such as near-duplicates. This detail comes from the press release, and I did not find it in the vendor documentation I reviewed.

Pricing and availability​

Version 10.4 is available for Windows 10 and 11. Pricing is $9.90 per month, $97 per year, or a $197 perpetual license. The order page says each license covers one Windows PC. It says the perpetual license includes one year of updates, and it lists a 30-day money-back guarantee. A 30-day free trial is advertised. These are the vendor's terms.

What to make of it​

This is a vendor announcement, and the performance claims are unverified. The documentation is candid about limits, though. The search doesn't read metadata, it needs English, and it only sees indexed scenes. Scenes are thumbnails sampled at an interval, not detected shot changes, so you're searching samples of the footage rather than every frame.

Even so, the design is practical for small teams who don't want footage in a cloud service. The vendor also positions it as a companion to Premiere Pro and DaVinci Resolve, not a replacement. If you have a large archive, the first job after upgrading is probably the AI Scenes pass. Run it, watch the coverage figure, and then try some deliberately vague descriptions.