Windows + NVIDIA · Local · Offline · One-Time

Per-game ranking, and what it actually changes.

CutBench ranks a long recording into short vertical clips. Two games have visual detectors trained on hand-labelled footage of them, which changes what the ranking can see in a session. Every other game runs the same pipeline on the signals that are not game-specific. This page explains the difference in mechanical terms, because "tuned for your game" is a phrase that usually means nothing.

01 · What tuning means

Same app, same pipeline, different detectors.

Tuning a game does not mean a separate app, a separate download or a separate engine. Every recording, tuned game or not, goes through the identical stack: the speech is transcribed, the audio is scanned for events like gunfire, explosions and voice, scene changes are detected, and trained classifiers read what is on screen.

The game-specific part is that last step, and only that step. Two of the visual classifiers were trained on hand-labelled footage, and the label set is per game. Concretely, that means three things:

None of that adds a stage or a new kind of detection. It changes what the ranking can make of the picture it was already looking at.

02 · How it gets applied

There is nothing to switch on.

No game dropdown, no profile to choose, no filename convention to follow. Both trained class sets are scored against every recording you hand over, and the one that recognises what is on screen is the one that ends up carrying weight. Point CutBench at a Rust session and the Rust classes respond to the wounded overlay and the loot window. Point it at something neither was trained on and they mostly stay quiet, which is the ordinary case described further down.

That is also why the list of tuned games is short and why no setting can extend it. Adding a game means labelling its frames and its clips by hand and retraining on them, not editing a config file.

03 · The trained games

Two games, and that is the whole list.

Two. Tuning a game means labelling its frames and clips by hand, training on them, then checking what the ranking picked against what a person would have picked. That is slow, so the list grows slowly. It is more useful to say so plainly than to publish a wall of game logos that all resolve to the same untrained fallback.

GameWhat the training gives the rankingPage
Rust Trained on Rust footage: the wounded overlay, the killed-by respawn screen, the loot window, the building menu and the world map, plus rockets, sulfur and high-tier guns read as valuable loot. On video, a firefight with another player lifts a moment, while swinging a tool at a node, road looting and standing at base push it down. Rust clip maker
ARC Raiders Trained on ARC Raiders footage: the downed overlay, the extraction-failed screen, rarity-framed loot, looting a body straight after a fight, and the zone map. On video, shooting and machine encounters lift a moment, while traversal and an idle camera push it down. ARC Raiders clip maker
Every other game No trained class set. The full perception stack still runs, and the ranking leans on the signals that are not game-specific: speech, audio events and scene changes. No page needed

Rust is by Facepunch Studios and ARC Raiders is by Embark Studios. CutBench is not affiliated with, endorsed by or connected to either studio.

04 · Everything else

An untrained game is the normal case.

Most sessions run through CutBench are not Rust or ARC Raiders, and that is fine. Every stage still runs: the transcript, the audio events, the scene changes, the 9:16 crop, the burned-in captions, the ranked list. What you do not get is a trained read on that game's own HUD and activity, so the ranking rests on the signals that work anywhere.

That works best where the good moments are loud and visible, which covers most shooters, extraction games and survival games. It is weaker where the drama is quiet or slow: a moment with no audio event and no scene change gives the ranking very little to go on. A puzzle or strategy game is the honest edge of what this tool is for.

Worth knowing before you buy, whatever you play: there is no text reading in the pipeline. Nothing parses a kill feed, a scoreboard, chat, a sign or a quest log. If the story of your clip is written on screen rather than heard in the audio or visible in the picture, expect the ranking to miss it. On Rust and ARC Raiders, a heavily modified or hidden HUD costs you a signal for the same reason. This is a limit of the design, and training on a game does not fix it.

Run it on the game you actually play.

The free demo runs the same ranking on your own footage, trained game or not, and marks the exports with a watermark. That is the cheapest way to find out whether the list it hands back matches the clips you would have cut yourself. If you want the tool-by-tool view instead, the comparison index covers the alternatives one at a time.

Windows 10/11 · NVIDIA GPU · one-time license
Last checked 14 August 2026.