Raw footage in. Better clips out.

The first update since launch, and it touches almost everything: which moments get picked, how they open, what the captions say, and how much the app tells you while it works. Runs locally, no subscription, no account.

Cards from the newest generation are no longer crawling

Everything except speech recognition was packaged against an Nvidia toolkit older than the newest cards. It ran, it did not crash, and it crawled: a recording that should take minutes could take hours. This build is rebuilt against the current toolkit, so those stages finally get real kernels.

If this was you, judge it by the per-stage timings rather than by the total, because this build also switches on a pass that installed copies used to skip.

Moments are picked by what happens, not by what is loud

• Every clip now needs a payoff. No more build-ups that lead nowhere.
• A long fight comes back as several clips instead of one shapeless cut.
• Quiet weapons count. A suppressed shot is still a shot.
• Menus, inventory, loading screens and dark stretches stop showing up in clips.
• A session with nothing in it returns nothing, an honest empty result instead of filler.
• The ten-clips-per-session ceiling is gone.

The first second decides, so it gets judged first

• Clips are ordered by how hard they open, not only by what happens later.
• A clip that opens on a frozen or empty frame is flagged instead of quietly sent to the top.
• The thumbnail is the frame a viewer actually sees first.

Captions that sound like a person

• Captions describe the clip instead of reporting what the camera saw.
• The title carries the point; the description sits under it as a footer, the way feeds actually read.
• A caption appears in the language of the recording, or not at all. No half-translated lines.
• The game's name and its hashtag are used only when the game is named on screen.

Subtitles in the language being spoken

• Speech is recognised in the language spoken on the recording, not only English.
• Subtitles stay up while the words are being said. No phrases cut in half, no lines glued across a pause.
• Languages written in the Latin or Cyrillic alphabet are supported end to end. Japanese, Chinese and Korean are transcribed but not yet burned into the frame.

The app stops going quiet

• A run says how long it should take, and each stage shows progress inside itself.
• Work that drops to the processor is reported, not silently slow.
• The window is up in about two seconds instead of half a minute of nothing.
• Cancelling no longer looks like a crash, and a crashed run closes itself.
• Diagnostics describe this machine, and every warning says what to do about it.

Any RTX card

Memory is the only limit. The card and the driver are checked and named, and nothing asks for memory that cannot be freed. Cutting, ranking, rendering and burned-in subtitles work from the minimum card on the store page. The written titles and descriptions need about 10 GB of video memory; below that the clips are still cut, ranked and described, and the wording falls back to templates. The run notes say which case you are in.

Publishing and calendar

• Only clips the app actually recommends get planned and pre-ticked.
• Past weeks keep the times posts really went out.
• Week chips show the clip's poster full height and the whole status line.
• Disconnecting a connected account now removes what was stored for it.

Fixes

Clips could run past the length set for them. The window would not resize and grew between monitors with different scaling. A stalled upload could sit in retry forever. Published links could fail to open. Interrupting a job now actually stops it.

As always: nothing is uploaded, nothing is transcoded by a stranger, and the licence checks once and then leaves you alone.