Có gì mới

Changed

• Strava account connection is being retired. In June 2026 Strava began charging developers a subscription for API access. We decided not to pass that cost on, so StintBox will stop connecting to Strava accounts. Nothing you have already imported is affected, and nothing about Strava data support changes: export the activity from Strava as a GPX, TCX or FIT file and StintBox reads GPS, heart rate, power, cadence and altitude straight out of it. FIT keeps the most, GPX is essentially position and time. The import panel now says so where you would actually look for it, rather than leaving you to find out that the file route existed.

Fixed

• Being at Strava's connection limit looked like a crash. Strava caps how many accounts an app may connect, and it does not refuse the sign-in: the authorization and the token exchange both succeed, and only the next call fails. StintBox treated that last step as unimportant, reported success, saved the tokens, and displayed "Connected" for a connection that could never work. It now stops at the point it actually failed and says which limit it hit.
• Failures said nothing at all in Preferences. The Strava card read every part of its own state except the error, so a failed connection simply stopped spinning. Every failure now has a sentence, in all 30 languages, and each one says what to do rather than repeating a status code.
• Disconnect never told Strava. It cleared the local token and left the authorization live on Strava's side, so the app looked disconnected while the grant was still open. It now revokes properly. Disconnecting from inside Strava also deletes what we hold, instead of only being logged.
• Signing in could hang forever. If Strava answered in the browser rather than sending you back, the app waited with the button disabled and no way out. It now gives up after ninety seconds and offers a way to cancel sooner.
• A saved sign-in was rejected for the wrong reason. Refreshing an expired Strava token had never been able to work in a released build, so the app blamed the user's authorization for what was a configuration gap on our side.
• Re-importing a Strava activity added it twice. The check for "already in this project" did not recognise Strava's own sessions, so the match banner kept offering an activity that was already there.

New

• Hardware encoding for clip exports. NVENC on NVIDIA, Quick Sync on Intel, AMF on AMD and VideoToolbox on Apple are now used for clips, not only for full exports. On a normal laptop that is the difference between a clip you wait minutes for and one that is finished before you have decided on the caption.
• Custom layouts export cameras you have not opened. A layout referencing a camera that is not currently loaded no longer blocks the export.

Improved

• The export dialog names the encoder it actually chose, so "why is this slow" has an answer on screen instead of in a log file.

Using the hardware people already own, rather than asking them to wait or to upload, has been the point since the first version.

New

• A layout builder. Split any cell horizontally or vertically, drag the dividers, drop a picture-in-picture wherever you want it, add or remove windows, and assign a camera to each cell. Start from a preset or from an empty frame.
• It exports what you drew. Custom layouts go down the same path as the built-in ones, so there is no class of layout that looks right in the preview and comes out wrong in the file.
• Previews with data in them. Every layout tile animates on hover with a live telemetry readout drawn into it, so you can see what a layout does with your data before committing to it. The same preview is used everywhere a layout appears.
• Undo covers the builder. Splitting, removing, changing aspect ratio and reassigning cameras are each a single step back.

This closes the arc that opened with multiple camera angles: several cameras, then a catalogue of layouts to put them in, and now any layout you can draw. It is also the clearest illustration of why we stopped assembling and started building. Nothing you add from the outside gets to redefine the frame itself.

New

• Live before and after. Data Doctor now slides in over whatever you were doing, with the original in dashed orange and the repaired version in solid blue on the same axes. Move any control and both the graph and the video preview update as you drag. No Preview button, no waiting to find out whether that was better.
• An auto-tuner. It sweeps smoothing strength across your actual data, scores each result on how much real structure survives, and tells you what it recommends, why, and how confident it is. On clean data it recommends no smoothing at all, which is the answer most tools are structurally incapable of giving.
• What changed, in numbers. Chips report exactly what was done: spikes fixed, samples smoothed, gaps filled, and gaps deliberately left alone because filling them honestly was not possible.
• Sport presets for motorsport, cycling, drone, running and walking, plus tunnel, indoor and urban variants, tuned from the benchmark rather than by feel.

Fixed

We had no honest way to tell whether a repair helped. Smoothness is easy to measure and easy to fake: flatten a signal hard enough and it scores beautifully while the driving disappears along with the noise. So we made tracks where the true line is already known, damaged them in eight controlled ways, and scored the repairs against the truth instead of against how tidy the output looked.

Three things were wrong.

• Doppler fusion was adding error, not removing it. Ground speed was being combined along a heading worked out from noisy position data, which corrupted the very thing it was meant to improve. On real files, position error dropped from 2.9 m to 0.6 m once that was corrected. Before the fix, leaving the data completely alone beat our own balanced preset.
• The smoother cut corners. It assumed constant velocity, so a tight circle came out about 208 m off line. That is the one place a racing line has to be right. A correction is now limited to what the measurement can actually support.
• Spike removal was deleting real driving. It discarded 32 to 38 percent of high-rate GPS as spikes, because the acceleration it derived magnified position noise enormously at high sample rates. It finds spikes a different way now, and on real files it flags none.

On real RaceBox sessions, track drift went from 134 m to 6.7 m on one and from 20 m to 4.3 m on another.

The uncomfortable half of the result is that less smoothing beats more, including less than we had been shipping. The defaults came down to match.

New

• The test bench runs on every build. A repair that quietly gets worse now fails here rather than on your footage.
• Over-smoothing is detected, not just noise. Two statistical tests look at what a repair threw away and check whether real structure went with it.

Improved

• Sources already loaded in Sync Studio open straight into Data Doctor, with no re-import.

New

• Every GoPro since the Hero 5. GoPro changed its telemetry format between generations, so StintBox carries a decoder for each: the older layout used by Hero 5 through Hero 10, Max and Fusion, and the newer self-describing one in Hero 11 through 13. Both come out as the same channels, so a project does not care which camera filmed it.
• Insta360, DJI and Sony, read straight out of the video file, plus DJI's subtitle-track logs as a fallback for the models that write telemetry there instead.
• Garmin and Wahoo FIT files, and Garmin TCX, which brings heart rate, power and cadence into the same timeline as the lap times.
• Strava activities, linked to your account and matched to a video by timestamp, so a ride you already uploaded can become an overlay without exporting anything.
• Betaflight blackbox for the drone side: gyro, accelerometer, GPS and motor RPM.
• KML and KMZ from Google Earth, NMEA straight off a GPS receiver, and CAMM as written by Android phones.

That is twenty-two formats, all read by us, none of them routed through a converter we do not maintain. This is the part that was never going to arrive from the outside: when you own every reader, "does it support my device" stops being a shrug and becomes a date.

New

• Master and linked variants. Choose the shapes you need up front. StintBox builds one master and links the rest to it, so a trim or a widget change carries across instead of being redone three times in three places.
• Each variant keeps its own framing. A 16:9 shot cropped to 9:16 has to be pushed around to keep the car in frame, and that framing belongs to the variant. Everything else stays shared.

New

• Channel source bindings. A widget is no longer tied to whichever source the project treats as primary. Put a speed gauge on your RaceBox, a second one on the GoPro's own GPS, and compare them side by side in the same frame. Preview and export resolve the binding through one shared code path, so a widget can no longer show one source on screen and a different one in the exported file.
• Per-widget styling. Accent, needle, foreground, background, border and secondary colours, plus glow, scanlines, blur, opacity and font, layered on top of the theme per widget. Every widget renderer honours it.
• Bulk layout. Select several widgets for six alignment actions, two distribute actions and three match actions.
• Auto-scale follows the binding. A gauge scales to the peaks of the source it is bound to, not to the project's.

Existing projects open unchanged.

New

• A template system for clips. A template like Best Lap Reveal carries everything the clip needs: aspect ratio, layout, a target length with a sensible minimum and maximum, which widgets belong on screen, and the detection recipe that finds the moment in your data.
• Templates know which sports they suit and which platforms they are shaped for, so a vertical 22-second lap reveal is offered for a phone-shaped upload and not for a landscape one.
• The picker reads your footage. It works out your source's real aspect ratio, including rotation, and marks the templates that fit as recommended, rather than letting you find the mismatch after the export.

New

• Racelogic VBOX and RaceBox files, in both the binary VBO format and RaceBox's CSV export.
• RaceChrono, including motorcycle lean angle as a proper channel rather than a column you have to explain to the app.
• TrackAddict and AiM Race Studio, recognised on sight. Their column headers are detected automatically, so the importer arrives already knowing what each column is.
• Harry's LapTimer database files, with laps and sectors intact.
• Torque OBD-II logs, which brings engine data off the car's own bus into the same timeline as the GPS.
• A column mapper for everything else. Any CSV can be mapped by hand, with suggestions for each column, and the mapping is remembered for the next file from the same device.

If you have a logger we do not name here, tell us what it writes. Adding a format is a normal week for us, not a roadmap item.

New

• Import your existing projects from DashWare. A guided wizard reads the project file, maps its gauges onto StintBox widgets, finds the video it belonged to, and keeps your layout and data bindings intact. Years of setup arrive as a working project rather than as a screenshot to copy by hand.
• You approve the mapping before it becomes a project. Anything the importer is unsure about is shown with an override instead of guessed silently, and you see the result before you accept it.
• Built on a shared intermediate format with one adapter per program, so the next program we add reuses everything except its own reader.

We spent years inside these programs. Nobody should have to abandon that work to move.

New

• A quality report for every session. Data Doctor scores what you imported and names the problems in plain language: how long the receiver took to get a real fix, where the signal dropped out, which samples are physically impossible, how far the track wanders while you are standing still.
• Repairs you choose, one at a time. Spike removal, a Kalman smoother for position, gap interpolation, cold-start trim, signal smoothing, heading stabilisation and derived channels are separate stages you enable individually, not one "clean up" button that does something unspecified.
• Non-destructive. The original is kept. Turn repairs off and you are back to exactly what the device recorded, to the sample.
• Honest about what it could not fix. A gap too long to interpolate is marked as missing rather than quietly bridged with a straight line. We would rather show you a hole than let you present invented data as measured.

New

• Event detection from telemetry. The app reads your data rather than your video, and scores every moment on six axes: how big it was, how rare it is for you, how surprising it was given what came before, whether several channels moved at once, how long it lasted, and whether you have already seen something like it. A late brake into a corner you normally coast into scores high. The forty identical laps around it do not.
• It picks a set, not a top ten. Ranking by score alone gives you six versions of the same corner. The selector optimises for a varied, watchable set inside the length you asked for, and opens on a strong moment instead of a chronological one.
• Cuts on the beat. Add a music track and segment boundaries snap to downbeats. Beat detection, tempo and song structure are analysed inside StintBox, with nothing uploaded anywhere. Choruses get hard cuts, verses get crossfades.
• Speed ramping that slows into the dramatic part of a segment and releases on the way out, scaled to how the moment scored.
• Six sport profiles, plus your own recipes if ours do not describe what you ride, drive or fly.

For years, finding these moments meant an analysis tool telling us that something had happened, and us scrubbing through the footage by hand to find out where. This is that job, done by the program that already holds both.

New

• Multiple camera angles. Load an on-board camera, a chase camera and a helmet camera into one project and cut between them on a single timeline.
• Alignment without a clapperboard. Getting two cameras onto the same clock is the genuinely hard part of multi-angle video, and it normally means clapping your hands on camera before every session. StintBox tries five methods in order of accuracy and stops at the first that works: embedded timecode, the motion signature in each camera's own accelerometer, the sound of the same engine picked up by both microphones, a shared transient like a door slam, and finally GPS timestamps.
• The audio method is the one we are proud of. Two cameras that never saw each other still heard the same engine. StintBox cross-correlates a 40 to 800 Hz band of both soundtracks and reads the offset off the correlation peak. On footage where nothing else lines up, it lands within a frame or two.
• Lock and confirm. Nothing is aligned behind your back. Every proposed offset comes with a confidence score you can overrule before you lock the group.
• Telemetry loggers, standalone recorders and GPS units go through the same pipeline. Anything with a clock can be a source.