Some interactions could not load. The video controls, direct video link and written walkthrough are still available.
FRAMEWISE / ONEDRIVE PHOTOS
We save videos because there’s something in them we don’t want to lose. But the thing worth keeping is rarely the whole video.
Five seconds of your kid landing the jump. A great reaction buried between everything before and after. One useful answer in a long interview. You scrub, replay, drag back, miss it, and try again.
The video is there. The moment is trapped inside it.
With Framewise, I wanted to explore a bigger role for video in OneDrive: not just storing it, but understanding enough to help people find the moments worth keeping.
A 10-day team prototype · Intelligent video & familiar editing
The intelligence finds a starting point.
The person decides what to keep.
2 MINUTES / NARRATED DEMO
The recording could not load. Try the direct video link above, or read the walkthrough below.
English captions could not load. The written walkthrough below explains the demo.
00:00 / The opportunity. The demo opens in a OneDrive Photos gallery. Videos hold movement, sound, people and stories, but often remain in storage rather than becoming something people use.
00:22 / A clip from a phone. A 21-second racing video opens in the viewer and then in Framewise. The editor has a large preview, a timeline, and Trim, Crop, Audio and Highlights tools.
00:39 / Finding a starting point. Highlight detection considers sustained action, people, visual quality, camera stability, audio energy and natural start and end points. Three suggestions appear beside the racing clip, with time ranges, preview actions and corresponding marks on the timeline.
00:56 / Shaping the sound. A music track is added to the clip. The audio panel offers a balance between music and original sound, with fade and loop controls.
01:06 / A longer, edited video. An interview running over 30 minutes opens in the same editor. The narration describes a small video-highlight model. A 15-second suggestion appears, can be previewed, and becomes a selected range on the timeline.
01:43 / Making it usable. The highlight can be trimmed or refined manually. Crop presets include widescreen, portrait, square and social formats. A portrait crop is previewed, while Save a copy stays available in the toolbar. The recording closes back in the gallery.
01 / Question the first signal
Our first hypothesis was simple: track visual entropy, looking for where the most movement and change happened. Movement meant action. Maybe action meant a highlight.
I wanted us to evaluate whether that assumption held up. Once we ran the results through evals, the gap became clear.
FromWhere does the video change the most?
ToWhat tells us something worth keeping might be happening?
That changed the question we were designing around. A useful highlight was not simply the busiest part of a recording.
02 / Let evals improve the definition
We layered in more signals, evaluated the results, looked at where they failed, and improved the system again. No single signal got to decide.
More temporal precision across the video. Dense resampling inside a promising region then refined its start and end points.
Look for continuing action, not isolated spikes. Suppress scene cuts so a transition is not mistaken for a highlight.
Prioritise moments with visible faces when detection is available.
Favour footage that is sharp, detailed and well exposed.
Separate meaningful subject movement from the whole camera shaking.
Listen for another clue that something interesting might be happening, rather than relying on pixels alone.
Avoid cutting into strong motion. Adjust the duration to the density of the action.
Combine the signals, with graceful fallback when a capability is unavailable.
03 / Different videos, different problems
A short phone clip and a long interview do not ask the same thing of AI. One needs help finding the action. The other needs help finding the meaning.
A clip from your phone
Find the action, remove the waiting, and offer a useful starting point. Motion, people, quality, stability and audio can get us surprisingly far without deep narrative understanding.
See the phone-clip flow 00:22 →A longer interview
No one should need to scrub through half an hour to find the sentence they remember. We explored a video-highlight model to suggest a meaningful section for someone to preview and refine.
See the 15-second suggestion 01:06 →Not every video needs the heaviest AI model. Not every moment can be found through movement alone.
Whatever the technique underneath:
AI suggests. People decide.
04 / Familiar by design
Finding the highlight was only half the experience. We could have built an explicitly “AI” editing tool around it. I didn’t want a technology experiment sitting beside OneDrive Photos.
I grounded Framewise in the existing Photos editor: direct manipulation, recognisable controls, previewing a change, and deciding what to keep. The intelligence could be new. The interaction should feel familiar.
Framewise suggests a starting point. You decide exactly where the moment begins and ends.
Add music and balance it against the original audio without leaving the editor.
Crop the moment for different formats and destinations.
Make saving a copy the path when the moment feels right.
The intelligence happens behind the experience.
The agency stays in front of it.
05 / The wider product ambition
Once we could ask “where is the good part?”, I started seeing uses far beyond highlighting. The real output wasn’t just a 15-second clip. It was an understanding of where useful moments live inside people’s videos.
A trip slideshow shouldn’t need an entire 40-second video between its photos. It could use the five seconds where someone dives into the pool, everyone cheers, or the candles are blown out.
For Memories and Recaps, knowing which video belongs isn’t enough. The useful question is which moment inside it helps tell the story of a summer, a trip or a year.
For Create, those candidate moments could become the ingredients, rather than starting with hours of raw footage. They could also open up generative experiences we haven’t designed yet.
That was the bigger opportunity: intelligence that could travel across the product, rather than end at the editor.
06 / Make the opportunity tangible
From a question about highlights
to an end-to-end prototype.
We built the intelligence, evaluated it, learned why entropy alone wasn’t enough, and made the detection richer. We explored deeper AI for long-form content and connected it to an editor that still felt recognisably OneDrive.
We weren’t presenting a deck about what AI might do someday. We could show a video coming in, a useful moment being suggested, and a person shaping the result.
The work helped our team earn investment to build more: deeper intelligence, continued quality improvement through evals, a stronger video-editing foundation, and more room to explore creation and storytelling across OneDrive Photos.
That tangibility mattered. People didn’t have to imagine what intelligent video could mean. They could experience it. We built enough of the future to make the opportunity visible.
The ambition
OneDrive already keeps people’s videos safe.
Framewise asks what happens when it begins to understand what’s inside them.
Less searching for the good part. More control over what happens next. More ways to turn videos people already have into moments worth returning to.
