Musubi Case study · Lead UI/UX Designer · Looking Glass

Editing a hologram on a screen that can't show you one.

Musubi turns ordinary photos and video clips into holographic memories for the 7-inch Musubi frame, using local AI to add the depth a flat photo doesn't have. The frame is the only screen that can actually show the result, which means every editing decision in the desktop app has to be made on a display that's lying to you, a little, on purpose.

RoleLead UI/UX Designer
ProductDesktop app (macOS/Windows) paired with the 7″ Musubi frame
This piece coversAI conversion trust, preview modes, Magic Moment, albums, hardware transfer
Musubi desktop app Edit screen: Simulated HLD and Canvas preview mode toggle, a baby-photo thumbnail in the frame preview, and the Magic Moment panel with background, duration, scale, position and rotation controls. Edit screen: Simulated HLD vs. Canvas, Magic Moment controls
The physical 7-inch Musubi frame on a wooden table, displaying a holographic bonsai tree with visible depth inside the frame. 7″ Musubi frame: a bonsai hologram on display

A converter, not a filter.

Nothing about a flat photo actually contains depth. Musubi's desktop app reconstructs a volumetric scene from the source material using machine learning, including Gaussian splatting, the same radiance-field representation professional 3D-capture pipelines use, and it does it entirely on the machine in front of you. Nothing uploads to the cloud, there's no Wi-Fi setup gating the conversion, and a typical photo or clip finishes in seconds on ordinary consumer hardware. That's a heavier claim than a photo filter, and the interface has to carry the weight of a real reconstruction step rather than an instant preview.

TechniqueML-based volumetric reconstruction, including Gaussian splatting
Where it runsEntirely local: no cloud upload, no account, no subscription
SpeedSeconds per photo or clip, on consumer Mac/Windows hardware
Why the GPU requirementThe reconstruction is real inference work, not a lookup; the minimum-spec GPU list exists because of this step
ORDINARY PHOTO OR VIDEO CLIP no depth data EXISTING GAUSSIAN SPLAT Unity · Unreal · Blender Premiere · After Effects local ML reconstruction already volumetric, direct import ONE FORMAT MUSUBI‑READY HOLOGRAM
Two very different source paths (a phone snapshot with no depth at all, and a professional Gaussian-splat capture that's already volumetric) have to land in the same editable object. The editor and the frame never need to know which path a piece came from.

Editing what the editing screen can't show.

A laptop display can't render a lightfield. So every editing decision (composition, background, motion) is being made partially blind, on hardware that physically cannot preview the thing it's producing. Musubi splits the difference with two honest preview modes instead of one dishonest one.

Musubi editing screen in Simulated HLD mode: the butterfly composition shown inside a spotlit recessed box, approximating how the piece will look with depth on the physical frame.
Simulated HLD

A best-effort approximation of how the piece will actually look on the frame, for judging depth and feel, with the honest caveat that it's a simulation, not the real thing.

Musubi editing screen in Canvas mode: the same butterfly composition shown flat on a plain white background, with the depth simulation switched off.
Canvas

A flat, neutral view with the simulation switched off, for judging composition, cropping, and position without a depth guess influencing the decision.

Once you accept neither view is the ground truth, the remaining editing controls stay deliberately small:
BackgroundOn / Off / Partial
DurationHow long the piece holds on screen
Scale & positionSize and X / Y placement within the frame
Rotation90° steps

Magic Moment: editing a held instant.

The frame doesn't play video the way a phone does; it holds a still hologram. So a 30-second clip needs to become one moment worth holding, not a looping window. Magic Moment turns that into an explicit editing step: scrub the timeline, pick the frame that's the memory, and it becomes its own holographic object, with its own background, duration, scale, position, and rotation, independent of the clip it came from.

Video timeline
0:000:30

A moment is chosen, not detected

No auto-picked "best frame." Scrubbing to the exact instant is a deliberate act, the same way choosing which print to frame from a roll of film always was.

Its own controls, not inherited ones

Once selected, the moment gets the identical control set as a photo (background, duration, scale, position, rotation) rather than borrowing settings meant for a playing clip.

An album that never touches your library.

An album in Musubi is a curated view into the library, not a second copy of it. Removing a photo from an album doesn't touch the original; deleting it from the library is a separate, deliberate, permanent action. For an app built entirely around handling people's family photos, that distinction is the whole trust model.

LIBRARY: source of truth ALBUM: a view, not a copy "Grandma's 90th"
Dashed reference lines, not solid containment: removing a photo from "Grandma's 90th" breaks the reference, nothing else.

The frame is an appliance, not a monitor.

Musubi's frame connects over USB‑C for power and data; there's no HDMI, and it can't be driven as a second display. Content has to be explicitly transferred, which turns "save" into a real decision point rather than something that happens automatically in the background.

ConnectionUSB‑C only: power and data, no video-out protocol
ModelPush to the frame, not mirror or extend
Musubi Display cardShows exactly what's actually on the connected frame right now
Editing after transferChange the local original, then re-encode to push the update; the frame doesn't update itself
Removing from frameFrees space on the device only; the local copy in the library is untouched
Treating the frame as its own small, honestly-represented state (rather than pretending it's just another window) is what makes "why isn't this updated?" answerable instead of mysterious.

What carried over.

Three working principles from designing an editor for a display the editor itself can't show.

Two honest previews beat one confident one

Simulated HLD and Canvas each admit what they're not. A single "preview" mode would have had to lie in one direction or the other.

A still is not a clip with a trim handle

Magic Moment succeeds by treating the chosen frame as its own object, with its own controls, not as a special case of video editing.

Curation and storage are different jobs

An album that can't accidentally delete a memory is what lets people actually use the album feature on photos they can't replace.

Your questions, answered fast

Quick info about my services and process.

Have more questions? Reach out anytime.