Skip to content
The model

Beehive edits live video as it streams

Point a camera, describe a change or show a product photo, and Beehive restyles every frame while the person keeps moving. No app, no 3D models, no rendering wait.

CONCEPT VIDEO

What Beehive edits

One model, six kinds of change

Each change comes from a sentence, a reference picture, or both. They all run on the same live stream.

A leather moto jacket product photo

Virtual try-on

Garments from their product photo, fitted to the body and moving with it.

Aviator sunglasses

Accessories

Glasses, earrings, caps and watches that stay in place as the head turns.

A tropical beach

Scene change

A new place behind the person, lit to match, while they stay in the frame.

Starry Night painting

Art style

Painting, anime, pixel art or animated-film looks applied to live video.

A marble bust

Material

Marble, clay, metal or toy bricks, keeping every movement.

Van Gogh self-portrait

Character

Become the character in a reference picture while keeping your own motion.

Performance

Measured, not promised

These figures come from our own deployment on a single 96 GB GPU. Faster and cheaper versions are on the roadmap.

2.7 s

from camera on to the first edited frame

18 fps

sustained live stream

960×552

output resolution

2 steps

of generation per chunk of video

Under the hood

A video model that never waits for the end of the clip

Beehive generates video one short chunk at a time, looking only at what it has already seen. That is what lets it answer a live camera instead of a finished file.

Explore the architecture arrow_forward
  1. 01Camera frames are compressed into short chunks
  2. 02The instruction and reference picture guide the change
  3. 03A large transformer edits each chunk in two steps
  4. 04A short rolling memory keeps the video steady
  5. 05Edited frames stream straight back to the screen

Questions about the model

What do I need to run Beehive?add

A browser and a camera. The model runs on our GPUs; your device only sends camera frames and shows the result.

Does it keep the person's face?add

For try-on, scene and accessory edits, yes: the instruction asks Beehive to change only the item and keep the face, body and motion. Character and style edits change the look on purpose.

How small can a product be?add

Clothing and glasses work well. Fine jewellery works best with a close camera framing and a photo of the piece being worn, because small details have few pixels at 960×552.

Can Beehive learn our products?add

Yes. A small adapter can be trained on a brand's own photos and videos, with their permission, to improve fit, colour and steadiness for their catalogue.

All questions arrow_forward