NOV 2026 · NOTE
Live video at 18 fps on one GPU
Chunked causal generation, a bounded attention window and two-step sampling: how Beehive keeps up with a camera.
Read the overview arrow_forwardOur work is about one problem: editing video while it is still being filmed, at a quality and cost that make sense for a shop window.
Technical notes
NOV 2026 · NOTE
Chunked causal generation, a bounded attention window and two-step sampling: how Beehive keeps up with a camera.
Read the overview arrow_forwardNOV 2026 · NOTE
A product photo has no sense of size. A photo of the piece being worn fixes most of it; combining two pictures in one reference does not.
Ask for the note arrow_forwardNOV 2026 · NOTE
Training a small adapter on a brand's own videos, at the live model's two noise levels, so it keeps its streaming speed.
Ask for the note arrow_forwardRoadmap
NOW
Studio, store embed and extension running at 18 fps.
NEXT
Small adapters trained on a brand's own products, with permission.
PLANNED
Our own retrained model, tuned on try-on data.
PLANNED
A smaller student model for 24 fps, several shoppers per GPU and lower cost.
How we work
We report numbers from our own deployment, say which figures are targets, and test changes side by side on the same clips before they ship.
The same products and people for every comparison: right piece, right place, right colour, steady in motion.
A change that improves quality but drops below live frame rates does not ship.
We train only on data we own, have licensed, or have written permission to use.
Yes, on request. Contact us and tell us what you are evaluating.
We are open to it, especially on evaluation of try-on quality and efficient video generation.
It is planned. We will announce dates once training and evaluation are complete.