Skip to content
Architecture

From camera to edited frame, in one continuous stream

Beehive is a streaming video model. It never waits for a whole clip: it compresses what the camera sees into short chunks, edits each one as it arrives, and sends it straight back.

  1. 01 Capture & encode
  2. 02 Condition
  3. 03 Generate · 2 steps
  4. 04 Rolling memory
  5. 05 Decode & stream

Animated illustration of the pipeline. Frames shown are from a concept video.

The pipeline

Five stages, running continuously

Every stage works on a small piece of time, so the first edited frame arrives in about 2.7 seconds and the rest follow at 18 frames a second.

01

Capture & encode

Camera frames are compressed into a compact latent space by a causal video autoencoder. Every eight frames become one latent chunk, the unit the model thinks in.

  • Chunk = 8 camera frames ≈ 0.44 s
  • Causal: never looks at future frames
  • 64 latent channels per position

02

Condition

The instruction is read by a 7-billion-parameter vision-language encoder. A reference picture, such as a product photo, is tokenised alongside it. Separate position codes tell the model which tokens are the live camera, which are the reference and which are the output.

  • Text: up to 1,024 tokens
  • Reference: product photo or character
  • Source-aware positional encoding

03

Generate in two steps

A dual-stream diffusion transformer, about 16 billion parameters across 40 blocks, edits each chunk. Image and text tokens flow in separate streams that meet in joint attention. A distilled sampler needs only two denoising steps per chunk.

  • 40 dual-stream blocks, hidden size 4,096
  • Flow-matching objective
  • 2 denoising steps per chunk

04

Remember, briefly

To keep the video steady, each new chunk attends to the three chunks before it and to one anchor chunk from the start. Older chunks are forgotten, so the cost of each step stays flat however long the session runs.

  • Window: 3 recent chunks + 1 anchor
  • Cached keys and values, reused
  • Constant cost per chunk

05

Decode & stream

The edited chunk is decoded back into frames and streamed to the browser over a WebSocket. Flow control keeps at most two chunks in flight: a live camera drops a frame rather than build up lag.

  • JPEG or H.264 frames
  • Acknowledged every 4 frames
  • ≤ 2 chunks in flight

Why it streams

Causal by design

Most video models look at a whole clip before producing anything. Beehive's attention only runs backwards in time, chunk by chunk, so it can begin editing the moment the camera starts and keep going for as long as the session lasts.

The rolling memory is what makes this practical: the model remembers enough to stay consistent from one moment to the next, but never so much that it slows down.

Attention at chunk t

Anchor chunk Recent window Being edited

Serving

Engineered for one GPU, in real time

memory

FP8 compute

The largest layers run in 8-bit floating point, roughly halving their cost.

bolt

Captured graphs

Each step replays a pre-recorded GPU graph, removing launch overhead.

hub

Fast attention

Modern fused attention kernels, with a portable fallback.

lock

Signed sessions

Browsers get short-lived passes; secret keys never leave the server.

MeasureValueBasis
First edited frame≈ 2.7 sMeasured
Sustained frame rate18 fpsMeasured
Output960 × 552Configured
Hardware1 × 96 GB GPUCurrent deployment
Sessions per GPU1 at a timeToday; multi-session is on the roadmap

What's next

Beehive-1 and Beehive-1 Lite

Beehive-1 is our own retrained model, tuned on try-on data. Beehive-1 Lite is a smaller student model distilled from it, aimed at 24 frames a second, several shoppers per GPU and a lower cost per session.

Read the roadmap arrow_forward

TARGET

≥ 24 fps

TARGET

≤ 1.5 s

TARGET

≥ 4 sessions

STATUS

Planned

Architecture questions

Why chunks instead of single frames?add

Video changes smoothly over time. Working on short groups of frames lets the model keep motion consistent and spreads its cost over eight frames at once, which is how it reaches 18 frames a second.

Why only two denoising steps?add

Standard diffusion models take dozens of steps. Beehive uses a distilled sampler trained to reach the result in two, which is what makes live editing possible on one GPU.

Does the session slow down over time?add

No. The rolling memory has a fixed size, and the memory is refreshed periodically on very long sessions, so each chunk costs the same at minute one and minute ten.

Is camera video stored?add

Frames are processed in memory to produce the edited stream. Recording is optional and should only be enabled with the shopper's consent.

All questions arrow_forward