Repository navigation
A 43-second showcase of Streaming Video Memory Buffer replaying one recorded run of the text pipeline on a Big Buck Bunny excerpt. The run uses the unchanged notebook benchmark preset, with Florence-2, Moondream2 and X-CLIP for perception and Qwen2.5-3B-Instruct as the reasoner.
What the video shows. It is recorded output, not live inference.
- At 02:23 the stream is asked what the rabbit picked up and held earlier. The apple was seen at 00:51–01:03.
- By then that moment had left the 60-second recent window, but it was kept in an episodic memory (00:41–01:05).
- The retriever returned that episode. Hierarchical memory answers "B. A red apple." The recent-only baseline answers "D. A carrot."
- Playback speeds and the jump back to the evidence frame are presentation choices, not inference time.
Scope. This is one qualitative example, not a benchmark result.
- On the clip's four questions, which were frozen before inference: hierarchical memory 2/4, baseline 1/4.
- A demo-only configuration that forces long-term consolidation inside the clip scores 1/4 each.
- The apple question was chosen for the video after the results were known; see
demo/showcase/RUN_PLAN.md. - The StreamingBench results are in the README.
Assets
| File | What it is | SHA-256 |
|---|---|---|
svmb-showcase.mp4 |
The showcase video: 1920×1080, H.264, 30 fps, 43.5 s, no audio | eb24bfba0cafd90738f3dc4954d99fd05dbe4c292eb3ff6cbe60c57e30c65531 |
bbb-excerpt-0033-0459.webm |
The exact 266 s source clip the recorded runs processed. python3 demo/fetch_showcase_media.py downloads and verifies it, then python3 demo/server.py --showcase replays the runs locally |
3435542701c60cf8593199f532a6a6ee76c50cfa7b131756aeba2e823a524712 |
svmb-showcase-first-attempts.zip |
The first attempts of both runs, kept for the record (see demo/showcase/README.md) |
b670256c64b31b13d69f22e72515e245fc6aa606299700c9e2695ab1ec5c2c24 |
Footage attribution. Big Buck Bunny, (c) copyright 2008, Blender Foundation / www.bigbuckbunny.org, licensed under CC BY 3.0. Source: https://peach.blender.org/.
- The clip, the video and the trace thumbnails use an edited excerpt: film time 0:33–4:59, audio removed, re-encoded to VP9/WebM 1280×720 at 24 fps. The video adds the viewer interface around it.
- The Blender Foundation does not endorse this project.
- The project's MIT licence covers its code, not the footage.