Live-events use: GPU regions, data retention, commercial licensing, session rollover, and what leaves the client

[GoodnightJames]

Hi — I'm evaluating Scope (Mac app v0.2.5, Apple Silicon, remote inference beta) for live AI visuals at small music events in Thailand, likely composited through TouchDesigner via Syphon/OSC. Before going further I have six questions I couldn't resolve from the docs. Written answers here would be hugely appreciated — several of these gate real-world use for us.

1. GPU regions: Where do remote-inference H100s (and Daydream API GPUs) run today? Is there Asia-Pacific capacity, and can a session request or prefer a region? We're in Thailand, so round-trip latency is a core feasibility question.

2. Frame retention / training: For live video streamed through Scope remote inference and the Daydream API — are frames retained after the session ends? Are they used for model training? Is there an opt-out?

3. Commercial use of hosted Scope: The repo is CC BY-NC-SA 4.0, which clearly restricts self-hosted commercial use. The Daydream service terms permit monetizing User Creations. What's the intended path for using Scope (hosted remote inference, or self-hosted) to produce paid live-event visuals? Is written permission or a commercial license available?

4. Post-beta pricing: What should builders expect for remote inference and API pricing after the beta — per-minute GPU metering, tiers, something else? Even a rough order of magnitude would help us plan.

5. Session rollover: At the one-hour remote-session limit, does reconnection reuse the warm GPU and already-loaded pipeline, or does it re-provision from scratch (the documented 2–3 min provision + 1–2 min pipeline init)? Can a second session be started before the first ends, to hand over without a gap during a live set?

6. What exactly leaves the client: In a pipeline where only a control map (depth/edges) feeds generation, is the original RGB camera image also transmitted to the remote GPU, or only the tensors the pipeline consumes? More generally: which pixels/tensors are transmitted, processed, retained, or logged? (This matters for privacy compliance when pointing a camera at event guests.)

Thanks — Scope on the Mac is genuinely exciting, and the free beta is a great on-ramp. Happy to share latency/fps measurements from Thailand once we run our tests if that's useful data for you.

[GoodnightJames]

Follow-up with the measurements we promised, plus two new questions that came out of testing (Scope 0.2.5, Apple Silicon, remote inference beta, tested from Thailand today):

Measurements (free beta, H100 runner, 79.2GB VRAM reported):

• Cold GPU provisioning: 5m02s from connect request to "Livepeer connected"; first pipeline load (LongLive + 2 LoRA permanent merges): another 4m41s. Total cold start ≈ 10 min. Warm restarts after stop: 15–35s.
• LongLive t2v @512×512 + RIFE: rock-solid ~30fps backend (34–35fps client), 8.2–8.9 Mbps, thousands of frames without a drop. Genuinely impressive.
• StreamDiffusion V2 v2v @512×512 (no LoRA/RIFE): 7.67fps sustained backend, 6.26s media TTFF.
• Network from Thailand: ~192ms ICMP to the orchestrator (US-West), ~400ms app-level pong. (Hence our Q1 about APAC capacity.)
• Live prompt push via POST /api/v1/session/parameters: applies instantly; visuals converge gradually (autoregressive momentum) — nice for transitions.

New Q7 — idle metering: after stopping all workflows, with zero active inference sessions but Remote Inference still connected, our balance fell from 25.00 to 7.50 credits over ~15.5 minutes. Is all cloud-connected time metered (i.e., the runner reservation), not just active generation? If so, a UI hint near the connect toggle would save testers a lot of credits.

New Q8 — test capacity: we're mid-evaluation for live-event use (Krea+VACE and the 1-hour rollover behavior still untested) and the remaining 7.50 free credits can't cover it. Is there any way to get additional evaluation credits for testing? Happy to keep sharing structured measurements from Thailand in return — glass-to-glass latency with camera input is next on our list.