LiveCanvas operator guide

This guide follows the same checks enforced by the product. A preview is not a public broadcast, and a campaign does not consume live minutes until its session reaches LIVE.

1. Prepare content, then run a preview

1. Create a workspace and select its service region. This affects data, payments, media and platform configuration and cannot be casually changed later.
2. Enter a name and topic in My live sessions to create a draft. No products, avatar, voice or GPU startup are needed at this stage.
3. In Prepare content, draft with AI or write manually. Review, save and approve content before using it live. AI availability depends on your plan and the page's current status.
4. Prepare authorized hosts and voices in Hosts and assets. Stock hosts can be selected when available; an empty catalog does not imply a gifted host. Complete the checks required for your session. All six motion types are not a universal prerequisite for basic live output.
5. Add product information only when discussing or selling products, and review prices, inventory, promotions and claims. Content-only sessions do not require products.
6. In Studio, run preflight. Resolve every blocked check. If capacity is off, request GPU capacity once and wait for the signed worker to become ready.
7. Generate and watch the 30-second preview. Click Enable audio, confirm the volume meter moves, and check the face, lip sync, framing, and gestures.

Do not approve a preview you cannot hear. Replacing a voice or avatar version invalidates the old preview and requires a new review.

Text PDFs are parsed locally. A scanned PDF uses only the OCR processor configured for the Workspace's physical market; there is no Global/China fallback. OCR output remains untrusted and every extracted field stays pending until a person reviews it.

2. Prepare an authorized real-person avatar

• Record one continuous 60–180 second MP4 or MOV in portrait framing.
• Use an adult subject who has granted the required commercial likeness, voice, and synthetic-performance rights.
• Keep one clearly visible face, even front lighting, a stable camera, and a clean background.
• Include natural speech from the same person. Avoid background music, another speaker, filters over the mouth, strong echo, clipping, and long silence.
• Keep the head and mouth unobstructed. Include deliberate but moderate body and hand movement rather than rapid gestures.
• Review source media and voice playback, then watch the generated output. Check motion-pack or lip-sync previews when using those enhancements. Machine approval does not replace visual review.

Failed assets remain separate versions. Replacing one does not overwrite media previously used by a completed session.

3. Connect OBS Browser Source

1. Open the session's Output panel and copy the clean capture URL. Treat this short-lived URL like a password; do not post it or place it in screenshots.
2. In OBS, add a Browser source and paste the clean URL.
3. Set the source canvas to 720 × 1280. Use 25 fps for the portrait output workflow.
4. If available in your OBS version, enable Control audio via OBS. Make sure the source is not muted in the Audio Mixer.
5. Right-click the Browser source, choose Interact, and click Enable audio if the preflight layer is visible. Browser autoplay policy may require this one interaction.
6. Confirm that the OBS audio meter moves while speech is scheduled. Record a local 30-second test and play it back before sending video to a platform.
7. Capture both a full scene and a clean/green-screen output if your production uses both routes.

The operator must remain in Studio during a live session. OBS captures the output; it does not replace LiveCanvas pause, stop, health, policy, or billing controls.

4. Go live safely

• Confirm the Studio header says Draft preview before testing and Live output only after the explicit start action.
• Check Now, Next, and Queue. Keep at least 60 seconds of approved speech available.
• Verify the persistent audio indicator, output connection, GPU health, expected cost, remaining minutes, and destination-platform rule.
• Start the session only after OBS preflight passes. Billing begins from authoritative server-side LIVE state, not from opening the output page.
• Use Pause for an intentional hold and Stop to end and settle the session. Safety controls do not require Agent approval.
• After stopping, review the immutable session summary for duration, billed time, assets, products, script, errors, and usage adjustments.

If the LLM or asset provider is unavailable, already approved media may continue. The system must not invent new product claims or silently start a paid task.

5. Platform rules

• TikTok Shop US/UK: Avatar Live is disabled. Use Commerce Copilot with a real human presenter and follow the platform's current disclosure and commerce rules.
• Other Global destinations: confirm the destination permits synthetic presenters and follow applicable AI-label, advertising, privacy, and consumer-protection requirements.
• China destinations: use the China workspace and the destination's official live companion. Apply required explicit and implicit AIGC labels.
• The V1 output page is designed for OBS or a platform live companion. Users control the destination account and final platform start action.
• Never publish an unapproved price, inventory statement, promotion, guarantee, health claim, financial claim, or safety claim.

Rules can change. A platform capability shown as blocked in Studio cannot be bypassed by editing a browser request.

6. Troubleshooting

The preview has no sound

Keep the session stopped. Click Enable audio and watch the output volume meter. Studio distinguishes a missing TTS result, a video without an audio track, encoded silence, a locked browser, and a muted browser. Follow the specific recovery message instead of repeatedly restarting the GPU.

OBS has video but no sound

Open Browser Source Interact, unlock audio, confirm Control audio via OBS when available, unmute the source in the Audio Mixer, and make a local recording. Do not begin the platform broadcast until the recording is audible.

Lip sync or gestures look wrong

Do not approve the version. Check face visibility, source sharpness, mouth obstruction, speech track, and the lip-sync report. Replace the source or motion candidate, then create a new preview. Faster motion is not automatically better motion.

The output says reconnecting

Keep OBS open while the output enters Safe Idle and retries. Monitor Studio for media and worker health. If it does not recover, pause or stop the session; fault time must not continue billing.

GPU capacity is off

Request capacity once from Studio. A retained node can take several minutes to restore and register. Do not create duplicate requests. Live minutes start only after the session reaches LIVE, and an unused on-demand node is automatically released after the configured idle grace period.

A session cannot start

Open preflight and resolve each specific block, such as required authorization, avatar or voice readiness, content duration, platform restrictions, plan, available minutes or output. Product information is required only for sessions that use it. Client-side changes cannot override server checks.

An account is connected but no comments appear

Account authorization, broadcast detection and receiving interaction data are separate states. Confirm that the connected account is the one broadcasting, then check data status and the last received timestamp. Retry failed requests or reconnect expired authorizations. A dash means data is unavailable, not zero.

The interaction list shows recently synced workspace data, which may include earlier broadcasts. It is not strictly limited to the current session.

How do I reply to a viewer?

Select a comment in the workspace, write a reply or request an AI suggestion, and review it. Save and approve it through the provided steps before using it in spoken output. Direct text posting to WeChat Channels is not supported; copy the reply and send it on the platform yourself. Automatic FAQ uses configured, confirmed rules rather than unrestricted AI publishing. Douyin is not yet connected; choosing its name does not enable a connector.

7. Get support

Create a ticket from Support or from the affected session. Include the session traceId, what you expected, what happened, and the approximate time. Never include API keys, provider credentials, output tokens, or private capture URLs.