Comet 异步视频收取

Comet 异步视频提交 returns a ticket in milliseconds. This node - "Comet 异步视频收取" - is where your actual video comes back, sometimes minutes later. It checks the pack's task list for completed video jobs, downloads the clips (or picks up locally cached files if the pack already saved them), and hands them out as video outputs you can preview, save, or feed into a 视频卡片.

It's a deliberately thin node, and that's the point. One input to care about, two outputs.

What you actually set

run_count (default 1, max 50) is the only real control. It's how many completed video tasks this run should collect. A run_count of 1 pulls a single finished job - enough for the common "I submitted one clip, now get it" case. Bump it to drain several submitted jobs in one go, which is exactly what you want after a batch of async video submits.

Outputs are video (the VIDEO asset - a file-backed object the pack's card and save nodes understand) and query_info (a STRING status readout: which task IDs were collected, the local filenames, any failures). Read that text output when a run looks wrong; it's the pack telling you what happened.

How it works

Under the hood receive() refreshes the task list, finds tasks whose status is succeeded, then for each one either reuses a locally saved copy or downloads from the task's video URL and saves it to your output folder (ComfyUI/output/cometapi_assets by default). Each collected task is marked downloaded so the same job won't be handed out twice. If nothing is ready, it returns a placeholder video with a "暂无可用视频结果" message instead of erroring - so re-running is safe and free.

The node always marks itself changed, which matters here: it means every queue run actually re-checks the task list instead of replaying a cached result. That's how "run the receiver again in a minute" works.

Install and the honest caveats

Standard pack install: ComfyUI Manager (search "ComfyUI-CometAPI") or git clone https://github.com/jieg9341-lab/ComfyUI-CometAPI into custom_nodes, restart. requests/aiohttp dependencies, no model downloads.

Three caveats worth carrying:

• Task state is process-local, so a ComfyUI restart between submit and receive can orphan a job. Video is slow, which makes this trap more likely than with images - give long jobs time, don't give them a server restart.
• Collected clips are billed clips. The receiver doesn't cost anything extra by itself, but the video you're pulling cost real money to generate, and re-running a submit because you didn't collect in time means paying again.
• The clip is whatever the cloud returned - same moderation, same resolution you asked for, same data-left-your-machine reality. Async changes timing, not content policy or price.