Comet 异步批量图像提交

Take everything Comet 批量图像 does - multi-prompt, multi-reference, concurrent cloud generation - and strip out the waiting. That's "Comet 异步批量图像提交." It's the batch node for the moment you realize a synchronous batch of a few hundred billed calls will freeze your ComfyUI queue for an hour. Submit it, walk away, and let the receiver pick up results whenever they're done.

This is the heaviest job in the pack, and it's also the one where the async split pays off most. A sync batch ties up a worker for the entire duration; an async batch just registers all the tasks and returns.

How it works

It inherits the batch machinery from CometAPIBatchImage but runs its submit() path like the single-image async node: upload refs, register the task (with per-subtask jobs), and hand back a query_info string containing the task ID and subtask count. The generation happens in the background; the node returns immediately.

Inputs are the batch image set with one notable change: there's no prompt widget because there's no single prompt - instead there's the batch_text input (COMET_BATCH_TEXT), which you source from the 批量文本卡片 node or an LLM's numbered/JSON list. The rest:

• channel (default grsai) and model (default nano-banana-pro, 28 choices) - where and what you're calling.
• pairing_mode - the same three-way choice as sync batch: all refs as one group per prompt, each image runs all prompts, or image-prompt one-to-one.
• concurrency (1–10) - how many subtasks run in parallel under this task.
• aspect_ratio, image_size, quality, reasoning_effort, background_mode - the standard image params.

References come through images, batch_text, and the virtual-wire image_1…image_16 inputs, same as the sync batch node. The output is a single query_info string - collect results later with CometAPIAsyncImageReceiver, which pulls ready tasks from the same internal list (in batch ordering) and returns them as an image batch.

When to use it over the sync version

Use async when the batch is big, the model is slow, or the workflow around it has other work to do. Use the sync Comet 批量图像 when you want one node that finishes in a single run and hands you images plus a summary - async moves the "did it succeed" question to the receiver, which is more powerful but a touch more to manage.

Install and gotchas

Same install story: ComfyUI Manager (search "ComfyUI-CometAPI") or git clone https://github.com/jieg9341-lab/ComfyUI-CometAPI into custom_nodes, restart, configure a channel key in the 设置中心. Only requests and aiohttp to install; no models to download.

Watch these:

• The cost multiplier is real and it's easy to misfire. pairing_mode = "每张图跑全部提示词" against a big reference list is a wall of billed calls. Decide the pairing before you hit queue, and check it again.
• The receiver only picks up completed tasks. Nothing is lost - tasks stay in the list until collected - but if you set the receiver's run_count too low, big batches need multiple receive runs to drain fully.
• Restarting ComfyUI can orphan in-flight tasks, since task state is held by the running process (and mirrored to the data folder). Don't submit a huge batch and then plan a server restart before you've collected.
• Your references all get uploaded to the reseller channel, so this is the node that makes your image set leave the machine in bulk. Keep that in mind for client work.