ZML_图像暂停选择

ZML_图像暂停选择 is the "decide after you see it" node: it freezes the workflow, shows you the generated image(s) inside the node, and waits for you to route each one to one of three output channels before anything continues. Like a manual gate from the plumbing layer, except the gatekeeper is you.

It's the descendant of the old timed pause node (which blocked for a fixed 15 seconds). The author rewrote it because a fixed timer was too short to think - now it blocks until you click. The mechanism is a signal file handshake: the backend writes a .signal file in ComfyUI's temp folder, the node polls it while the frontend shows you the images and collects your channel assignments, and once the UI writes the file the node reads which images go where. Unselected channels get an ExecutionBlocker, so nothing downstream of them runs - that's how the three outputs stay clean instead of emitting placeholder garbage.

The interface

• 图像 - IMAGE input. Feed it the sampler/preview output (batches work; it'll show every frame and let you distribute them individually).
• Three outputs: 图像_1, 图像_2, 图像_3. You assign each incoming image to one of them in the node's UI, and that's the only channel that fires.

There's a 60-second timeout in the code: if you never click, it falls back to routing the first image to channel 1. And this node plays well with the pack's resolution-unify node - the README's exact recipe is "input 1-5 images of different resolutions, then manually decide upscale / save / discard."

Why you'd use it

The classic workflow: a sampler produces a batch, and you don't want to commit all of them to a slow save/upscale chain. Pause, eyeball the differences, send the keeper to the save branch, send the rest to a cheap discard or preview branch. It's also how you build a manual "quality gate" - nothing auto-saves until you've looked. Impact-Pack-style manual intervention, done the ZML way.

The honest limitation: it's manual. If your goal is automation, this node is the opposite of it. Use it where a human in the loop is the point - picking between variations, reviewing upscale candidates - not as part of a headless batch.

Install

Part of the ComfyUI-ZML-Image pack. ComfyUI Manager → search ComfyUI-ZML-Image → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/zml-w/ComfyUI-ZML-Image

then restart. No models, no heavy deps. English users want the README's translation patch - the UI and channel labels are Chinese. If the preview doesn't show up, it's usually a stale temp-dir signal file after a hard stop; rerun once and it clears itself. And remember the pack's own saving node (ZML save image) is the one designed to pair with this, because it correctly skips 1x1 placeholder images - a normal save node won't.