ꘇ image preview

A preview that leaves no trace

ComfyUI already has Preview Image. ꘇ image preview is what you reach for when you want to see an image mid-graph without the standard preview's side effects - files in your temp directory, the preview-method setting in Comfy's options, and viewers that behave differently depending on which frontend you're on.

The trick is that it never writes a file. It encodes the frame in the backend, hands the browser a base64 data URL, and draws it in the node. Nothing lands in temp/, nothing shows up in an output folder, and there's nothing to clean up. For a node you'll sprinkle around a graph while debugging, that's the right default.

The mechanism

From the shipped source, in order:

1. Take frame zero of the batch and scale it to 8-bit.
2. Check whether the image has any real transparency. If the alpha channel isn't uniformly opaque, keep it and encode PNG; otherwise drop to RGB and encode JPEG at quality 85, because JPEG is much smaller and this is a preview.
3. Thumbnail bilinear to your preview_size - the default 512, in steps of 64, up to 2048.
4. Emit the data URL to the frontend, plus a timing value in milliseconds so you can see what the encode cost.

It's flagged not_idempotent, which is Comfy's way of saying "this node's output isn't a pure function of its inputs" - the right classification for a live view, and the reason it updates whenever something upstream executes rather than sitting on a cached result.

Passthrough preserves the entire batch. That's the important line, because it means the node is invisible to the rest of the pipeline: the image output is what you put in, all of it, not the downscaled preview. So the standard move is to wire it inline - texture → preview → next operator - and it costs you nothing but the encode.

Inputs and outputs

• image - IMAGE, required.
• preview_size - INT, default 512, 64–2048 in steps of 64. The only knob. Bigger is more legible and slower; 512 is a sensible ceiling for a sidebar-sized panel.
• image - IMAGE out, the untouched input batch.

It's an output node, so it always runs when you queue - you don't need to wire it to anything downstream. Which is also how you should treat it: a probe you leave in place.

One caveat that the pack's own docs are careful about: a frozen or minimised preview skips its local drawing work, but connected consumers and open output viewers keep receiving frames. Freezing a display is not the same as muting a branch.

Install

Genereti only. Manager → search Genereti (publisher liuboto) → install → restart → refresh the browser tab, because the display is a frontend extension and a stale page shows a blank box. Manual install:

cd ComfyUI/custom_nodes
git clone https://github.com/languel/genereti.git genereti
cd ..
./.venv/bin/python -m pip install -r custom_nodes/genereti/requirements.txt

Use ComfyUI's interpreter - .venv on a checkout, python_embeded\python.exe on Windows Portable, the bundled one on Desktop. The requirements file is one line (comfyui-frontend-package>=1.53.6), so this is not a heavyweight install: no model downloads, no inference server, no extra node packs. ComfyUI 0.38.2+ and Python 3.11+.

Worth saying plainly, since it's the natural objection: a node this simple does not require a 49-operator performance library. If all you want is a preview that avoids temp files, you're dragging in a lot of unrelated alpha code to get it. That's a legitimate reason to skip it and keep using Preview Image.

Gotchas

• It's still browser-side, despite the name. The name LiveImagePreview and the "no files" behaviour make it feel like a headless-safe node, but the display is JavaScript. Queueing from a headless/API client works - the node executes - you just don't see anything.
• JPEG quality 85 on the preview. Don't use it to judge fine grain or noise. Look at the real output for that.
• Frame zero again. Batches pass through whole; the panel shows one.
• Alpha. The README says "super alpha - early classroom testing… back up your workflows before updating," and the pack has essentially no community discussion to fall back on. Its own docs are the best reference you'll get.