我想看看丨详细

When a workflow misbehaves, the first question is always the same: what is actually coming out of that node? ComfyUI makes you answer that with preview nodes you have to wire up, unplug, and forget to remove. LetMeSee - the pack calls it "我想看看" (let me take a look) - is the debug viewer that does it properly: you attach it to any output, it shows you a detailed breakdown of what's flowing through, and then it passes the data straight through so your workflow keeps working while you look.

It's from ComfyUI-QING, and it's the first node I'd teach a beginner to reach for when a pipeline goes sideways.

How it works

The design is deliberately minimal: one optional source input that accepts any type (ComfyUI's * wildcard - images, masks, latents, strings, tensors, whatever), and one source output that passes the data through untouched. In between, it runs a full analysis of whatever came in:

• data type and shape (for tensors: dimensions, dtype)
• value summaries (min/max/mean where meaningful)
• runtime of the node itself
• system context - memory in use, GPU name and VRAM (via psutil if it's installed, and torch's CUDA introspection otherwise)

That last part is the differentiator. Most debug nodes print "here's a value." This one also tells you your GPU's state at the moment the wire fired, which is exactly the info you want when something is OOMing or a tensor is on the wrong device.

Why it earns its place

Two reasons, and they're both about not breaking your flow. First, it's inline: because the output passes the input through, you can insert it in the middle of an existing wire - look, then keep going - instead of branching off and remembering to clean up. Second, it's an output node, which means it renders its analysis in the UI and in the workflow - right-click a node, attach this, and you've got an on-canvas debugger for that specific output. The pack's README explicitly frames it that way: right-click any node to attach its dedicated debug panel.

It also pairs naturally with the pack's LogPrinter - LetMeSee for the visual/UI inspection, LogPrinter when you want a timestamped line in the terminal instead.

The honest take

There's a reason every ecosystem ends up with a node like this: the plumbing layer of ComfyUI is full of type mismatches and shape surprises, and "what's actually in this wire" is a question you ask constantly. LetMeSee answers it without ceremony. It's not fancy - it's the tool you use a hundred times and never think about. That's the compliment.

One caveat about it being an output node: output nodes always run, so dropping several of them into a hot loop will slow things down. Use it to diagnose, bypass it to generate.

Installing

Part of ComfyUI-QING, so it appears with the whole pack. ComfyUI Manager: search "ComfyUI-QING". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/GAO-SHIQING/ComfyUI-QING
cd ComfyUI-QING
python install_dependencies.py   # or: pip install -r requirements.txt

Restart ComfyUI. Base deps are Pillow, opencv-python, scipy, scikit-image, cairosvg. psutil is optional - the node degrades gracefully without it (memory columns just show zeros), so it's worth a quick pip install psutil for the full system readout. No model downloads. China mirror: --mirror --auto.

Gotchas

The only real gotcha is the "it always runs" behavior above. And a small one: because it accepts anything, a completely unconnected source input doesn't error - it just shows you an empty analysis, which is easy to mistake for a broken node. If your debug panel is blank, the wire isn't connected, not the node.