Tulut YOLO Preview

Every automatic detailer is a black box until the first time it eats a render and you realize it masked the wrong thing. Tulut's pack has a honest answer to that: Tulut YOLO Preview, the debugging node you wire in before the detailers to see what their YOLO detector is going to find, and in what order. It draws a green box with a big index number on every detected region and hands the annotated image straight back to you.

What it does

You feed it an IMAGE and pick a yolo_model from the dropdown - which is just the list of .pt files sitting in ComfyUI/models/ultralytics/bbox/. Set conf_thresh and hit run. Out the other end comes the same image with green bounding boxes and numbers.

The numbers aren't decoration. The detailer nodes sort every detection left-to-right by the center of the box and use character_index to pick which one to refine. The index Tulut YOLO Preview draws is exactly that sort position. So a "character_index = 2" in the Face Detailer corresponds to the box labeled 2 here. If your shot has three characters and you only want to fix the one on the right, this node tells you which index that is without any guesswork.

That's the real use: it's a confidence-and-order inspector. You tune conf_thresh up until the false positives disappear and down until nothing you care about drops out - then you know the detailer will run on the regions you actually want.

The inputs, all three of them

• image - the render you're about to detail.
• yolo_model - which detector to run. Defaults to whatever's in your bbox folder; there's a "none" entry if the folder is empty.
• conf_thresh - detection confidence, 0.01 to 1.0, default 0.35. This is the one you'll actually fiddle with.

Output is a single IMAGE - the annotated picture. Feed the original image onward to the detailers; this node is a look-don't-touch diagnostic.

Install and setup

Same as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Tulut114/tulut_comfyui_Gadget

Restart ComfyUI. The node needs the ultralytics Python package, which the pack's requirements.txt does not install - run pip install ultralytics opencv-python in your ComfyUI environment or the whole detailer half of the pack (this node included) silently won't register. Drop your .pt detectors into models/ultralytics/bbox/ and they appear in the dropdown.

Gotchas worth knowing

• If the folder is empty you get a "none" option and nothing to detect with. The pack doesn't auto-download models - this is a BYO-detector setup.
• A low conf_thresh catches more, but on anime renders it also catches false positives (a face-shaped cloud, a hand-shaped shadow). The KB's detailing doc makes the point: a detector that masks a rock it decided was a face is a failure mode you debug here, before it re-renders anything.
• The preview only shows detection - it won't tell you how the crop, upscale, and redraw will come out. Think of it as the aiming stage.

It's a thin node, but it's the difference between poking at a detailer blindly and knowing what's going to happen. If you run any of the Tulut detailers, wire this in first.