Load Image & Crop (obvpm)

The stock Load Image gives you a file picker and nothing else. So there's a whole little ritual: load the image, feed it into an Image Crop node, type pixel coordinates into four widgets, guess wrong, reload, look up the actual dimensions, type four more numbers. For a workflow where you're pulling a face out of a reference sheet, you do that every single time.

Load Image & Crop (obvpm) folds the two steps into one node with a crop rectangle drawn directly on the preview. Drag to draw the area, drag inside it to move it, drag a corner to resize, click outside to clear it. No crop drawn means the full image. The crop is stored in normalized coordinates, so it survives reloads and a workflow moved to another machine.

The fields

• image - the file to load. Upload, drag & drop, paste, or pick from your input folder.
• crop - managed by the editor on the node; you never edit this by hand.
• max_megapixels - if the crop (or full image) is larger than this, downscale it to fit, aspect preserved. 0 disables it, which is the default.
• aspect - free, or a fixed ratio: 21:9, 2:1, 16:9, 3:2, 4:3, 5:4, 1:1, 4:5, 3:4, 2:3, 9:16 and friends.
• Outputs: image, and mask from the image's alpha channel, cropped and scaled exactly the same way - same behaviour as the stock Load Image.

The label above the selection shows its size in source pixels, and a row under the preview shows the full image size. If max_megapixels is doing work, the preview also shows the resulting size as Downscaled To:, so you can see the actual number going downstream rather than guessing.

The aspect lock is the interesting bit

Set aspect to a ratio and the crop rectangle keeps that shape as you draw, move and resize it. Switch ratios and an existing crop snaps in place - same center, same area - rather than jumping. With no crop drawn, a dashed rectangle shows the largest centered cut of that ratio, and that's what gets output.

One design decision worth flagging because it's the opposite of what most nodes do: if the stored crop disagrees with the ratio - you hand-edited the workflow JSON, or something changed aspect by wire at runtime - the node refuses at run time rather than silently reshaping your crop. Slightly annoying for thirty seconds, considerably better than a workflow that quietly crops a different region than the one you framed.

Masking still works

Right-click → Open in MaskEditor runs ComfyUI's own mask editor on the loaded file, exactly like the stock node. Saving writes a painted copy under the input folder's clipspace directory and points the node at it. The crop stays where it was, because the painted copy has the source's dimensions - and both image and mask come out cropped together. That's how you do a detail pass or an inpaint prep off a hand-painted region without a second Load Image.

Where it fits

This is the front door of most reference-image workflows. Crop the face out of the portrait, crop the outfit out of the sheet, crop the pose out of the screenshot - and if you're feeding several of them at once, the pack's Load Images & Compose node takes over from here, layering multiple cropped images into one sheet. For a single reference, this plus max_megapixels set to something sane is the whole setup, and it's a real improvement on doing the crop in a separate node with typed coordinates.

Install

ComfyUI Manager → search comfyui-obvpm, or:

cd ComfyUI/custom_nodes
git clone https://github.com/chanon/comfyui-obvpm

Restart ComfyUI. It lives under obvpm/image; searching the node menu for obvpm shows the whole pack. No Python dependencies and no model downloads. The editor works in both the classic canvas renderer and Nodes 2.0, which hasn't been true of every pack in 2026. All class ids carry the (obvpm) suffix since 0.2.0 - the fix for an id collision with another pack - and old workflows migrate on open.

Two notes

Changing the crop re-executes the node on the next run, so a re-crop does cost you a re-run of everything downstream. And the pack's pre-0.2.0 workflows are migrated automatically, but if you've copied a workflow JSON by hand between machines and something looks off, re-drop the node - that's the author's own advice for migration edge cases.