Image Crop + Rotate + Pad

Outpainting in ComfyUI has an annoying amount of prep: extend the image onto a bigger canvas, mask only the new area, and land on a size the model will accept. Most people do it with a chain of ImagePad, MaskComposite and ImageScale nodes and get it slightly wrong the first three times.

Image Crop + Rotate + Pad (QG_ImageCropRotatePad) compresses that into one node with a visual editor on it. Load an image, drag it around until the framing is right, drag a canvas edge to add padding, and the node hands you the transformed image, the mask of what needs generating, and a stitcher for putting the result back. It's the most interesting node in this pack.

How it works

The panel writes into normal node parameters - rotation_degrees, the crop_* fields, the pad_* fields - so save, undo and API submission go through ComfyUI's native path. You're dragging sliders, just with a picture attached, which means everything you do by hand has a parameter a script can set.

The mask is the clever bit. White means "the model may paint here", and three things land in it:

1. Padding you added - the fill-colour edges.
2. The void corners a rotation leaves behind.
3. See-through parts of your source image - anything under 90% solid, i.e. the soft edge of a cutout, treated as paintable.

Point three catches people out, pleasantly or otherwise: feed an RGBA cutout into an outpaint and the transparent parts become generation targets. If you'd rather keep your picture under the paint and only mark it with the mask - the right call for a low-denoise refinement pass - set painted_area to keep picture. On fill (the default) the model never sees what was there, which is what you want when the area genuinely needs inventing.

canvas_multiple is the small feature that saves the most runs. Set it to 32 and the output becomes a multiple of 32: a padded side grows a few pixels, an unpadded side gets trimmed by a few. VAEs are much happier on multiples of 8 and up.

The stitcher is a AUSBOSS_STITCHER value - not an image, but instructions for pasting a generation back onto the original, aligned to the transform you did. Wire it into a stitch node (AusBoss's Stitch Inpaint is the intended partner) and it restores kept source pixels over the result. stitch_blend is the ramp width at the seam; stitch_grow shifts the seam itself - positive lets generation eat a strip of real image, negative keeps more of the source.

The inputs you actually touch

Fourteen required fields and seven optional ones, most of them panel-driven. The ones worth knowing:

• image (required, enum) - the file picker, reading from ComfyUI's input folder.
• rotation_degrees (required, −180 to 180) - clockwise, applied before crop and padding.
• crop_aspect_ratio (required) - free, source, or a fixed ratio like 16:9; locks the crop handles.
• crop_x / crop_y / crop_width / crop_height - the crop rect on the rotated image; 0 width or height means "keep the full available side".
• pad_left / pad_top / pad_right / pad_bottom - fill pixels per side, all of which land in the mask.
• feather (default 24) - softens the mask into kept pixels; the image keeps a hard edge.
• canvas_multiple - the divisibility fix above; 1 disables it.
• fill_color - hex, R,G,B or a CSS colour name; unparseable values fall back to mid-grey.
• resize_to_megapixels + megapixels + resize_method + resolution_steps (optional) - core's Scale Image to Total Pixels semantics, aspect preserved.

Outputs: image, mask, stitcher, original, width, height, prompt_image. That last one is the one people miss: same picture with transparent regions shown on white instead of the fill, so a captioning node sees a backdrop rather than a grey slab.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/live1617/ComfyUI-QG-tools

Full restart; the console should say [QG-tools] v0.9.2 加载完成 | 节点数 24. No pip install needed. The node imports av for its video-frame path, but av is a ComfyUI core dependency, so any working install already has it. scipy and pymatting are optional with built-in fallbacks - their absence is not an error.

Search for it under crop, rotate, pad or outpaint; it was ported from AusBoss, MIT.

Where people get burned

Expecting alpha to be preserved. Transparent areas are paint targets, not transparency to keep. Set painted_area to keep picture when the source should stay under the paint.

A stitcher with nothing to plug it into. The stitcher output only means anything to a stitch node that understands the type. Skip it and composite manually if you must - but you then lose the "unmasked pixels never touch the VAE" benefit the KB names as the reason crop-and-stitch beat every other inpainting approach.

Rotation before crop. crop_x and crop_y are measured on the rotated image. Rotate first, then read the coordinates off the panel; do it the other way round and your crop numbers land somewhere you didn't choose.