🍒Image_Square_Pad⬜图片正方形填充

Every image model wants squares. Your images aren't. Image_Square_Pad fixes that the non-destructive way: it extends the shorter side so the canvas becomes a square, centering the original and filling the gap with transparency or a color - without scaling, cropping, or touching a single pixel of your content. The author's own docstring describes it as "similar to Photoshop's canvas size function," and that's exactly right: it's the canvas that grows, not the image.

Why you'll reach for it: square-cropping a portrait loses content; squishing it distorts; this pads it. Feed the result to a square-latent pipeline, a tile-based upscaler that wants square inputs, or a model that rejects non-square resolutions, and you keep every pixel of the original. The README example shows exactly that - a wide banner image becoming a square without losing the sides.

How it works

This is one of the few nodes in the pack that's pure tensor math - no PIL round-trip, no cv2. It takes the image tensor, computes the long edge, creates a new square tensor filled with your background, and copies the original into the center. Because it never re-encodes the pixels, it's genuinely lossless.

The alpha input (a MASK) lets you bring in an existing alpha channel: it gets merged into an RGBA output, with invert_alpha (default on) flipping the mask first - the tooltip says "usually needs to be enabled," because ComfyUI masks are typically inverted relative to what you'd expect as opacity. The color input takes a hex string like FF0000 or #FF0000; leave it empty and the padding is transparent. If your image is already square, the node returns it unchanged - no wasted work.

The inputs that matter

• image - what to pad.
• color - hex fill for the pad; empty = transparent.
• alpha + invert_alpha - bring in transparency from a mask when you need RGBA output.
• Everything else is automatic - it always centers, always uses the long edge.

Output: image, the padded square (RGB if no alpha/color, RGBA otherwise).

Installation

Standard KimNodes pack install - Manager → search "ComfyUI_KimNodes" → Install → Restart, or git clone https://github.com/wjl0313/ComfyUI_KimNodes into custom_nodes. No extra dependencies - it's torch/numpy only.

Common issues

• Transparent padding shows up black downstream - you're getting RGBA but a downstream node that ignores alpha renders the transparent area black. Either feed the result into something that handles alpha, or set color to a solid hex like FFFFFF.
• The mask looks inverted - that's the invert_alpha default (on) doing its job. If your alpha appears wrong, flip invert_alpha off.
• "Invalid hex format" fallback - a malformed color string (e.g. 3-digit hex or missing letters) makes the node fall back to transparent rather than erroring. Use 6 hex digits.

For anything that needs square inputs without pixel loss, this is a quiet workhorse - and it behaves exactly the way you'd hope a canvas-padding node behaves.