Image Resize 🆎

Image Resize 🆎 is the resize node that thinks about the whole job, not just "make it bigger." It resizes an image to a width+height, a longest or shortest edge, a megapixel budget, or a scale factor - and when the aspect ratio changes, it can stretch, fit inside, center-crop to cover, or pad with a fill color. It optionally carries a mask through the exact same geometry, snaps the result to a clean multiple, and hands back the final dimensions as numbers you can wire into latent nodes. In a pack, that's usually a chain of three or four nodes plus a mask detour; the pack's own changelog says the point was to kill the most common reason a shared workflow drags in a heavy node suite - one dependency-free node instead.

The divisible-by detail is the one that makes it feel built for modern models. Many samplers, VAEs, and video models behave best on cleanly divisible sizes - 16 for WAN video, 8 for most latent spaces - and a resize that lands on an odd dimension can quietly produce artifacts at the latent stage. divisible_by (default 1) snaps each output dimension to the nearest multiple, and the node is explicit that the snap is resolved by a real sub-half-step resize, never by inventing pixels. So "divisible by 16" means a resize happens to land on the multiple, not that extra pixels get smeared into the edge.

How it works

target_mode picks which widgets matter: width+height reads width and height; longest_edge/shortest_edge scale until that edge hits edge_length; megapixels scales to a pixel budget (about that many million pixels, aspect preserved); scale_factor multiplies the source (0.5 halves both dimensions). A 0 in a size widget keeps the source dimension - and if only one of width or height is set, the other follows to preserve aspect.

When the target box doesn't match the source aspect, keep_proportion decides: stretch distorts to the exact box; fit shrinks the box to the source aspect (output may be smaller than you asked, no bars); cover_crop fills the box and center-crops the overflow; pad fits inside and fills the rest with fill_color - and here's the clever bit, the new bars come out as 1.0 in the mask output, so you get a ready outpaint mask for free. The other target modes derive their box from the source itself, so they never crop or invent pixels regardless of this setting. interpolation defaults to lanczos (the sharp all-rounder), with bicubic, bilinear, nearest (for pixel art and masks), and area (best for strong downscales).

Outputs: image (the resized batch), mask (the optional input mask through the identical transform; all zeros if none wired), and width/height as INTs - wire those straight into an Empty Latent or a sampler size.

Install

ComfyUI Manager is the easy route: search for ComfyUI-AusBoss under Custom Nodes and hit Install, then restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ausboss/ComfyUI-AusBoss.git

Restart ComfyUI and Image Resize 🆎 sits under 🆎 AusBoss/Image. No extra Python dependencies - the resize runs on Pillow and Torch that ComfyUI already ships. Minimum ComfyUI 0.27.1, and hard-refresh with Ctrl+Shift+R after frontend updates.

Common issues

The traps are the two defaults you might not read closely. keep_proportion only crops or pads in width+height mode - the edge/megapixel/scale modes derive the box from the source, so don't expect a pad there. And megapixels is a budget, not a guarantee: it targets "about" that many million pixels, so don't be surprised by a rounding sliver. If you're feeding a WAN video model, set divisible_by to 16 and don't rely on remembering to. For masks, nearest interpolation is usually the right call - a soft lanczos feather on a hard mask edge will bleed into your inpaint region.