🦾 Image Scale Down By

Image Scale Down By shrinks an image by a factor - half size, quarter size, whatever you set - and returns the smaller image. It's the "make this lighter to look at or feed into a preview" node: generate big, downscale for a thumbnail, or downscale before a cheap first pass to save memory.

The name says "By" because it's the sibling of the same pack's ImageScaleDown (which takes absolute width/height). This one takes a scale_by factor instead, which turns out to be the more convenient API most of the time: "0.5×" is easier to reason about than "give me 768 wide."

How it works

It's a thin wrapper around the pack's ImageScaleDown logic. The node computes new dimensions by multiplying the input's width and height by scale_by, then runs the resize with a center crop mode - meaning it center-crops to the target aspect before resizing, with Lanczos resampling. Because scale_by is capped at 1.0 (min 0.01), this node only ever shrinks; it won't upscale.

The center-crop behavior is worth knowing: because it scales both dimensions by the same factor, there's no aspect mismatch and usually nothing to crop - the crop only kicks in if your input is a batch of oddly-shaped frames. In practice it's a clean proportional downscale.

The inputs that matter

• images (required) - the IMAGE (or batch) to shrink.
• scale_by (required, FLOAT, default 0.5, range 0.01–1.0) - the factor. 0.5 halves, 0.25 quarters, 1.0 is a no-op.

Output: IMAGE - the downscaled batch, same count as the input.

Install

Part of ComfyUI-Beyond_nodes. ComfyUI Manager → search "Beyond_nodes" → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/beyondprompting/ComfyUI-Beyond_nodes

Restart. No extra dependencies.

Common issues

• It won't go above 1.0. Correct - the max is 1.0 by design. For upscaling use a real upscaler (the KB's upscaling doc is the map for that territory).
• Quality looks soft. Lanczos is a solid default, but if you're downscaling for a final thumbnail that's fine; if you're downscaling as part of a pipeline that re-upscales later, accept the loss or keep the original.
• The output shape changed unexpectedly. Batch inputs are handled per-frame; only width/height change, count stays the same.

Where this node earns its place: preview thumbnails, VRAM-cheap control inputs, or pre-downscaling before feeding an edit model that prefers smaller inputs. It's a one-knob utility - set the factor and move on.