Image Resize

Image Resize is Eclipse's answer to "which resize node do I grab today?" Instead of five nodes for five jobs, it folds them together: constrain by longest side, shortest side, width, height, or total pixel count; override the aspect ratio; pick how to fit (resize, crop, pad, pillarbox-blur, stretch); and snap to a multiple - all in one node. It's the node you reach for when you know what the output should be but don't want to do the arithmetic.

The first decision is scale_to. This is the whole personality of the node:

• longest / shortest - constrain the big or small side. The classic "make every image fit under 1024 on its long edge" for a batch.
• width / height - pin one dimension, keep ratio.
• total_pixels - target a pixel budget, e.g. 1024 here means ~1M pixels (it's in kilo-pixels). Great for normalizing a mixed batch to a consistent cost before a model.
• custom - exact width × height via custom_width/custom_height; 0 means "keep original" on that axis.

Then size carries the value for the chosen mode (1024 by default), and aspect_ratio overrides the ratio (1:1 through 16:9, or original).

The fit modes - where people get lost

fit decides what happens when the target doesn't match the input ratio:

• resize - scale proportionally; output may be smaller than target. No distortion, no bars.
• crop - scale to fill, then crop the excess (with crop_position choosing center/top/bottom/left/right).
• pad - scale to fit, then pad with pad_color (hex).
• pad_edge / pad_edge_pixel - pad with the mean color of the nearest edge, or by replicating edge pixels. Better-looking bars than flat color.
• pillarbox_blur - the modern letterbox: blurred, desaturated, darkened background. This is the one that makes a 16:9 video out of a 1:1 image without hard black bars, and it's worth knowing it exists.
• stretch - distort to the exact target. Almost never what you want, but it's there.

Rounding out the controls: method (lanczos default), divisible_by (default 8 - round output to a multiple, the VAE-friendly guard), and device (cpu default; GPU is faster for large images but lanczos isn't supported on GPU and falls back to bicubic - that fallback is the kind of silent behavior that confuses people comparing sharpness later).

An optional mask resizes in lockstep, so mask and image stay registered. Outputs: image, mask, and width/height ints for downstream arithmetic.

The honest take

If Eclipse already lives in your workflow, this is the one resize node you need - it replaces half a dozen single-purpose utilities. The pair with Image Rescale is: Rescale when you think in factors ("2x"), Resize when you think in targets ("fit under 1024, pad the rest"). This one's the workhorse.

Installing

ComfyUI_Eclipse, r-vage's pack. Manager → search ComfyUI_Eclipse, or:

cd ComfyUI/custom_nodes
git clone https://github.com/r-vage/ComfyUI_Eclipse

Restart; under Eclipse → Image → Transforms. Standard deps; Eclipse targets current ComfyUI (frontend 1.51.2+) and v4.0 removed legacy RvTools_v2 nodes (Migration Tool included).

The classic gotcha: total_pixels is in kilo-pixels, so "1024" is a megapixel, not 1024×1024. And if your GPU output looks softer than CPU output, that's the lanczos→bicubic fallback, not your imagination.