Image Rescale

Image Rescale does one job and does it with options: take an image, make it a different size, either by multiplying its dimensions (rescale mode) or by forcing an exact width × height (resize mode), and optionally run a supersampling pass so the result looks better than a plain resize.

This is the "more pixels, no invented detail" lane of the KB's upscaling doc - job one, not job two. It won't add texture or fix softness; for that you want an ESRGAN or SeedVR2-class model. What it gives you is fast, clean geometric resizing with filters tuned for the job.

The settings that matter

• mode - rescale multiplies the current size by rescale_factor (default 2); resize targets resize_width × resize_height (defaults 1024, rounded up to the nearest 8).
• resampling - the interpolation filter. Default lanczos is the quality pick for enlarging. The tooltip's tip is the one to remember: area is ideal for downscaling video frames - it averages pixels instead of sampling them, so downscaled video doesn't shimmer or alias.
• supersample (default on) + supersample_factor (2x–8x) - this is the interesting bit. The node upscales to an intermediate that's factor× larger than the target, then downsamples with your chosen filter. That two-step dance beats a single direct resize for anti-aliasing - it's why "resize in one step" images look slightly jaggy and these don't. Cost: a bigger intermediate, so more VRAM and a bit more time. 8x is overkill for most; 4x is the sweet spot if you care, off if you don't.

The node also accepts a list of images (the schema is list-input enabled), and it's explicitly ported from WAS Node Suite (MIT, original author WASasquatch) - so if you've used WAS's Image Rescale, this is the same behavior in the Eclipse universe.

Where it fits

Slot it in front of anything with a size requirement, or use it to normalize a batch to a common size before a concat or a model that wants fixed input. It's the "get to the right pixel count cheaply" node - the KB's post-processing point again: deterministic, instant, no generation.

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 wants current ComfyUI (frontend 1.51.2+); v4.0 removed legacy RvTools_v2 nodes with the Migration Tool for old saves.

The common mistake: leaving supersample on 8x and wondering why a batch resize is slow. Supersampling only pays when quality matters - for previews or intermediate steps, turn it off. And remember resize mode snaps to multiples of 8, so don't be surprised if your 1000px request comes back 1000 or 1008.