Upscale By Size

Ask any regular about upscaling and the first thing they'll say is: there are two different jobs hiding behind the word. One is "add detail" - that's a generative upscaler like an ESRGAN model or SUPIR, expensive and hallucination-prone. The other is "make it bigger with no new detail, and make it exactly this size" - that's a resize, and it's what ◎ Radiance Upscale By Size is for. It's the deterministic, instant, free-in-spirit end of the spectrum.

This is the node you reach for when a deliverable needs to be, say, 2048x2048 and your render is 1832x2011. It gets you to a target dimension with the aspect ratio respected, using a proper high-quality kernel, in linear light, without invoking a diffusion model. If you only remember one thing from the KB's upscaling essay, it's this: don't throw a generative pass at a problem a Lanczos resize solves in milliseconds with zero risk of inventing a face.

How it works

You give it a target width and height. If maintain_aspect is on (default), it computes a fit inside that box using aspect_mode:

• fit - the whole image fits inside the box; one side may come out shorter than the target (pillarbox/letterbox shape).
• fill - the image fills the box, overflowing on the shorter side (you'd crop the excess yourself).
• stretch - forces exactly target W×H regardless of aspect (distortion risk, but sometimes that's the brief).

The method dropdown is the kernel pick: lanczos, lanczos4, bicubic, mitchell, catrom, hermite, gaussian, bilinear, nearest. For upscaling, Lanczos or Lanczos4 is the standard recommendation - sharp without the ringing you get from aggressive bicubic. There's also a sharpening amount (default 0.2) applied after the resize.

Where this node gets geeky and good: process_in_linear (default on) does the resize in linear light rather than gamma-encoded sRGB, which keeps highlights honest, and input_color_space (sRGB / Linear / Auto) tells it whether to convert first - set it to Linear or Auto when feeding HDR or already-linear input so you don't double-convert.

The inputs that matter

• width / height - target dimensions, 64–16384, stepped to 8px.
• method - the interpolation kernel; lanczos or lanczos4 for clean upscale.
• maintain_aspect + aspect_mode - the trio above that decides fit vs fill vs stretch.

Outputs: upscaled_image plus final_width and final_height - the actual numbers, which differ from your target when maintain_aspect is on, so you can wire them into downstream nodes that need to know the real size.

How to install it

It's in the radiance pack. ComfyUI Manager → "Radiance", or:

cd ComfyUI/custom_nodes
git clone https://github.com/fxtdstudios/radiance.git
cd radiance
pip install -r requirements_windows.txt

Restart, and find it under FXTD Studios/Radiance/Image.

Common issues

Two classic stumbles. First, expecting it to add detail: it won't, and that's correct behavior - if you need detail, this is the wrong node (the pack has Radiance AI Upscale and Radiance Pro Upscale for that). Second, the process_in_linear default means feeding it already-linear data without setting input_color_space to Linear/Auto can double-convert and slightly shift your colors - it's the one setting worth checking when output doesn't match input. And because it resizes, not regenerates, upscaling past 2x with a kernel shows its limits fast: at big factors you'll see softness, which is exactly when you'd switch to a generative pass or tiled diffusion with ControlNet.

For exact-size delivery, it's the one I'd reach for. Cheap, precise, and it leaves your pixels alone otherwise.