Resampler | Mitchell-Netravali

Resampler | Mitchell-Netravali is the filter you want when Lanczos is too sharp and Triangle is too soft - which, it turns out, is a lot of the time. It's the famous cubic family from the 1988 Mitchell & Netravali paper, the one that image libraries like ImageMagick and many game engines use as their balanced default. It sits between bilinear and Lanczos on the sharpness scale and has the genuinely useful property of not ringing on hard edges.

Here's the mental model for choosing between this and Resampler | Lanczos. Lanczos is sharper but can draw faint overshoot halos along high-contrast edges - the "ghosting" you sometimes see on text or architecture after an aggressive upscale. Mitchell-Netravali trades a little of that edge pop for a smooth, artifact-free result. For faces, organic subjects, and anything where a halo would be noticed, it's often the better-looking upscale. For crisp text and graphics, Lanczos wins.

How it works

This one takes two inputs instead of a radius:

• b (FLOAT, default 0.33) - the "blurring" coefficient.
• c (FLOAT, default 0.33) - the "ringing" coefficient.

The b=1/3, c=1/3 default is the famous "Mitchell" setting that maximizes smoothness with no ringing - that's the pair most people mean by the name. The other landmark you'll see in the wild is b=0, c=0.5, which reduces to Catmull-Rom, a sharper variant that does ring a bit. You can sweep between those by nudging the sliders; for a beginner, the takeaway is simple: leave both at 0.33, and only touch them if you want more sharpness (raise c) or more smoothness (raise b). Output is a resampler wire for Resample Image, Resample Mask, or Resample Latent.

Where it fits in this pack

This pack's resampling family is a small ladder, and Mitchell-Netravali is the middle rung: nearest (hard), triangle (softest), Mitchell-Netravali (balanced), Lanczos (sharp), Jinc-Lanczos (sharpest, slowest), area (downscaling). If you're doing a lot of organic-image upscaling and you find Lanczos leaves halos, this is the rung to settle on. It's also a great general-purpose pick when you're resizing once and can't be bothered to compare variants.

Installing

It's one of the six resamplers in comfyui-some-image-processing-stuff:

cd ComfyUI/custom_nodes
git clone https://github.com/wmpmiles/comfyui-some-image-processing-stuff

Restart ComfyUI, or install via ComfyUI Manager by searching "Some Image Processing Stuff". No models, no extra pip deps. And as with every node in the pack, wire its resampler output into a Resample node or nothing happens - these filter nodes are ingredients, not results.