Multi Resize

Upscaling in ComfyUI usually means one of three things: a hi-res fix pass, a dedicated upscaler model, or just making the tensor bigger so a node will accept it. Multi Resize is that third kind - it resizes images, latents, and masks by a scale factor with your choice of interpolation mode. The README sums it up: it's like LatentUpscale but you give it a multiplier instead of absolute width/height.

What it is

A scale-factor resize across all three tensor types at once. scale runs 0.01–10 (default 2), and mode picks the interpolation: area, bicubic, bilinear, nearest, nearest-exact. The output dimensions snap to multiples of 8 (via a center crop after resizing), which keeps downstream samplers and VAE decoders happy.

The mode you actually want

The default is nearest-exact, which is a fair choice for latents - it avoids the weird color fringing that bilinear/bicubic can introduce when resampling a compressed latent space. That's also why core ComfyUI's LatentUpscale defaults to "nearest-exact" by area. The practical guidance:

• Latents: nearest-exact or area - interpolation modes that don't invent structure. This is the "safe" resize for before a sampler.
• Images you'll look at: bicubic or bilinear - smooth, no pixel edges. area is good for downscaling (it averages, so it anti-aliases).
• Masks: nearest - hard edges, no fuzzy antialiasing bleeding into the mask.

So yes, you'll want to switch modes depending on what you're resizing - the default isn't wrong for latents, but it's not what you want for viewing an image.

Why you'd reach for it

The classic use is the pre-sampler upscale: take a 512×512 latent, upscale 2× before a second-pass sampler at higher resolution (the "latent upscale before resampling" pattern that's lighter than a full img2img pass). Because it handles images and masks in the same node, it's also great for keeping a whole pipeline aligned - crop, resize, and feed everything at matching dimensions before a KSampler or ControlNet.

The center-crop-to-multiple-of-8 step means the output isn't exactly scale × in edge cases - a 100×100 image at 2× lands at 192×192, not 200×200, because 200 isn't divisible by 8. For nearly all downstream uses that's the correct behavior (samplers want multiples of 8); just don't expect pixel-perfect scale factors.

Install

Part of Zuellni/ComfyUI-Custom-Nodes, via ComfyUI Manager (search "Zuellni") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Zuellni/ComfyUI-Custom-Nodes

No extra requirements, no downloads - pure torch interpolation. The pack is archived but this is dependency-free, stable code. If you need aspect-ratio-aware resizing, per-type modes, or exact dimensions, core nodes plus this cover it; if you want the scale-factor convenience in one node for all three tensor types, this is the one.