PDImageResize

Core ComfyUI's ImageScale will stretch an image to any size you ask, ratio be damned. PDImageResize is the node for when "any size" isn't good enough - it gives you four distinct resize behaviors, interpolation choices, an optional multiple_of constraint to keep dimensions latent-friendly, and it hands back the final width and height so the rest of your graph knows what actually happened.

If you've ever resized a batch and then discovered the images came out 1022×1026 instead of 1024×1024 (and your model grumbled), this node's multiple_of is the thing you didn't know you wanted.

The four methods (the part that matters)

• stretch - force to exactly width × height. Distorts the aspect ratio. There are better nodes for most jobs, but sometimes you genuinely want the distortion.
• longest size - landscape images scale to width (height follows), portrait scale to height (width follows). The aspect-ratio-preserving default for most use.
• fill / crop - scale up so the image covers the target, then center-crop the overflow. This is the "cover" behavior you know from CSS and social platforms: no distortion, no empty bars, but content gets cut.
• pad - scale to fit inside the target, then pad the leftover space with black to hit width × height exactly.

On top of that, multiple_of (default 0 = off) rounds the target dims down to a multiple of your number - set 8 and you're guaranteed 1024, not 1022. This is the knob that keeps VAE and latent paths happy, and it's worth turning on by default.

The inputs that matter

• image (IMAGE) - source.
• width / height (INT, default 512) - target. 0 means "keep original" for that axis.
• interpolation - nearest / bilinear / bicubic / area / lanczos (defaults to bilinear in code; for upscaling prefer lanczos, for downscaling area).
• method - one of the four above.
• multiple_of (INT, default 0) - force output dims to a multiple of this.

Outputs: IMAGE, plus width (INT) and height (INT) - the actual output size, which differs from your request whenever longest size/pad/multiple_of adjust things. Wiring those back out is the tidy part: you can read the real dims for a filename or downstream sizing instead of guessing.

Where it fits

Standardizing batches for model input is the headline job: pick pad + multiple_of=8 and every frame comes out exactly 1024×1024 with ratio intact and bars added. Social-platform covers, thumbnail grids, and "make everything fit this box" jobs are its natural territory. lanczos is the interpolation to reach for when you're upscaling and quality matters; area when you're shrinking a 4K render down.

Gotchas

The longest size semantics are subtle: it uses width for landscapes and height for portraits, so if you set width=1024, height=1024, a 2:1 image becomes 1024×512 and a 1:2 becomes 512×1024. That's usually what people want, but if you expected both to be 1024 on their long edge, you'd actually want PDIMAGE_LongerSize. And the pad mode always pads with black - if you need a configurable background color, PDImageResizeV2's letterbox fit has a background_color input for that.

Install

ComfyUI Manager: search PDuse (repo 7BEII/Comfyui_PDuse), install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/7BEII/Comfyui_PDuse
cd Comfyui_PDuse
pip install -r requirements.txt

Restart and it appears as PDImageResize under PDuse/Image. Four behaviors, real output dims, and a multiples safety - the workhorse resize.