Image: Downscale to Total Pixels (lab)

Image: Downscale to Total Pixels (lab) resizes an image so its total pixel count lands on a target in megapixels - 1 MP, 2.5 MP, whatever you ask for - instead of forcing a specific width × height. If the image is already at or below the target, it's passed through untouched. If it's bigger, it shrinks, cropping a little to keep the aspect ratio rather than stretching.

It's the downscale-only sibling of ComfyUI's core ImageScaleToTotalPixels, and there are three situations where you'll actually reach for it.

One: pre-processing a soft source before an upscaler. The KB digs into this - Comfy Org's own guidance for SeedVR2 is to drop the input to 0.35 megapixels first when the source is out-of-focus rather than genuinely small, on the theory that an upscale from blurry-full-res just keeps the blur. Any upscaler that wants a normalized input benefits from this. Two: normalizing wildly different resolutions before a batch, so every image in a queue hits the model at the same pixel budget. Three: taming a huge image before it blows past your VRAM.

How it works

The target is megapixels × 1024 × 1024 actual pixels. The node computes a scale factor from the square root of target / current, which is what keeps the aspect ratio intact, then resizes with your chosen method and center-crops to the final size. That crop is the deliberate part: the description says it plainly - "image may be cropped, to prevent stretching." Better to lose a strip of edge than to squish everything.

The multiple_of input (default 0, off) rounds both dimensions up to a multiple of that number, which matters when you're feeding a model whose latent space wants multiples of 8, 16, or 64.

The inputs that matter

• image - the tensor you want shrunk.
• megapixels - target size, a float from 0 to 16 in 0.1 steps, default 1. Set it to your model's comfortable budget; 1 MP is a fine default.
• downscale_method - lanczos (default), bicubic, bilinear, area, nearest-exact. For normal photos leave lanczos. For pixel art, nearest-exact keeps the crisp edges; area is the classic "honest" downscale and rarely a bad call.
• multiple_of - 0 to disable, otherwise dimension rounding.

Outputs: the resized image, plus width and height as ints and actual_size - the real MP after cropping, which is usually a hair below what you asked for. The actual_size output is genuinely useful for sanity-checking a pipeline.

Install

ComfyLab Pack is one install for its whole node family (nodes display with a "(lab)" suffix). ComfyUI Manager: search ComfyLab Pack, install, restart.

cd ComfyUI/custom_nodes
git clone https://github.com/bugltd/ComfyLab-Pack.git
cd ComfyLab-Pack
pip install -r requirements.txt

Restart after. No models to download; the pack's requirements (opencv-python, pyyaml, jsonschema, and a few small ones) install cleanly.

Gotchas

• Cropping eats edges. If you can't lose content, letterbox the image first or feed one already near the budget.
• megapixels of 0 is legal and dumb. You'll get a 1×1-ish sliver. Stay at 0.1 or above in practice.
• It never upscales. Smaller-than-target images pass straight through. If you need "make it exactly this size either direction," this isn't the node - that's the core ImageScaleToTotalPixels.

For a node with one job, it's refreshingly hard to misuse. Set a budget, wire it in front of whatever's picky about input size, move on.