Scale To Multiple 🐲

Every diffusion model has a native resolution and a set of dimensions it wants to see. SDXL likes multiples of 64 and roughly a megapixel; video models want multiples of 16 or 32; generate off-grid and you get duplicated anatomy, tiling, or a latent that refuses to cooperate. Scale To Multiple is the node that makes your numbers and your images agree: it snaps a width/height (or a whole image) to the nearest multiple you specify.

What it does

Give it a multiple (default 64), and it rounds a width and height up or down to a multiple of that value. The clever part is it works with or without an image:

• With an image - it scales the image to the snapped dimensions.
• Without an image - it just does the math and outputs the numbers, plus a placeholder image so the graph stays wired.

That second mode is the sleeper feature. Wire the scaled_width and scaled_height outputs straight into a KSampler or latent node and you've got a pure arithmetic helper: "whatever the workflow says, make it a multiple of 64."

If you supply the optional width and/or height inputs, those act as overrides - the image gets scaled to those values (snapped to the multiple) instead of its own dimensions. No image plus overrides = pure number scaling. That's the whole input matrix, and it's flexible without being fiddly.

The controls that matter

• multiple - the snap target. Default 64 (the SDXL sweet spot); set to 1 to effectively disable it.
• rounding_mode - nearest (default), floor, or ceil. Floor matters when you're feeding a VRAM-strapped latent and want to guarantee you never exceed a budget; ceil when you need minimum coverage.
• crop_mode - how the image fills the target box: stretch, center, fill (scale + crop), uniform (fit inside), or uniform fill (fit with padding).
• resize_mode - lanczos is default and the right call for downscaling. There's also bilinear - tensor, which uses torch.nn.functional.interpolate() and looks noticeably different from PIL's bilinear - worth knowing if you've ever seen the same mode render differently between two node packs.
• scale_factor - multiply width and height by this before snapping. Handy for pre-scaling a hires pass.

Outputs

• scaled_image - the resized image.
• scaled_width / scaled_height - the integers. The image-independent mode is where these shine.

Installing it

Same pack, same command:

cd ComfyUI/custom_nodes
git clone https://github.com/DraconicDragon/ComfyUI-RyuuNoodles

Restart or use ComfyUI Manager (search "RyuuNoodles"). No models, no extra deps beyond the pack's color-matcher requirement which this node never touches.

Gotchas

Two things. First: if image, width, and height are all left unconnected, the node raises a ValueError - it needs at least one source of truth. Second: when you run it image-less, the placeholder it outputs is a black image with the text "No Image" drawn on it. That's a stand-in so the graph stays connected, not something to save. It's easy to forget and accidentally save a black PNG if you're routing the image output onward without checking. Beyond that, if your workflow runs through this node a lot and the console logs about "using image width" annoy you, the pack's settings page has a log-level control.