Image Resize By Megapixels (Swwan)

Every model has a resolution it likes. SD 1.5 is happiest around 512, SDXL around 1024, Flux a bit bigger - and the usual way to get there is a resize node where you type the exact width and height, or you do the square-root math in your head. ImageResizeByMegapixels from the ComfyUI_Swwan pack does the math for you: you say "I want about a megapixel," and it figures out the dimensions that get you there while keeping the aspect ratio.

It's part of a growing habit in ComfyUI of thinking in total pixels instead of WxH. When a workflow asks for a "1.0 MP" input, or you're standardizing a dataset so every image has roughly the same information content, this is the node. A 512x512 and a 1024x512 are the same total pixels but wildly different crops; targeting megapixels is the honest way to compare them.

How it works

The math is simple: target pixels = megapixels × 1,000,000, then it works out a scale factor so width × height lands near that total while keeping your chosen ratio. Then it floors the dimensions to divisible_by (16 by default), which is the part that actually matters to diffusion models - most latent spaces want multiples of 8 or 16 to avoid VAE tiling artifacts.

Under the hood it has two paths. If you keep aspect_ratio on "default" and leave it on CPU, it uses a fast direct torch resize that matches the "resize" mode of the pack's ImageResizeKJ v2 node. If you're forcing a new aspect ratio or passing a mask, it hands off to that same node and lets keep_proportion decide how to handle the mismatch.

The inputs that matter

• megapixels - target total pixels. 0 means "skip, return the image untouched," which is handy for a pass-through in a workflow you don't want to disable.
• aspect_ratio - "default" keeps the source ratio; otherwise pick 1:1, 16:9, 4:3, etc.
• keep_proportion - what to do when the new ratio doesn't match: crop cuts, pad fills with pad_color, resize distorts, and pillarbox_blur blurs the background bars.
• divisible_by - floor the output to a multiple. Default 16 is the right call for most pipelines.
• upscale_method - bilinear default is fine; lanczos if you're enlarging and want it sharp.

You also get mask and device as optional inputs. Outputs are the resized IMAGE plus width, height, longest_edge (all INT, handy for feeding other nodes) and a resized mask if one came in.

Installing it

It ships inside the ComfyUI_Swwan pack, which bundles a grab-bag of nodes migrated from LayerStyle, rgthree and KJNodes. Either search "ComfyUI_Swwan" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/aining2022/ComfyUI_Swwan
pip install -r ComfyUI_Swwan/requirements.txt

The requirements are heavy (torch, opencv, scipy, scikit-image, onnxruntime) but you'll already have most of them.

Where people get burned

The "target" megapixels is a floor-and-round, not an exact value - after the divisible-by rounding you can land a few percent under. That's usually fine, but if you genuinely need exact dimensions, this isn't the node. And the pad modes only matter when you're forcing a ratio; with "default" they never fire. The one thing to check: when you enlarge a lot, bilinear gets soft - bump to lanczos or resize in two steps with an upscaler model in between (that's a different job, covered in the KB's upscaling essay).