Image Scale Range From MP (CRT)

Upscale is a one-way street: once an image is too big for your workflow, scaling it back down loses the point. ImageScaleRangeFromMp is the node that thinks in megapixels instead - you give it a floor and a ceiling, and it normalizes whatever comes in to sit between them. In a batch where inputs vary from tiny thumbnails to wallpapers, it's the difference between a sampler that chokes on mixed resolutions and one that never notices.

The mechanism, and why it's smart about it

It measures the incoming image, and:

• below min_megapixels → upscales to the floor
• above max_megapixels → downscales to the ceiling
• already in range → passes through untouched

The scaling is the part worth reading twice: it resizes both axes by one uniform factor (so no aspect-ratio stretch), then does the final quantization with a centered crop, never a second resize. That crop-first approach means the output stays step-aligned without the softness a second interpolation pass would add. Small downside to know: because it crops down, the final megapixels can land slightly below the bound you asked for.

The inputs

• image - your IMAGE tensor.
• upscale_method - nearest-exact, bilinear, area, bicubic, or lanczos. For downscaling, area is the classic pick; for upscaling, lanczos/bicubic look better than nearest.
• min_megapixels - floor (default 0.25).
• max_megapixels - ceiling (default 1.0).
• resolution_steps - the step to quantize the final dimensions to (default 1, so no quantization). Set it to 8 or 16 if your sampler needs multiples.

Outputs: the resized image, plus width and height as INTs - the pair you can wire into a KSampler's EmptyLatentImage or into CRT's own Resolution (CRT) world.

Where you'd reach for it

Mixed-resolution batches, mostly - a folder of reference images, video frames, or scraped data where every input is a different size and the downstream model has a sweet spot. It's also a handy normalization layer before an upscale pass: clamp inputs to a sane MP range so the heavy model only sees what it's good at. It is not a quality upscaler - it's a range normalizer with good taste.

Gotchas

• Cropping, not stretching. If you need every pixel preserved, this will center-crop edges instead. Know that going in.
• Tiny inputs get upscaled with real interpolation. min_megapixels above the source size means invented pixels, not recovered detail - keep the floor close to your actual content.
• Step 1 = no quantization. The default resolution_steps of 1 is fine for most downstream nodes, but model-specific multiples (8/16/64) are your job to set.

Install CRT-Nodes via ComfyUI Manager (CRT-Nodes) or git clone https://github.com/plugcrypt/CRT-Nodes.git into custom_nodes, install requirements.txt, restart.