👑 FRED Auto Crop Image Native Ratio

Every diffusion model has native resolutions it was trained on, and SDXL especially was trained on a discrete set of aspect ratios. Feed it something off-ratio and you get stretched bodies, doubled limbs, repeated patterns. That's exactly the problem FRED_AutoCropImage_Native_Ratio exists to kill: it takes whatever image you throw at it, crops it to a proper SD/SDXL ratio, and - here's the part that makes it worth reaching for - it can preserve a mask so the crop doesn't lop your subject's head off.

What it actually does

It ships with a baked-in table of 25 standard ratios, from 5:12 portrait (640x1536) up to 12:5 ultra-wide (1536x640), plus square 1:1. You have three broad modes on the aspect_ratio dropdown:

• Auto_find_resolution - it picks the closest native ratio to your input image. Zero thought required.
• A named ratio like 3:4 portrait 768x1024 - force that exact one.
• custom - use the custom_width / custom_height inputs instead.

On top of the ratio crop you get the useful extras. mask_preserve keeps the area under mask_optional inside the final crop, which is the killer feature for faces or products - feed a subject mask in and it slides the crop window so nothing important gets cut. Precrop_from_input_mask is different: it crops the image down to the mask's boundaries first, then applies the ratio crop, which is handy when your source is a big scene and you only care about the masked region. crop_x_in_Percent / crop_y_in_Percent shift the crop window off-center when you want a subject kept at a specific spot.

Then there's the resize half. resize_image turns on scaling to the target dimensions, with separate interpolation choices for upscale (bilinear default) and downscale (area default - that's the right pick, area is what you want when shrinking). multiple_of rounds the result to a multiple of 2/4/8/16/32/64, which matters because many upscale models and some samplers are happiest on multiples of 8. prescale_factor scales before cropping when enabled, useful when the source is small.

The outputs you'll actually use

• modified_image - your cropped (and optionally resized) result, straight into a KSampler or a save node.
• modified_mask - the mask transformed to match the new crop, so downstream inpainting/Detailer-style passes line up.
• scale_factor, output_width, output_height, native_width, native_height - bookkeeping; scale_factor is the one worth wiring if you're feeding an upscaler.
• preview - the crop with a colored mask overlay, so you can eyeball where it's about to cut before you commit.

sd_aspect_ratios is just the list of ratios as text. And like every node in this pack, there's a help output that dumps the whole reference card into a text box.

Installing it

It ships in the FRED Nodes v2 pack, so install once and every 👑FRED node shows up under a 👑FRED menu:

cd ComfyUI/custom_nodes
git clone https://github.com/Poukpalaova/ComfyUI-FRED-Nodes_v2.git

Then restart ComfyUI (or hit Refresh). You can also search "ComfyUI FRED Nodes v2" in ComfyUI Manager. The pack pulls in opencv-python, numpy, Pillow and a git dependency on ocampor/image-quality, so give the install a moment.

Gotchas

Mask inputs need to match the image's spatial dimensions; the mask output only makes sense if you fed a mask in. And remember mask_preserve is a constraint, not a guarantee - if your target ratio is far enough from the source, something has to give, so pick a closer ratio or let Auto_find do it.