Scale SDXL Image MXD

Here's a common img2img situation: you've got an image at some arbitrary size, you want to work on it in SDXL, and you want it to be about a megapixel - because that's the resolution SDXL was trained at and where it behaves itself. The core ImageScaleToTotalPixels node does exactly this: pick a target megapixel count, keep the aspect ratio, done. Scale SDXL Image MXD is that node with a SDXL-flavored twist - if your image is already at one of the five SDXL-safe resolutions, it passes it straight through untouched rather than resampling for no reason.

Why the "SDXL safe" framing exists

SDXL's five training resolutions (1024², 1152x896, 1216x832, 1344x768, 1536x640) are where the model is happiest, and the community has internalized that going far outside them costs quality. This node encodes two bits of that wisdom: it knows those sizes, and it refuses to nudge an image that's already there. So when you're prep an input before a generation or a detail pass, you get the target-megapixel convenience and a guard against pointless resampling.

How it works

The math is ComfyUI's standard megapixel scaling, which is as clean as it gets:

1. Compute scale_by = sqrt(target_pixels / (width * height)) - a single factor for both axes, so aspect ratio never distorts.
2. Resize with your chosen upscale_method (bilinear default; bicubic and lanczos are the other sane picks).
3. If the input already matches an SDXL-safe resolution (in either orientation), return it unchanged.

One deliberate quirk: the node hard-rejects nearest-exact and area as upscale methods with an error telling you to pick something like lanczos. It's a blunt hint, but a fair one - those two give visibly worse results for this job.

Inputs and outputs

• image - what you're scaling.
• total_megapixels - the target, as a float (1.0 = 1MP). Default 1.0 is the SDXL sweet spot.
• upscale_method - bilinear / bicubic / lanczos (use these) or the two it'll refuse.
• Output: IMAGE, ready to feed a sampler at the right footprint.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Maxed-Out-99/ComfyUI-MaxedOut.git

Restart, or ComfyUI Manager → search "Maxed Out" → install. No model downloads - this is plain tensor resampling, nothing generative.

Troubleshooting

• "nearest-exact/area gives poor results" - that's the node refusing a bad choice, not a bug. Switch to lanczos.
• Image comes back the same size when you expected it to change - it was already an SDXL-safe resolution, so it skipped. If you genuinely need a different size, change the target megapixels.
• Aspect ratio drifts slightly - round() can make the output a pixel or two off from exact. That's normal and harmless for img2img.

One honest caveat: this is a pixel resize, not a detail upscale. If your goal is "make it bigger AND add detail," you want a generative upscaler (SeedVR2/SUPIR territory), not this. For getting an image onto SDXL's home turf, it's exactly right.