Quick Size (SDXL)

The node that encodes the "use a trained ratio" rule

SDXL's whole trick - and its whole curse - is that it was multi-aspect trained, meaning it's happiest at a specific list of resolutions: 1024×1024, 1216×832, 1152×896, 1344×768, 1536×640 and their rotations. Generate on one of those and you're golden; generate at 1920×1080 and you get the stretched-ribcage SDXL look that's become a running joke. Quick Size (SDXL) from the ComfyUI-QuickSize pack is essentially that list turned into a dropdown that outputs width and height integers. If you run any of the massive SDXL finetune/LoRA/ControlNet archive - Juggernaut, RealVis, Pony, Illustrious - this is the one you'd actually reach for.

Like its siblings, it's a lookup-table node: pick a preset, pick an orientation, get two INTs. And like the SD1.5 node, it uses a 1.5x toggle instead of a megapixels tier. Off, you're on the exact trained-ratio set - 1:1 at 1024×1024, 16:9 at 1344×768, 21:9 at 1536×640, each one a ratio SDXL was literally trained on. On, you jump to the ~1.5MP table: 1536×1536 at 1:1, 2016×1152 at 16:9, 2304×960 at 21:9.

The inputs that matter:

• preset - 1:1, 3:2, 2:3, 4:3, 3:4, 16:9, or 21:9.
• orientation - horizontal or vertical.
• 1.5x - the size booster, default off.

Outputs are width and height INTs, wired into Empty Latent Image.

The honest caveat is the 1.5x tier. At 1:1 it's 1536×1536 - still reasonable, and SDXL finetunes like Illustrious happily handle 1536². But the 21:9 1.5x entry (2304×960, ~2.2MP) is well past anything SDXL was trained on, and the 16:9 1.5x (2016×1152) is pushing it. Fine for Illustration-style finetunes; softer and riskier on the strict realism checkpoints. The trained-ratio purist's workflow - and honestly the correct one - is generate at 1x, then upscale or hires. The toggle is there when you want to gamble.

Installing it

Same featherweight story as the rest of the pack - no models, no extra dependencies, just Python:

cd ComfyUI/custom_nodes
git clone https://github.com/rh6423/ComfyUI-QuickSize.git

Restart ComfyUI. Or ComfyUI Manager → search "QuickSize" → install.

Verdict

This is arguably the pack's best node, because SDXL is where the "remember the trained ratios" problem is sharpest - there are seven of them and nobody keeps them straight. The 1x table alone is worth the install if you swap checkpoints between realism and anime finetunes, since they share the architecture but punish off-ratio sizes differently. Just treat 1.5x as the "live dangerously" button the author's announcement thread calls it, not a recommendation.