Best-Res (area+scale)

The hi-res fix is the most reliable way to get a big, clean image: generate at the model's native resolution, then do a second pass at 1.5–2x with low denoising to add detail. The fiddly part isn't the concept - it's the arithmetic. Both the initial resolution and the upscaled resolution need to stay divisible by your step (8, or better, a bigger planning step), and if you pick the initial size badly, the upscaled size ends up in a latent-incompatible hole and you're cropping or outpainting to survive. Best-Res (area+scale) is the pack's answer: one node that plans both sizes together so the whole chain stays clean.

It's the Best-Res (area) node plus an upscale section. You give it the model's native square size, aspect ratio, orientation, and your target upscale factor - it outputs the init width/height and the HD width/height, plus the exact upscale value to use.

How it works

Same core as the area node: normalize your aspect ratio to unit area, multiply by square_size to get the float init dimensions, then multiply by upscale to get the float HD dimensions. The difference is the priority setting, which decides what gets rounded first:

• original (default) - init size is rounded first, then the HD size is derived from the rounded init and rounded again. The HD image follows the init's ratio most faithfully, but may drift furthest from your desired aspect ratio.
• desired - both sizes are rounded independently toward the desired ratio. Closest to what you asked, but the two sizes' ratios can differ the most from each other, which can force crop/pad later.
• upscaled - HD size is rounded first, and init is back-tracked from it. Pick this if the final HD dimensions are what you truly care about.

After rounding, the node checks whether the HD size can be reached by uniformly scaling the init size. If yes, its upscale output is the precise factor (which may be 1.4783 instead of a clean 1.5 - that's fine). If no, it outputs your original upscale value untouched, and the status readout shows a ⚠️ - that's the signal to route things through the Upscaled Crop/Pad node rather than scaling blindly.

The inputs you actually set

• square_size - model's native square: 1024 for SDXL/Flux, 512 for SD 1.5.
• step - init-side divisibility, default 48.
• HD_step - upscaled-side divisibility, default 144 (= 8 × 2 × 3 × 3, so a 1.5x init upscale stays clean and the next 1.333x after that stays clean too).
• upscale - your target HD-fix factor, default 1.5.
• landscape, aspect_a, aspect_b - ratio and orientation, same as the area node.
• priority - the rounding order above.

Outputs and the workflow it feeds

• init_width, init_height → your first KSampler / Empty Latent Image.
• upscale → the exact factor for the latent upscale (or image upscale) between passes.
• HD_width, HD_height → your second KSampler - and into Upscaled Crop/Pad alongside the init sizes, so it can tell you whether crop/pad is needed and how much.

The intended flow is: generate at init size → upscale by the upscale output → optionally crop/pad per the companion node → second KSampler at HD size with low denoise. The node's on-face status shows both sizes with their step counts and the ✅/⚠️ uniformity verdict, which is genuinely useful while you're tuning.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Lex-DRL/ComfyUI-BestResolution

or search "Best Resolution" in ComfyUI Manager, install, restart. Only frozendict as a dependency, no model downloads. Like the rest of the pack it's pure arithmetic - it computes, it never draws. Which is exactly why the numbers it gives you are worth trusting: no VRAM, no model quirks, just math that stays clean at every step.