Resolution Master

If you've ever generated an image with a stretched body, two heads, or a weirdly repeated horizon, you already know the problem this node solves. Every model was trained at a native resolution and a handful of aspect ratios, and when you wander off those, it falls apart. The most-upvoted answer to a thousand "why does my person look melted" posts is blunt: "You are using a model trained on 1024x1024 images trying to make images with ridiculously different aspect ratios." Resolution Master is a visual, all-in-one dimension picker that keeps you inside the lines - and hands you an empty latent on the way out so you can delete one node from your graph.

What it actually does

Here's the thing to get straight up front: Resolution Master does not touch your pixels. It's a calculator with a nice interface. You drag a dot around a little 2D canvas or pick a preset, and it outputs numbers - a width, a height, a rescale factor, and a ready-made latent. Those numbers go into the nodes that do the real work. The README says it plainly, and it's worth repeating because people trip on it: "Resolution Master doesn't process images directly - it calculates dimensions." If you want to actually resize an image, you still wire its width/height into a Resize Image node.

Where it earns its keep is the preset library and the model-specific math. It ships buckets for SDXL, Flux, Flux.2, WAN video, Qwen-Image, Z-Image Turbo, plus print, social, and cinema ratios. SDXL's real trained ratios - 1024×1024, 1152×896, 1216×832, 1344×768, 1536×640 and their rotations - are baked in as presets, so you stop guessing. Turn on Custom Calc and it'll clamp whatever you've got to something the model is happy with: Flux mode, for example, forces 32px increments and caps you at 4 megapixels. It's the difference between "I think 1216×832 is legal for SDXL?" and just clicking the button.

The inputs and outputs that matter

Most of the controls live in the on-node widget - the canvas, the preset dropdowns, the buttons - so you rarely type into fields directly. The handful worth knowing:

• input_image (optional) + auto_detect: connect a Load Image and flip Auto-Detect on, and the node reads that image's dimensions as your starting point. This is the resize-workflow entry point.
• batch_size: how many latents to make (1–4096). Bump it if you want a batch at one resolution.
• latent_type: leave it at the default latent_4x8 for SD, SDXL, and Flux. Only switch to latent_128x16 if you're on Flux.2, which uses a 128-channel latent.

On the output side you get five slots. width and height go to your checkpoint/sampler (or a resize node). latent is the sleeper feature - it's a correctly-sized empty latent built from your width, height, batch, and type, so you can wire it straight into KSampler and drop the separate Empty Latent Image node entirely. batch_size passes through. And rescale_factor feeds an upscaler.

One gotcha on rescale_factor: it is not the ratio between your input and current size. It's your scaling intent for downstream upscalers. Set a 1080p target and swap in a different input image, and the factor recalculates so you still land on 1080p. That's by design - it makes your workflow resolution-independent - but it surprises people who expect it to reset to 1.0.

Installing it

Easy one. In ComfyUI Manager, search Comfyui-Resolution-Master and hit Install, then restart. Or do it by hand:

cd ComfyUI/custom_nodes/
git clone https://github.com/Azornes/Comfyui-Resolution-Master.git

Restart ComfyUI and it's there under utils/azToolkit. No models to download, no heavy Python dependencies - it's a JavaScript widget plus a bit of backend math, so the install is genuinely painless.

When it breaks

Two known issues, both from the author and both quick fixes. If the node shows up completely blank on the canvas, that's a conflict with comfyui-mixlab-nodes - mixlab overrides a drawing method other nodes rely on. Disable or uninstall mixlab and it renders fine (there's a one-line manual patch in the README if you can't live without mixlab). Second, it does not work with ComfyUI's newer Vue Nodes rendering; if the node looks broken, go to Settings, search "Vue Nodes," and turn off "Modern Node Design (Vue Nodes)."

And the recurring confusion, not a bug: it outputs numbers, not images. If your picture didn't change size, check that you actually wired width/height into a resize node downstream.