📐 分辨率输出器 Magic Resolution Output

Every workflow has that moment where you need the width and height in three different places - an Empty Latent, a resolution note, a crop - and you're typing 1024 over and over, praying they stay in sync. Magic Resolution is the value node that ends that: pick a preset from one dropdown and it outputs the width, the height, and a ready-made empty latent in one shot. It's the "one source, many consumers" pattern from the node-plumbing playbook, applied to resolution.

The preset dropdown is the headline feature, and it covers the usual suspects: SDXL's full set of aspect ratios (1024×1024, 1152×896, 1216×832, 1344×768, and the verticals), a couple of 2K options (1024×1536 and its swap), and SD1.5's 512×512. There's also a preset management window so you can add and delete your own resolutions, and the preset data is shared with the pack's Magic Resolution Resize node - set your list once, both nodes see it. The V1.2.5 changelog even added double-click-to-edit on a preset name, which is the small thing that makes preset lists feel maintained instead of abandoned.

Inputs and outputs

• dim_preset - the resolution dropdown; choosing one fills the width/height fields
• width_px, height_px - the actual values used (INT, 64–8192). The preset just fills these; edit them directly for anything custom
• batch_size - how many copies of the latent to make (1–64)

Outputs are width (INT), height (INT), and latent (LATENT). The latent is created at width/8 × height/8 with 4 channels and your batch size - the standard empty-latent shape the VAE expects. So you can wire the width/height into a resolution-dependent node while feeding the latent straight into a sampler.

What to know before you trust it

The latent is an empty latent. It sets the canvas the sampler will fill; it doesn't carry any image content. For txt2img that's exactly what you want - pick SDXL_1024x1024, wire the latent to the sampler, done. The one thing to keep straight: the width/height outputs are in pixels (the real values), while the latent is the downscaled-by-8 version. They're meant to be used for different things, and mixing them up is the classic error - if you connect the latent to something expecting pixel dimensions, everything downstream computes wrong.

Also note the V1.1.3 design decision: the preset directly fills the input fields, the one-click swap button exchanges width and height, and the outputs always use whatever is currently in the fields. That's a simplification that makes the node predictable - what you see in the boxes is what you get out, no hidden state.

Installing

It's part of the Magic-Assistant pack:

cd ComfyUI/custom_nodes/
git clone https://github.com/shigjfg/ComfyUI-Magic-Assistant.git
cd ComfyUI-Magic-Assistant && pip install -r requirements.txt

Restart and it's under "✨ Magic Assistant". No models, no heavy dependencies - this is pure plumbing, so it runs anywhere ComfyUI does.