Anima Resolutions

Anima is picky about its canvas, and the difference between a first pass at 1024x1536 and an awkward 1014x1539 is real. It's a 2B diffusion transformer trained up to 1536px on Cosmos-Predict2, and it composes differently depending on where you land - that's not superstition, it's the model. Anima Resolutions is a glorified picker that removes the guessing: two dropdowns, three clean integers out, done.

That's the whole thing, and I mean that literally. It loads no model, changes no sampler settings, and generates nothing. The name is not a lie - it's just a resolution selector that outputs numbers. Which sounds trivial until you've fat-fingered 1280x720 as 1279x721 and watched the VAE punish you for it.

How it works

The node has exactly two inputs: resolution (a tier: 1024, 1280, or 1536) and ratio (about 35 aspect-ratio presets, from 1:1 through 9:7, 4:3, 3:2, 16:9, and the cinematic 21:9). A little JS frontend keeps the two honest with each other - pick the 1024 tier and the ratio dropdown only shows the sizes that exist in that tier, so you can't select a 2048x1152 preset while you're on the 1024 tier by accident.

The outputs are the three integers you'll actually wire into the graph:

• resolution - the tier you picked, as an int (1024, 1280, or 1536)
• width and height - the real pixel dimensions from your ratio choice

Wire width and height into an Empty Latent Image's matching inputs and you're off. Every preset lands on a multiple of 8 (most on 16 or 32 - the cinematic 1280x720 breaks the 32 streak, but that's just how 16:9 works out), so the sizes stay on the grid Anima's latents actually like. No rounding surprises downstream.

Which size should you pick?

The README's rule of thumb is right: aim for 1 to 1.6 megapixels on the first pass. 1024x1024 is the general default, 1344x896 / 896x1344 are the landscape/portrait workhorses at 1.2MP, and 1024x1536 or 1536x1024 are the high-tier options at 1.57MP. Go bigger and you pay double - Anima is slow for its size (over a minute on a 4060, around ten seconds on a 5090), and more pixels means more wall-clock per step.

Here's the take I'd actually hand you: don't default to the biggest tier. One systematic user comparing 1280/1360/1536 found multi-character composition held together best when the first pass was at 1280, then upscaled afterward. Square canvases also encourage Anima to add an extra character you didn't ask for; if you're prompting a group, match the aspect to the subject count. The 1280 tier is frequently the sweet spot, and this node makes switching tiers a single dropdown click so you can A/B without rebuilding anything.

Installation

No models, no heavy dependencies - pyproject.toml declares dependencies = []. It's pure Python plus a small frontend file.

Via ComfyUI Manager: search "Anima Resolutions", install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/cyberdeliaAI/ComfyUI-anima-Resolutions.git

Then restart ComfyUI.

Common issues

The one thing that trips people: resolution output is the tier, not the width. If you wire resolution into a width input you'll get 1024 where you wanted 1344. Grab width and height and leave resolution alone unless a node genuinely wants the tier value.

Also - this node does not resize an image, ever. If you're looking for an IMAGE output on it, you're holding the wrong node; that's what Anima Image Scale By Multiple is for. It's a picker, and that's the point.