Resolution Presets

👾 Resolution Presets is the node that exists because "just type 1344x768" stops being fine once you're juggling three model families in one workflow and you've memorized the wrong number. It's a dropdown that hands you the dimensions a given model was actually trained at, so you stop generating 1920x1080 on a model that thinks in 1024 squares and then wondering why the anatomy looks like a melted chess set.

Why this node exists at all

Every architecture has a native resolution, and the closer you stay to it, the less your images fall apart. SD 1.5 was trained at 512x512; SDXL at 1024x1024 with a small set of trained aspect ratios (1152x896, 1216x832, 1344x768, 1536x640, plus rotations); Flux at 1024x1024 and up; Qwen Image at its own higher-resolution set. Go far off those and you get double heads, stretched limbs, and repeated patterns - the single most common resolution complaint in the community is someone feeding a 1024-based model a ridiculous ratio.

This node encodes those numbers so you don't have to. The presets here are the real ones: SDXL and Flux share the classic trained set, SD 1.5 stays in 512-land, and Qwen Image gets its official 1328x1328 / 1664x928 / 1584x1056 family.

How it works

It's a dictionary lookup and nothing more. The source maps each of the four models (Qwen Image, SDXL, Flux, SD 1.5) to a table of aspect_ratio -> (width, height), the node grabs the row you picked, and returns it. No API, no model weights, no VRAM, no Python dependencies beyond the standard library - this is about as safe and boring a custom node as exists.

One trap worth knowing before it bites: the aspect ratio dropdown is the union of every ratio across all models, so it will happily show you "21:9" while SD 1.5 is selected. The node doesn't silently fudge it - it raises an error at queue time listing the ratios that model actually supports. Annoying at first, but it's honest validation, and the error message tells you exactly what to pick instead.

The inputs and outputs that matter

• model - which of the four families you're generating with.
• aspect_ratio - the ratio, shown as e.g. "16:9" or "3:4".

That's the whole input side, and it's the whole point: two dropdowns, done. The outputs are ratio (STRING), width (INT), height (INT), and model_name (STRING). Wire width and height into the Empty Latent Image node that feeds your sampler - that's the 90% use case. ratio and model_name are handy for logging or for routing (pipe them into a text node as metadata, or use them to switch prompts per model).

Installing it

ComfyUI Manager: search for "ComfyUI_resolution_presets" (or "Resolution Presets") and install. If Manager can't find it - it's brand-new and not yet in the official Comfy Registry, only a contributor PR for that exists - clone it instead:

cd ComfyUI/custom_nodes
git clone https://github.com/CypherNaught-0x/ComfyUI_resolution_presets

Then restart ComfyUI. There's no requirements.txt and nothing to download, so this install is genuinely one step.

Where it fits

It's a small utility by an unknown author (new repo, basically no community footprint), so temper expectations: it doesn't upscale, it doesn't pick your sampler, it doesn't fix composition. It just stops you from hand-typing dimensions. If you need to scale a preset - say, run SDXL a bit past native - skip the sibling Resolution Presets (Advanced) node in the same pack, which adds a scale multiplier and a pixel-count output. But for "give me the right size for this model," this is the one you'll reach for, and it's cheap enough to keep in every workflow.