Qwen 2512 t2i resolutions

When you run a text-to-image model at a resolution it wasn't tuned for, you get either noise or a picture that quietly ignores your prompt. Every modern model has a short list of resolutions it actually likes - Qwen 2512's text-to-image pipeline is no exception - and the list isn't printed anywhere obvious. SBL_Qwen2512ResT2I is a dropdown that contains it: pick a preset, get Width and Height integers out, wire them into your Empty Latent (or resolution input), done.

The presets are the model's own recommended buckets, and you'll recognize the shape of them if you've worked with the Qwen-Image family: a ~1.8–2.5 MP sweet spot that the 20B base was trained around, kept out of the "just use 1024x1024 and pray" territory that wastes the model's larger native canvas. The full list:

• 1328×1328 (1:1)
• 1664×928 (16:9) / 928×1664 (9:16)
• 1472×1104 (4:3) / 1104×1472 (3:4)
• 1584×1056 (3:2) / 1056×1584 (2:3)

Seven options, all of them divisible by 8, so they slot cleanly into latent space without extra snapping - the node returns the exact dimensions, no rounding.

How it works

There's no real computation here, and that's the point. The node holds a fixed map of display string → (width, height) in its source. You pick 1328×1328 (1:1) from the resolution enum and it hands you 1328 and 1328 out the Width/Height INT sockets. Zero math, zero chance of typoing 1468 when you meant 1472, and no need to remember which of these the model was actually trained on.

This is the useful edge of the resolution-plumbing problem: models change, their native resolutions change with them, and packs like this exist to pin the correct set to a dropdown so your workflow doesn't silently drift.

Install and notes

Same story as every SBL node - it ships in the one pack:

cd ComfyUI/custom_nodes
git clone https://github.com/SoftByteLabs/SBL

then restart. Or use ComfyUI Manager and search SBL. No pip dependencies, no model downloads; the node is just a lookup table, so there's nothing to break.

One honest caveat: a preset is a starting point, not a law. If you're upscaling or doing an img2img pass, you'll still reach for a dedicated resolution/upscale node after this one. But for a first-pass text-to-image at the native sweet spot, this dropdown is the whole answer - and if you're on the broader SBL pack, its sibling Resolution node carries these same Qwen 2512 presets inside a much bigger general-purpose dropdown if you want everything in one place.