LongCat Size Picker

Most models these days accept any resolution inside a megapixel band and just degrade softly when you wander. LongCat isn't one of them. It was trained on a discrete set of aspect-ratio buckets, and LongCatSizePicker exists so you never have to guess which ones those are. It hands you a dropdown of the model's supported sizes and produces the empty latent the sampler expects.

The dropdown has 60 entries, each labeled like 1024 - 1.0 (1024x1024) - that's base-pixel tier, aspect ratio, then width×height. The sizes come straight from the model's MULTI_ASPECT_RATIO_1024, _512, and _256 tables, so what you pick is literally what LongCat saw in training. The 1024 tier is your full-res bucket list (portrait 704×1408 through landscape 1408×704, roughly); 512 and 256 are smaller tiers that matter if you're on a VRAM diet or running edit passes at lower cost.

Under the hood it's doing something almost comically simple: it parses the WxH out of the label and creates a zero tensor shaped [1, 4, H//8, W//8]. That's a placeholder - the sampler overwrites it with noise at the start of the denoise loop, so this node is really a size carrier. Don't be confused if you preview it and see pure zeros; that's the point.

The outputs:

• latent - the empty latent, wire it into LongCatSampler's latent_image socket.
• width / height - the chosen dimensions as ints, handy for wiring into text or into LongCatImageResizer so your reference matches the generation.

There's exactly one required input (size) and no knobs beyond it. That's the whole node, which is refreshing - most of this pack is early-stage rough edges, and this one is just a clean utility.

Worth a workflow note: because the bucket tables are what LongCatImageResizer snaps images into too, the clean pattern for an edit flow is to pick a size here, run your reference through the resizer with keep proportion, and let both agree on the same bucket. Pick an off-bucket size and you're fighting the model instead of working with it.