LTXV resolutions mc

LTX Video (LTXV) is genuinely picky about resolution, and that's the whole reason this node exists. The model's defining trait is speed - Lightricks shipped it as the first video DiT that renders faster than it plays back - and that speed is tied to its low native resolution: the launch spec was 24 FPS at 768×512, and the default moved to 1216×704 at 30 FPS from v0.9.6 on. Feed it some random size and you'll either waste VRAM on a resolution it was never trained on or trip the kind of artifacts that make you think the model is broken when it isn't.

LTXVsize_mc is a dropdown of 22 resolutions that are known-good for LTX Video, and it hands back the Width and Height as two integers. That's the entire job, and honestly that's all it needs to be.

How it works

Pick a preset string like 768x512, and the node splits it on the "x", applies swap_hw if you checked it, and returns the two integers. Nothing else happens - no math, no resizing, no model interaction. It's a labeled lookup table, which is exactly what you want a resolution picker to be.

The two inputs that matter:

• resolutions - the dropdown of 22 presets, default 768x512 (the classic budget resolution LTX was born on). The list runs from 1216×704 down to 544×352, covering the sizes the community actually runs, from the modern default down to low-VRAM territory.
• swap_hw (boolean) - flips the pair for portrait, e.g. 768×512 becomes 512×768.

Outputs are Width and Height as integers. Same wiring as any size node: convert an Empty Latent's width/height widgets to inputs and drop these in.

The trap: three presets are broken

Here's where people get burned, and it's a genuine bug in the shipped code, not user error. Three entries in the list - 640×416, 672х384, and 608×384 - were typed with a Unicode multiplication sign (×) instead of an ASCII "x", and one even sneaks in a Cyrillic "х". The node does a plain .split("x"), so picking any of those three throws a ValueError and the run dies. The author never fixed it.

Workaround: just don't pick them. The ASCII siblings nearby (640×384, 608×352) do the same job, or use AspectSize3_mc from the same pack if you want your exact ratio instead of a preset. It's one line in the source - if you're comfortable editing, that x in the string is all it takes to patch.

Where it sits in a workflow

Size node → Empty Latent → sampler, with the LTX Video model loaded downstream. If you're on a 3060 or 4060, the bottom of the ladder is your friend - the LTX ecosystem docs have people running 81 frames at 768×768 on a 3060 in under 30 seconds, which is exactly the "fast iteration" workflow LTX exists for.

Install

Part of the mc_cascade_res pack - install once, get all four nodes. ComfyUI Manager: search mc_cascade_res. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/tester4488/mc_cascade_res

Then restart. No models, no Python dependencies. It does use ComfyUI's newer V3 extension API, so keep ComfyUI updated - an old install won't even find the comfy_api.latest module it imports.

The one real criticism, besides the Unicode bug: it only offers presets, so your exact aspect might not be in the list. For anything off-ladder, the pack's own AspectSize3_mc is the companion node - same install, no preset limitation. Together they cover the two ways people actually pick video resolutions: "known-good size" and "exact ratio I want."