LtxResolutionSelector

LTX Video 2.3 is genuinely great at what it does, but it has a hard constraint that turns every resolution decision into homework: width and height must be divisible by 32, and the frame count must be divisible by 8 plus 1. Feed it off-grid values and it fails rather than rounding - the model card is explicit about this. LtxResolutionSelector is the node that does that arithmetic for you, and it also handles the sneaky part: when you're using the distilled upscaling workflow, the input resolution you set is not the output resolution you get.

It's a tiny utility node from the ComfyUI-rogala pack, built to feed the pack's own FMLFLTX + SamplerLTXV pipeline, but there's no reason you can't use it to size any LTX 2.3 graph. You set the clip's duration and target resolution; it hands you the exact width, height, and frame length the model wants.

How it works

Three modes, and the distinction matters:

• Dev (no upscale) - picks a standard LTX Dev resolution from a preset list and rounds up to the nearest multiple of 32. If you ask for 1920x1080 (Landscape) you get 1920x1088, because 1080 isn't a multiple of 32 and the model won't negotiate.
• x1.5 Distilled - returns the input size that lands on your chosen target after a 1.5x latent upscale.
• x2 Distilled - same idea for a 2x upscale.

That last pair is the actual value. The LTX-2.3 distilled workflow renders low then upscales in a refinement pass, so if your final target is 1080p you don't feed the sampler 1080p - you feed it roughly half that. Get that wrong by hand and you're either wasting VRAM or upscaling from a resolution that makes the refinement pass pointless.

Inputs and outputs

The inputs are self-documenting: mode (the three options above), dev_target (14 presets for non-upscaled work), upscale_target (8 presets for distilled modes, landscape and portrait), fps (1–120), and duration_sec (1–300) for the clip length.

Outputs are four bare numbers: width, height, length, and a passthrough fps. Frame count follows 1 + 8 × round(fps × sec / 8), which is exactly the divisible-by-8-plus-1 rule. Wire all four into FMLFLTX's width, height, length, and fps pins and you never touch the math again.

Concrete example from the README: 5 seconds at 1920x1080 with x2 Distilled, 24 fps → the node outputs width=960, height=544, length=121, fps=24. The sampler runs at 960x544, upscales to your 1080p target.

Installing it

Part of ComfyUI-rogala:

cd ComfyUI/custom_nodes
git clone https://github.com/rogala/ComfyUI-rogala

Restart, then find it under rogala → Video → LTX Resolution Selector. ComfyUI Manager works too - search "rogala". No dependencies or downloads.

Where people get burned

The trap is assuming the Dev presets are the exact pixel values you asked for. The node rounds up to a multiple of 32, so 1920x1080 becomes 1920x1088 and 1280x720 becomes 1280x736. That's correct behavior for LTX - the model wants valid dimensions - but if you're feeding a fixed-size downstream node that expects exact 1080p, decode and check. And if you switch models mid-project, remember the distilled workflow is a two-stage pipeline: your visible output resolution is the upscale_target, not the width/height the selector emits. Mix those up and the refinement pass in SamplerLTXV will quietly do nothing interesting.