LTXResolutions

LTX is the speed tier of local video - where Wan takes minutes, LTX takes seconds - but it has the fiddliest resolution situation of the popular open models. LTX-2/2.3 renders internally at half the requested resolution before upscaling, its VAE grid demands dimensions divisible by 32, and the official workflow is a two-stage dance: generate a low base resolution, then spatial-upscale to the final size. LTXResolutions is a preset picker that does all that arithmetic for you, and it's the smarter sibling in this pack because of one clever trick: when you set an upscale power, it doesn't output the big final number - it outputs the base resolution you should generate at.

How it works

Same skeleton as the pack's WanResolutions: pick an aspect_ratio (seven options, 1:1 through 16:9), pick a resolution tier, get two INT outputs (width, height) that are guaranteed divisible by 32. The presets are tuned to current LTX 2.3 guidance rather than the old 0.9.x sizes - so you get official fast-iteration anchors like 640×640, 512×768, 768×512, 512×704 and 704×512, a balanced tier, and up to 1920×1088 at the top.

Six tiers per aspect: Stage 1 Preview, Fast Iteration, Balanced, HD Output, High Detail, Full HD Output. The naming isn't marketing - it's the two-stage workflow in dropdown form. Stage 1 Preview is your cheap pass to nail prompt and composition, and the higher tiers are what you finish on.

The input that matters most is upscaler_power (options none, x1.5, x2). This is where the node earns its keep. Pick, say, Full HD Output at 16:9 and set x2, and the node returns 960×544 - not 1920×1088. That's the point: you generate at 960×544, upscale 2× with a spatial upscaler, and land on exactly 1920×1088, which is the current official ComfyUI workflow for LTX 2.3. The preset is treated as your final target; the output is the base that gets you there. Set none and it just passes the tier through as-is.

Two more controls:

• image_bypass (default off) - the optional IMAGE input auto-matches the nearest aspect ratio from your source frame, like it does on the Wan node. Flip this on and the connected image is ignored, which you want when the source aspect is misleading but you still want it wired in.
• The upscaler_power combo also accepts a direct connection from the pack's LTX Upscaler Power node, or from any compatible combo switch - so one switch can drive the upscale mode on several LTXResolutions nodes at once.

Installing

Same as the rest of the pack: ComfyUI Manager, search "WanResolutions" (display name "Wan/MiniMax/LTX Resolutions"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/boobkake22/ComfyUI-WanResolutions.git

Restart after. No Python dependencies, no model downloads - pure dropdown logic, Apache-2.0.

Where people get burned

The one real trap: expecting the x2 output to be the big number. If you wire LTXResolutions into a resolution input expecting 1920×1088 and get 960×544, that's not a bug - that's the base resolution for a two-stage render. Generate at the base, upscale 2×, and the pair lands exactly on the target. It reads backward until you've internalized the workflow, and then it's the most useful behavior in the pack.

One honest caveat from the wider LTX story: LTX still loses to Wan on faces, bodies and complex motion, so this node's real job is the draft-then-refine pipeline - cheap LTX passes for iteration, Wan for the final. LTXResolutions just makes the cheap passes cheap in the right shape.