Model Sampling LTXV (Ray)

LTX Video is the speed-first model in the video stack - the one you reach for when you want ten generations in the time Wan does one. But it's also a flow-matching model, which means it wants its shift set correctly, and it's resolution and token-count aware in a way most nodes aren't. RayModelSamplingLTXV is the LTXV-specific sampling patch for Raylight, and it has one trick the other Ray sampling patches don't: an optional latent input.

Here's what that latent input buys you. When you leave it disconnected, the node assumes 4096 tokens and computes shift from that. When you wire in the actual LATENT you're about to sample (i.e., the one from your EmptyLatentVideo node), it counts the real tokens and sizes the shift to your actual resolution and frame count. For a video model where resolutions and lengths vary run to run, that's the difference between a schedule that's right and one that's guessed. Wire the latent in - it's one connection and it costs nothing.

The other two inputs are the shift pair you've seen on Flux and SD3:

• max_shift (default 2.05) - how much denoising effort concentrates on the early, structure-forming steps.
• base_shift (default 0.95) - the effort floor at the detail end.

LTXV's defaults skew higher than Flux's, which matches its character: it's a model that needs its composition locked in early or it drifts. The 0.9.x → 2.x lineage in the KB's LTX guide all treat shift as a real lever, so don't treat these as sacred - but do treat them as starting points rather than things to zero out.

Where it fits

After Ray Init Actor, before the Ray KSampler, in any LTXV workflow running through Raylight. It's a distributed port of ComfyUI's LTXV sampling patch, applied per-worker. If you're on a single GPU, use the built-in node - this one exists because the model lives on Ray workers.

Install

Part of raylight, no extra models:

cd ComfyUI/custom_nodes
git clone https://github.com/komikndr/raylight
cd raylight
pip install -r requirements.txt

Restart, or grab it via ComfyUI Manager.

Gotchas

The whole game is the latent input. Leave it unhooked and you get the 4096-token default, which is roughly right for a 512×512-class latent but visibly off for long or high-res sequences. Also note this node lives under Raylight/extra, not the LTXV guider section - it patches sampling, and the guiders (STG, Multimodal) are a separate family you'll also need for a complete LTXV-Raylight graph. Don't conflate the two; you usually want both.