ModelSamplingLTXV

If you've ever opened the official LTX Video workflow, this node is the quiet one in the middle of it - and it's doing more than it looks like. LTX is a flow-matching model, which means it has the same timestep-shift knob as Flux, except LTX doesn't just need it set per-resolution; it needs it per-video, because the token count changes with both spatial size and length. ModelSamplingLTXV computes that shift for you, and if you hand it your latent it does it exactly right.

The mechanism is nearly identical to ModelSamplingFlux - same linear interpolation between a base_shift and max_shift, over a token range - but tuned for LTX's numbers. The defaults are max_shift 2.05 and base_shift 0.95, interpolating between 1024 and 4096 tokens. The clever bit is the optional latent input: when you wire in the latent you're about to sample, the node reads its spatial dimensions (height × width in tokens) and picks the shift from the real thing. Leave the latent disconnected and it assumes 4096 tokens, which is right for a 768×512-ish frame and wrong for anything bigger.

The input that changes everything

• latent - feed it your actual latent. This is the "just works" move: the node follows whatever resolution and frame size you use, so your shift never goes stale when you switch from a draft to a final render.
• max_shift / base_shift - the defaults are the model's own training values. Leave them unless you're chasing a specific look; they're not a quality dial in any intuitive sense.

One MODEL output, straight into the sampler. That's the whole graph side of it.

Where it came from

The node landed with LTX-Video support itself, on November 22, 2024 - day one, alongside the LTXVConditioning and LTXVScheduler nodes that make up the rest of the official setup. LTX's whole identity is speed on consumer hardware, and the default workflow is carefully tuned around that; this node is the resolution-correctness part of the tuning.

The gotchas worth knowing

LTX is famously finicky about resolution - the model was trained around specific base resolutions, and the community consensus is to stick close to them (768×512-ish, or the newer 1216×704 at 30fps from 0.9.6) rather than push wild aspect ratios. This node doesn't fix that; it fixes the shift so that at whatever resolution you do choose, the sampling trajectory isn't silently wrong. If output comes out looking off after you've changed a video's size, the first thing to check is whether the latent is actually connected here. It's the kind of node you never think about when it's wired right - which is exactly the point.