LTXV Duration Predictor

Every LTX user has sat there with length set to 97 frames again, wondering if the shot actually wants to be 97 frames. The LTXV Duration Predictor removes the guess: it reads your prompt's text embeddings and predicts the natural duration of the shot in seconds - no sampling, no VAE, no waiting through a generation to find out your clip is either a still life or chopped off mid-motion. It's the "how long should this be" knob, answered by the model itself.

This is a very new node - it landed in ComfyUI core in mid-August 2026 alongside LTX 2.5 support - and it's a surprisingly elegant bit of plumbing. It's built for the current-generation LTX model (Lightricks' 2.4/2.5 era, not the 0.9.x 2B or the 13B), so it also hints at how Lightricks now treats duration as a property of the prompt rather than a number you pick.

How it works

The trick is that the head runs on the caption connector output - the same processed text embeddings the diffusion model consumes - not on the video. In the source, the node calls the model's preprocess_text_embeds exactly the way sampling does, splits the result into video tokens and audio tokens, and feeds both into a tiny network: two projection layers, one cross-attention pooler with a single learnable query, and a small MLP that outputs a number in log space (the exp of a scalar). One forward pass, no denoising steps, done.

That head is genuinely small - a few hundred thousand parameters - so this costs nothing on top of already having the model loaded. What you get back is a raw duration in seconds, which the node then converts using your frame_rate and clamps to [min_seconds, max_seconds] before snapping it to the VAE's causal 8k + 1 frame grid (97, 121, 145…). Snapping floors down, and if that undershoots your minimum it bumps up to the next grid point. The num_frames output is always a valid LTX frame count, which is the whole point.

The inputs that matter

• positive (CONDITIONING) - the prompt. This is the entire basis of the prediction; the head never sees your image, your motion, or your audio. Long, detailed LTX-style prompts give it more to work with.
• duration_head (MODEL_PATCH) - loaded with ModelPatchLoader, which reads from ComfyUI/models/model_patches/. The file is ltx-2.5-duration-head-bf16.safetensors from the Lightricks/LTX-2.5 HuggingFace repo. Plug something else in and the node throws a ValueError - it checks the patch is actually a DurationHead.
• frame_rate (default 24), min_seconds (1), max_seconds (20) - the guardrails. Note the default ceiling is 20 seconds, which caps the frame count hard even if the model "wants" more.
• model - your loaded LTX model; it's what owns the caption connector.

Outputs

Two: num_frames (INT), the clamped, grid-snapped frame count you wire into EmptyLTXVLatentVideo's length input (or any LTX latent creator), and seconds (FLOAT) - the raw, unclamped prediction. That second output is the honest one: if seconds says 34 but num_frames gives you 20 seconds' worth, the model wanted a longer shot than your max_seconds allowed. Watch it when you're tuning the limits.

Getting it

It ships with ComfyUI core - no install. You only need the head checkpoint itself, and it lives in models/model_patches/ (a folder that ModelPatchLoader, still marked experimental, watches). Because this is week-old tech, expect the workflow examples to evolve quickly; check Lightricks' official 2.5 workflows if ComfyUI's example gallery hasn't caught up.

Gotchas

Two real ones. First, the prediction is prompt-only - it can't know your subject is a race car or a sleeping cat, so the duration it picks reflects your text, not your content. Treat it as a smart starting point, not a contract. Second, the version naming is a mess (a recurring Lightricks habit): the node docs call it the "2.4 duration head" while the checkpoint and PR are tagged 2.5. Don't get cute trying to mix it with older LTX models - it's built for the current generation. And if you're on an older LTX, this node simply doesn't apply to you.