🅛🅣🅧 LTXV Extend Sampler

LTX makes short clips. That's the whole trade - you get a few seconds in a few seconds, which is glorious for iterating and annoying when you actually want length. This node is the fix: it generates a fresh batch of frames that continue an existing latent, so you can chain segments into something longer than one pass gives you.

How it works

Instead of denoising a clip from scratch, the Extend Sampler seeds the new frames with the tail of your existing video. It reuses the last frame_overlap frames as context, then denoises num_new_frames of new content that grows out of them. Run it once and your clip is longer; feed that result back in and it's longer still. Because LTX drafts in seconds rather than the minutes Wan takes, iterating this way is genuinely cheap - that speed is the reason the whole extend workflow is worth doing at all.

It's a custom-sampling node, which is why the input list looks heavy. It wants the SamplerCustomAdvanced trio - a guider, a sampler, and sigmas - plus a noise source, on top of the model, vae, and your latents.

The inputs that matter

• latents - the clip you're extending. The tail of this becomes the seed for the new frames.
• num_new_frames (default 80) - how much video to add on this pass.
• frame_overlap (default 16, and 16 is the minimum) - how many frames of the old clip carry over as continuity context. More overlap means a smoother seam but fewer genuinely new frames per pass.
• strength (default 0.5) - how much freedom the new frames get to diverge from the seed. Too low and motion stalls into a slow drift; too high and continuity breaks at the joint.

There are optional optional_cond_images / optional_cond_indices / cond_image_strength inputs if you want to pin a specific frame inside the extension. The outputs are denoised_video (a LATENT, send it to your VAE decode), and positive / negative conditioning passed straight through so the next node in the chain has them.

Installing it

It ships in the official Lightricks pack, so grab the whole thing:

• ComfyUI Manager - open the Manager, Install Custom Nodes, search LTXVideo, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo, then restart.

You'll need an LTX checkpoint and its VAE (they download on first use). Budget the disk - the LTX-2 family wants 100GB+ free and a 32GB-class card is the comfortable target, though quantized builds run on far less if you've got 64GB of system RAM.

Common issues

A visible seam or color shift where segments join. That's not enough shared context - raise frame_overlap. It costs you new frames per pass but the transition smooths out.

Motion goes limp across the join. Nudge strength up. At 0.5 the new frames sometimes inherit the last frame's stillness and just pan.

Quality drifts the longer you chain. Each extension inherits the previous one's flaws and adds its own, so error accumulates. If you're going long, generate segments and clean up the worst joins in post rather than chaining ten passes blind. And mind LTX's grid rules - on 2.3, frame counts must be divisible by 8 plus 1, and off-grid values fail outright rather than rounding.