LTX-2.5 Multishot Sampler (Joy-LTX 2.5)

The LTX-2.5 Multishot Sampler is this pack's other engine lane: it runs the LTX-2.5 two-pass pipeline once per shot and joins the shots with an AV-extend - the previous shot's tail pinned as raw latents at the head of the next - so a take can be as long as you like, and even a hard cut keeps the voice. It's the whole "multishot" idea applied to Lightricks' speed-first video family rather than to MiniMax H3.

LTX's role in the ecosystem is speed: where H3 (and Wan, for that matter) are minutes-per-clip affairs, LTX drafts in seconds. That makes it the right tool when you're iterating on a long scene or want a fast version of a take before committing to the H3 render. The node is a self-contained sampler - model, clip, both VAEs, prompts, and all the dials are on the node itself - so you can build an LTX lane without any of the H3 chain machinery.

How it works

For each shot it runs LTX's two-pass pipeline: pass one at the requested resolution, then (with two_pass on and an upscale_model connected) an upscaled refine pass on the output grid. Then it joins:

• continue (default) - AV-extend: the previous shot's tail (video and audio, raw latents) is pinned at the head of the next via a noise mask, so the take reads as one continuous shot and the voice carries across.
• cut - the voice extends but the picture is fresh, like a hard cut in a scene where the dialogue continues.
• fresh - independent shots, no carryover.

The overlap input (default 3, i.e. 17 pixel frames ≈ 0.7 s) sets how much of the previous tail gets pinned - more overlap means a smoother join but less new content per shot.

The inputs that matter beyond those: prompts (the writer's {"prompts":[...]} JSON or ----separated blocks), frames_per_shot (8n+1 at 24 fps; 193 = 8 s), shot_count (0 = every prompt the writer produced), the two sigmas_pass1/sigmas_pass2 schedules (the distilled 8-step defaults are the measured recipe), and video_cfg/audio_cfg (1.0 is the LTX norm). Optional inputs include start_image for I2V on shot 1 and shot_images - one image per shot used as the first frame in cut/fresh mode, which is how you carry identity from your reference plates.

Outputs are images, audio, and an info string.

Installing it

Part of the ComfyUI-H3-Multishot pack:

cd ComfyUI/custom_nodes
git clone https://github.com/jlucasmcrell/ComfyUI-H3-Multishot

Or search H3 Multishot in ComfyUI Manager. Requires ComfyUI v0.30.0+. You'll need the LTX-2.5 model, its text encoder, and the video/audio VAEs; two_pass also needs a latent upscale model. The pack itself adds no Python dependencies.

Gotchas

Two honest warnings. First, LTX's cost model is tokens = (W/32)(H/32) × latent frames, and pass two runs on the output grid - so the upscale factor decides how much a take costs: x2 = 4x the pixels of pass one, x1.5 = 2.25x. The pack's own measurements: 960x544 → x2 → 1920x1088 at 193 frames renders on 24 GB, while 481 frames hangs a 32 GB card fully offloaded. Keep shots short for long takes. Second, the join modes are not quality-neutral - continue is the seamless mode but spends content on the overlap; if a shot feels like it's rehashing the previous one, lower overlap. And as with everything LTX, it loses to H3 on faces and complex motion - use it for speed and iteration, and don't expect the facial fidelity H3's reference rows give you.