JoyLTX Multishot Sampler (LTX-2.5, AV-extend joins)

This is the node the whole pack is built around, and the idea is genuinely good: render each shot as its own generation - so VRAM only ever holds one shot's worth of latents - then join them into one video with an AV-extend. The previous shot's last frames and its sound are pinned into the head of the next one, so a take can be as long as you like with no visible seam, or you can cut between angles and keep the same voice. One story in, one video out, and the length of the final clip is not limited by your card.

The inputs split into two camps. The settings you'll touch on every project: prompts (the writer's {"prompts": [...]} JSON, or blocks separated by ---), width/height (960x544 default), frames_per_shot (8n+1 frames at 24 fps - 193 is 8 seconds), shot_count (0 = every prompt the writer produced), and join. The three join modes are the heart of the thing:

• continue (AV extend: seamless take) pins the previous shot's last overlap latent frames - video and audio - at the head of the next. Same voice, same room, one continuous take across generations. The replayed head is trimmed on output.
• cut (voice extends, new picture) pins only the audio tail: the voice carries straight across a picture cut, the picture is free to be a new angle. This is how you get dialogue continuity across a cut.
• fresh pins nothing: independent shots, plain and simple.

Then the identity machinery, which is the second reason people buy into this pack. identity_ref (default "cuts only: frame from shot 1") attaches a frame of shot 1 to every later shot as an in-context keyframe reference - LTX-2.5 AddGuide, appended tokens that condition but never appear - so a cut to a new angle keeps the same face, clothes and room without any reference photos. identity_strength (0.6) is how hard it pulls: 0.4–0.7 gives you the same person with free composition, 1.0 is close to a first-frame copy. "All shots" also anchors continue mode to shot 1's look, which the tooltip notes fights slow texture drift on long takes.

The optional inputs are where reference photos enter: start_image (first frame of shot 1, image-to-video), shot_images (one image per shot, used as each shot's first frame in cut/fresh mode), end_images (one keyframe per shot at that shot's last frame - first→last-frame shots when combined with shot_images), and ref_images + ref_mask from Refs by Name or RefImage, which attach a reference photo per shot and lock each character's face and voice. two_pass (on by default) upscales each shot with the latent upsampler and refines - it needs upscale_model wired - and both passes pin the previous tail, so the refine never redraws the join. The distilled schedules ship as sigmas_pass1 (8 steps) and sigmas_pass2 (3 steps), with video_cfg and audio_cfg at 1.0. Leave those alone.

Outputs are images, audio, and info (a string logging what happened). save_every_shot writes each raw shot untrimmed to output/video/JOYLTX_SHOTS/ if you want the pieces.

Gotchas, from the pack's own code and README. Fewer, longer shots mean fewer joins: the README notes 4 × 30 s single-pass shots on 24 GB is a 2-minute take with three joins, so think in long takes. The "voice" in continue mode is pinned audio, so if a new speaker appears, the sampler has a mechanism to pin the audio from their own earlier shot instead of continuing the last voice. And one real trap the node defends against: an empty prompts input usually means the workflow file is a stale browser copy where ComfyUI dropped the link - reopen the workflow from the sidebar, don't rebuild it. Set up via the JoyLTX25 pack's install; if you're using the writer, expect occasional 503/empty responses from cloud models - pick another in the writer's list and rerun.