FunPack LTXAV Scene Chain Sampler

This is the node that turns a shot list into a movie. FunPack LTXAV Scene Chain Sampler takes multi-entry scene conditioning - one conditioning entry per scene, produced by Studio or Video Refiner V2 with split-by-transitions - and samples the whole thing as one smooth LTXV/LTXAV continuation chain. Each scene gets its own chunk, each chunk starts from the previous one's tail, and the result comes out as a single continuous latent. It's the difference between "a folder of five clips" and "one take."

How it works

The first scene samples from a fresh copy of latent_template. Every following scene copies the previous output's tail into the start of the next chunk, masks that overlap so it's preserved during denoising, samples the new frames with that scene's conditioning, then blends the overlap in latent space. Audio and video continue together on LTXAV - audio overlap is derived from the audio/video latent length ratio.

The required inputs you'll set: model + vae (LTXV/LTXAV), positive (the scene conditioning list - one entry per scene, in order), negative (shared), sampler + sigmas (from a FunPack sampler), seed, latent_template, num_frames_per_scene (default 97), frame_overlap (default 16 - pixel frames carried from the previous scene), cfg (default 1 - LTX is a CFG=1 model), max_scenes (default 8), transition_duration (default 16 - extra fade frames past the blend zone on each side of a boundary).

The inputs to actually care about

Three controls dominate a beginner's experience:

• frame_overlap - this is your continuity budget. More overlap = smoother seams, costlier. The warning in the tooltip is worth taking literally: frame_overlap=0 combined with carry_i2v_guides=True is confirmed to produce bad results. Use it only for testing.
• carry_i2v_guides (default off) - appends protected frames from latent_template's noise_mask as hidden guide tokens into each chunk. Reference as context, not as visible frames.
• use_same_seed (default off) - off means each scene uses its per-scene seed metadata (from Studio/Refiner split mode) falling back to seed + scene_index; on means every scene reuses the first seed. Off is what you want once Studio is attaching scene seeds; on is for debugging.

Then the continuity extras, all default off and all worth reading the tooltips on: mid_scene_guide (appends the previous scene's middle frame as an LTX guide - keep mid_scene_guide_strength at 0.25 minimum or audio degrades), joyai_memory (JoyAI-Echo cross-shot memory bank - a managed set of clean prior-shot frames injected into every scene, superseding mid_scene_guide), embed_guidance (applies the Refiner's learned liked-quality direction at every step - adds ~20–30% overhead), score_slider (FreeSliders-style taste guidance running 3 forward passes on quality steps), dynashift (a negative prompt at CFG=1 driven by your bad ratings instead of text), bounded_attention (spatial subject splitting for two figures in one frame), and plateau_cache (step-caching that skips ~3-4 of 8 transformer forwards on the default schedule - marked UNVALIDATED LIVE, A/B before trusting it).

Outputs: latent, images, status, scene_count (INT), scene_report (per-scene text/seed log), scene_boundaries (JSON with overlap diagnostics and contamination zones).

The big gotcha: it's heavy

The docs say it plainly: "this sampler is resource heavy." Long chains produce very large final latents, and you can OOM during VAE decode even when every sampling chunk succeeded. The fixes: decode_tile_size (set ~512 if decode OOMs), lower max_scenes, reduce num_frames_per_scene, or decode shorter chains. This is the pack's most memory-hungry node and there's no shame in tiling the decode.

Install

It's part of ComfyUI-FunPack:

cd ComfyUI/custom_nodes
git clone https://github.com/olivv-cs/ComfyUI-FunPack
pip install -r requirements.txt

or ComfyUI Manager → search "ComfyUI-FunPack". It leans on ComfyUI-LTXVideo, which you'll need anyway to load LTX models.

Where it fits

The intended flow: Studio or Refiner V2 with split_by_transitions → one conditioning entry per scene → this sampler → stitch/export. Manual workflows work too - encode each scene separately and combine with ConditioningCombine (not ConditioningConcat, which merges entries into one scene). Keep the character description before the first transition so it anchors every scene. If you want the full montage without wiring a graph, the Cutting Room editor drives this same engine from a real timeline.