💜 LTXV Scene Extender (MVP) ErosDiffusion

LTXVSceneExtenderMVP is the stripped-down sibling of the pack's flagship LTXVSceneExtender. Same idea - timestamped scene scripts, image guides, extend an existing video - but cut down to a single temporal chunk, with the audio machinery left out. It's literally the "minimum viable product" the author shipped first to prove the script parsing and image-guide resolution worked before building the multi-chunk, synchronized-audio version.

Which makes it the friendlier place to start. One chunk, no audio VAE to wire, no overlap audio knobs to fiddle with. If you want to learn this pack's scene-script format or test whether image guides behave, the MVP is where you do it without dragging the whole machine along.

The caveat from the pack README applies to all three nodes: pre-alpha, in progress, "not meant for usage." The MVP is the most stable of the three by construction (fewer moving parts), but "most stable pre-alpha" is still pre-alpha.

How it works

Mechanically it's the same wrapper pattern as the full node: feed it the standard LTX plumbing (model, video_vae, sampler, sigmas, noise, guider), and it parses scene_script, resolves image references against the guide_images batch, and runs a single chunk through the LTXV sampler path. It even checks whether you've fed it an audio-video model via the guider - if you have, it prints a tip to route its output through LTXVSeparateAVLatent to split video and audio, since the MVP itself doesn't produce an audio output.

The big gotcha is in the tooltip: it uses the first chunk of the script only. The script format is the same as the full node:

[00:00-00:03] A woman speaks | audio:"Hello" | first:$0 | end:$1

…but only [00:00-00:03] matters. Lines after the first are ignored, not an error - which is exactly the kind of silent behavior that wastes an afternoon.

The inputs that matter

• model, video_vae, guider - standard LTX; guider wants an STG-type guider. No CLIP input here, unlike the full node.
• num_frames - pixel frames to generate, default 97 (the classic LTX "121 frame" lineage, just shorter).
• frame_overlap - overlap frames when you're extending an existing video, for continuity. Leave it alone for new generations.
• scene_script - your shot list. Remember: first chunk only.
• guide_strength (0.9) and overlap_strength (0.5) - how hard the guides bite, and how much the overlap region conditions the output when extending.
• latent (optional) - an existing video latent to extend; leave empty for a fresh generation.
• guide_images (optional) - the batch your $0, $1 refs point at.

Outputs

Three: latent (decode it through the video VAE to get frames), plus positive/negative conditioning if you're routing further. No audio output, no separate video/audio streams.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/erosDiffusion/ComfyUI-Erosdiffusion-LTX2

or search "Erosdiffusion LTX2" in ComfyUI Manager, then restart. No pip dependencies in the repo, but the node uses ComfyUI's newer v3 node API and comfy_extras.nodes_lt, so run a current ComfyUI. You need an LTX model on disk - for video-only MVP work, even the older LTXV checkpoints work, since audio isn't required here.

Troubleshooting

The most common failure is the silent-ignore trap: your 40-second script generates a 3-second clip, because everything after chunk one didn't exist as far as the node was concerned. Check the console - the node prints which chunk and prompt it's using. Otherwise the usual suspects apply: wrong model wiring, OOM if you push resolution up, and general pre-alpha roughness. If you genuinely need multi-shot output with audio, this node isn't done - that's what the full LTXVSceneExtender is supposed to become.