Muse Collective LTX Timeline V3

The most frustrating part of long-form video generation is that your character is only consistent until they're not. You're five chunks in, the plot's moving, and suddenly the protagonist has a different nose. MuseDirectorSamplerV3 takes V2's director and adds Face ID - a reference-image identity lock that pulls the sampled face toward a real face you provide, chunk after chunk.

V3 is a WIP fork, and the pack is honest about it: V2's full feature set is intact (Seed Hunt included), and the delta is the face machinery. If identity drift is your number-one enemy, this is the version to try.

How Face ID works

The node uses an identity-overlap conditioning pass (under the hood it drives LTX's LTXIdentityOverlapConditioning, the same identity conditioning Lightricks ships for 2.3) to bias the sampled face toward a reference image. The reference goes in through the optional face_reference_image input - and the tooltip gives you the good advice: feed it a close-up face crop, not a full-body shot. The MuseFaceLock node in this same pack exists precisely to produce that crop from any reference image, via SAM3 text-prompted segmentation.

The controls are the required face_id_enabled toggle plus a small cluster of identity knobs:

• face_id_enabled - master switch. Default off; V3 behaves like V2 when it's off.
• identity_projector / source_id / arcface_mode - which face-embedding projector pipeline to use and how it reads the reference. Defaults are the sane starting point; changing these is for people who've already seen the defaults fail.
• phase_scale and id_strength - how hard and at what stage of sampling the identity is pushed. id_strength is your dial: too low and identity drifts anyway, too high and the face starts looking pasted-on and stiff.

The rest of the panel

Since V3 inherits V2, you get the full director: the timeline editor with MAIN/AUDIO/BG AUDIO/MOTION tracks, [SPEECH]/[SOUNDS] tags (uppercase only), Seed Hunt with seed_hunt and the use_seed_hunt_1..4 toggles, chunking (chunk_duration_seconds, carry_frames), and the two-stage sampling (stage1_steps 8, stage2_steps 4, stage2_denoise 0.42, cfg 1). Outputs are last_chunk_frames, audio, stage1_frames, and the four seed-hunt preview/audio pairs.

Installing it

Same pack, same drill:

cd ComfyUI/custom_nodes
git clone https://github.com/muse-collective-26/muse-ltx-timeline

Restart and pip install av torchaudio soundfile. You'll need the LTX 2.3 stack plus the talking-head LoRA if you're doing lipsync. If you want Face ID plus the automatic face crop, add MuseFaceLock - which needs the comfyui_sam3 custom node package, a dependency the README doesn't advertise. The pack loads V3 as a WIP module in a try/except, so a missing dependency won't take the whole pack down, but it will silently skip the node.

Gotchas

• Face ID is not magic. A close-up reference with good lighting gives you a chance; a mid-shot of a face at an angle gives you a coin flip. Crop tight, crop well.
• Identity conditioning costs VRAM and time on top of an already-heavy 22B pipeline.
• V3's Face ID sits alongside V4's Ghost Mask as two different answers to the same problem - V3 locks the face, V4 locks whole-character references across the timeline. If you're fighting full-character drift rather than just faces, V4 is the one.

It's WIP, it's fiddly, and the identity knobs reward tuning. But if you've ever watched a character's face quietly change over a ninety-second clip, you know why this node exists.