Muse Director V10

The name tells you what it is: this is the node that directs. Muse Director V10 is a single custom node that takes LTX 2.3 AV (the 22B video+audio model from Lightricks) and turns a drag-and-drop timeline into a finished, long-form video clip - per-segment prompts, speech, ambient audio, lipsync, and a character that stays the same person from the first frame to the last. If you've ever stitched five LTX generations together in ComfyUI and watched the face drift or the scene cut hard at every seam, this is aimed at you.

It's not a beginner's first node - it expects a real VRAM budget and a model you've already got running. But if your goal is "90 seconds of coherent talking-head video with matching audio, without hand-wiring a hundred nodes," nothing else in the open-source stack does it in one graph.

How it works

LTX 2.3 can't natively render minutes of video, so Director does what every long-form workflow does: it splits the timeline into overlapping chunks, samples each one, and locks the tail of the previous chunk into the start of the next via carry frames (carry_frames, default 73 ≈ 3s at 24fps) so the seam disappears. Chunks can align to your timeline segments (auto_chunk_by_segment) so a "cut" lands where you put it instead of straddling a boundary.

Under the hood it runs a two-stage pipeline: a cheap low-res Stage 1 draft (8 steps, euler), then a Stage 2 pass that upscales and refines it (stage2_denoise 0.42). The distilled model runs at CFG 1.0, which silently kills the classic negative-prompt branch - so the node routes negative prompts through NAG (Normalized Attention Guidance) via KJNodes' LTX2_NAG instead. That's why negative_prompt and nag_scale exist.

The identity piece is the clever part. Drop up to three character images into the timeline's character-card panel and pick a reference_mode. Ghost Mask (End) appends the photos as hidden guide frames past the end of each chunk, then crops them off before decode - the model sees real reference photos as identity anchors without them ever appearing in your output. Licon MSR (Prefix) is the newer IC-LoRA identity guide (via ComfyUI-LTXVideo), injected at the front; it needs vae connected. Either way, tag a slot in a prompt with @ref1/@ref2/@ref3 to say which character appears in that segment.

The inputs and outputs that matter

The required inputs are the full LTX stack: model (with LoRAs applied), clip, audio_vae, vae, and spatial_upscaler. Most people wire loaders → this node → video combine and never touch the rest. Optional ref_images adds non-character reference material (an object, a set), and bg_audio mixes ambience under everything at bg_volume.

Outputs: last_chunk_frames (your finished clip - wire it into a VHS Video Combine), audio (the mixed track), stage1_frames (a debug view of the draft pass), four seed_hunt_preview_N/seed_hunt_audio_N pairs, and reference_image (handy for color-matching the result to the source photo).

The most useful feature for LTX specifically is Seed Hunt - LTX 2.3 is far more sensitive to seed than to prompt wording, so flip seed_hunt on, watch four cheap Stage 1 candidates, then flip the matching use_seed_hunt_N on. The node carries that candidate's actual cached latent into Stage 2, so what you previewed is what you get. The cache lives in memory - restart the server and it falls back to re-rolling with the widget's seed.

Installing it

ComfyUI Manager (search "Muse Director V10"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/muse-collective-26/muse-director-v10-final
pip install av torchaudio soundfile

Then restart. You'll also need three companion packs - VideoHelperSuite (saving/preview), ComfyUI-LTXVideo (Licon MSR mode), and KJNodes (NAG) - plus the model files: the LTX 2.3 distilled transformer (diffusion_models/), video and audio VAEs (vae/), the Gemma 3 text encoder (text_encoders/), the spatial upscaler (latent_upscale_models/), and the talking-head LoRA for lipsync. Download links are in the README; budget several GB before anything runs.

Gotchas

• Licon MSR crops with the stock LTXVCropGuides, not MuseCropGuides - the node warns about this, and it's a real trap: MSR's guides sit at the front, Ghost Mask's at the back, so the crop needs to know which end.
• Ghost Mask artifacts near the chunk tail. If you see faint hallucinated junk in the last few visible frames, raise ghost_anchor_buffer (default 2) - the developer hit exactly this while debugging.
• Echoed voice in generated-audio mode. enable_ambient_pass runs a second audio pass that watches the video for context; it can duplicate speech faintly. Flip it off to test.
• [SPEECH] and [SOUNDS] tags must be uppercase - lowercase is silently ignored.
• single_stage_mode is experimental - it skips Stage 2 and samples once at full res, and the author hasn't A/B'd it. Keep the default unless you know why you're switching.
• Use individual angle shots per character slot, not one composite reference sheet - a multi-panel sheet gets resized as one image and hurts identity fidelity.