Muse Seed Scout

LTX 2.3 is seed-sensitive - more sensitive to the seed than to your prompt wording, as the community discovered the hard way. So the fastest path to a good clip is usually generating a handful of cheap candidates and picking the winner. MuseSeedScout is that workflow as a standalone node: it runs four short, low-resolution, Stage-1-only candidates using Muse Director V1's own Stage 1 machinery, lets you eyeball the previews, and hands you the winning seed to feed into a full Director run.

Here's the important caveat up front, straight from the pack README: Seed Scout is superseded by V2's built-in Seed Hunt for anything with more than one timeline segment. The Scout is the earlier, standalone version of the same idea, and the README is blunt that it's the lighter-weight option for single-image setups. If you're on the Director V2+, use Seed Hunt. If you're running a plain single-segment graph, the Scout is simpler.

How it works

The Scout imports V1's actual Stage 1 helper functions (prompt relay, guide application, audio/lipsync mask) rather than duplicating them, so its previews use the same recipe as the real run - not a separate, inconsistent approximation. You wire it the same inputs you'd give a Director: model, clip, vae, audio_vae, prompts, and your start image.

The flow:

1. Run with all four use_seed_1..4 toggles OFF. It generates and caches four candidates at your chosen seed_1..4 values, then purges VRAM.
2. Flip exactly one use_seed_N toggle ON. Re-running doesn't regenerate the candidates - they're cached against your other inputs - it just re-selects that candidate's seed instantly.
3. Wire the seed output into Director V1's seed input. (Right-click Director's seed widget and "Convert seed to input" first.)

Outputs: seed (INT, your picked seed) plus preview_1..4 images.

The settings are the Director's Stage 1 controls in miniature: stage1_steps, cfg (1), frame_rate, custom_width/custom_height, duration_frames (the tooltip says it plainly: "Keep this short - it's a scouting preview, not the final clip"), plus the guide/IC-LoRA settings and the audio toggles. Optional inputs are start_image and custom_audio.

Installing it

Same pack:

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

Restart, pip install av torchaudio soundfile, and the LTX 2.3 model stack. The Scout has no extra dependencies beyond what the Director needs.

Gotchas

• It only takes a single reference image. No timeline tracks. That's precisely why V2's Seed Hunt replaced it for multi-segment work - the Hunt scouts using the real timeline data (images, audio, motion guide) and carries the picked candidate's actual latent into Stage 2, so the preview and the final share real content, not just a seed integer.
• The seed output is just a number; it doesn't carry latent state. A full Director run from that seed is a fresh roll - which is fine for a Scout (that's the point: cheap search), but don't expect frame-for-frame match with the preview.
• Cache is in-memory and keyed to your other inputs; change the prompt and the old candidates are stale.

For a single-image workflow where you're not yet on the Director path, the Scout is a clean, cheap way to find the seed before committing. For everything else, the pack itself will point you to V2 - and it's a rare thing when a project's README is that honest about its own roadmap.