H3 Segment

---

When a MiniMax H3 timeline renders, the strip of shot cards you wrote on the Creator node has to become actual sampling work. MiniMax H3 Timeline Segment is one shot of that timeline as a node - written into the graph by the Timeline/Creator node, one per card, and not meant to be placed by hand. It's the load-bearing bit that turns "a card with a prompt and some references" into a model, conditioning, and a latent the sampler can chew on.

Its most interesting design detail is about caching, and it's worth understanding even if you never touch the node. It takes a self-contained payload (segment_data) rather than the whole timeline plus an index. Why? Because its cache key changes when this segment changes and not when any other one does. Re-shoot card 3 and cards 4–9 aren't re-sampled from scratch - ComfyUI's input-hash caching sees their segment_data unchanged and skips them. That's the mechanism behind the padlock/"shoot one pass at a time" workflow in the pack: locked cards aren't rendered, and untouched cards stay cache hits.

The inputs that matter

• clip - the text encoder, passed through. Required.
• segment_data - the self-contained JSON for this shot: prompt, assets, LoRAs, duration, checkpoint. This is what keys the cache.
• vae / audio_vae - optional, and deliberately so. A text-only segment encodes no picture, so the video VAE is only reached for when there's a keyframe or a visual reference to turn into a condition latent; the audio VAE is only touched when the request carries reference audio or a sound seam. Leaving them unwired when unneeded keeps the loaders a decode-time cost. If one is needed and missing, the node raises rather than reaching a None inside the encoder.
• vae_name / audio_vae_name - the VAE filenames, so cached references can be keyed to them.
• model_fl2va / model_ref2va - the two checkpoints the segment may land on; the mode routing decides which one actually runs.
• prev_image / prev_audio - an earlier segment's last frame and soundtrack tail, when this segment continues from it. next_image / next_audio - the opening of the supplied clip this segment runs into. These are where the seam nodes' outputs land.
• hold_lora - a LoRA to leave off the lead model output - the distillation, for a turbo lead-in.

The outputs that matter

• model - the patched model for this segment.
• positive - the conditioning, built per-segment.
• LATENT - the starting latent (noise, or the inherited keyframe).
• lead model - the model with hold_lora left off, for turbo lead-ins.

Installing

It ships with the pack and appears only inside expanded graphs:

cd ComfyUI/custom_nodes
git clone https://github.com/roadmaus/ComfyUI-MiniMax-Creator

Restart ComfyUI, no pip install, standard H3 weights. If you ever expand a Creator workflow and see a row of these, they're the individual shots - and the reason re-shooting one card doesn't re-render the whole strip is sitting right there in how segment_data keys the cache.