MiniMax H3 Timeline Segment

The Creator Palette's storyboard isn't a second renderer - it's the same workflow, sliced into shots. MiniMax H3 Timeline Segment is one of those shots as a node: the Creator's job for a single segment, written into the graph by the Timeline node. You won't add it from the menu; it's the per-shot workhorse that lets a whole piece be one Timeline instead of a row of manually-wired Creators.

It takes a self-contained payload (segment_data) rather than the timeline plus an index, which is a quiet design win: its cache key changes when this segment changes and not when any other one does. Re-edit shot 7 and shots 1–6 replay from cache instead of re-sampling.

How it works

Each segment compiles its own prompt, assets, LoRAs and duration into a payload, then does what the Creator does for one shot - model, positive conditioning, and a latent for the sampler - and exposes the model twice: model (with the shot's LoRAs patched) and lead model (the same model with hold_lora left off, for the opening steps of a turbo lead-in). Without a hold_lora set, lead model is the first output - the same object, not a second patch - so a graph that never wires it pays nothing for it.

The seam inputs are where the timeline magic lives: prev_image / prev_audio carry the earlier segment's last frame and soundtrack tail, so this shot continues them; next_image / next_audio carry the opening of the clip this segment runs into. Both VAE inputs are optional because they're only reached when the segment actually encodes a keyframe or reference - a text-only segment leaves them unwired, and the node raises a clear error if it needs one and it's missing.

The inputs that matter

• clip and segment_data - required; the CLIP for text encoding and the JSON payload that defines the shot.
• vae / audio_vae - optional but needed the moment the segment encodes a keyframe or a visual/audio reference.
• prev_image / prev_audio - "an earlier segment's last frame, when this segment continues from it" (tooltip), and the same for sound.
• next_image / next_audio - the opening of a supplied clip this segment runs into.
• model_fl2va / model_ref2va - optional; H3's FL2VA and Ref2VA are separate weights, and only the routed one gets loaded.
• hold_lora - optional; the LoRA left off the lead model output, for a turbo lead-in.

Outputs: model, positive, LATENT, lead model.

Why it's built this way

A segment starts from the decoded last frame of the one before it, a dependency that only exists downstream of sampling - you can't express it by returning conditioning and re-wiring. So the Timeline builds a graph: each segment node, then the reel chain that decodes and spills each pass, then the save. The segment node is the per-shot brain in the middle of that chain. If you ever hand-build one and a segment complains about a missing VAE, read the message - it names the exact input the encode path reached for. Install is the shared pack install: ComfyUI Manager → search "H3 Creator Palette", or git clone https://github.com/z3rofeels/ComfyUI-H3-Creator-Palette into custom_nodes/, restart, hard-refresh. ComfyUI 0.34.0+.