FL MiniMax H3 Beat Shot Planner

If FL MiniMax H3 Prompt Timeline is the manual-edit path, this is the automated one. FL MiniMax H3 Beat Shot Planner takes an exact FL_PROMPT_SCHEDULE - the output of FL Audio Beat Prompt Schedule in ComfyUI Fill Nodes - and turns it into a shot_plan: a list of independent H3 render units, each with its own latent, conditioning, and matching slice of audio. That's the machinery behind the "make a whole music video, one beat-driven render at a time" workflows.

The key idea is render grouping. Prompt sections that are marked as one render group in the audio sequencer are encoded and rendered together, so H3 can model internal cuts with shared context - one continuous render for a verse with a couple of internal cuts. Ungrouped sections become independent renders, each sampled separately. You get explicit control over which cuts share generative context without being forced to render the whole track as one long clip. The planner even slices the timeline audio per shot, so each render only sees the audio that belongs to it.

The render units

Everything downstream works off the shot_plan output - Beat KSampler samples it, Shot Motion Context augments it, Shot Assembler decodes and stitches it back in order. For convenience the node also exposes scheduled, latent, and semantic for the first render, so you can build a plain single-sampler workflow off a connected schedule without pulling in the whole beat-sampling stack.

Two ways to use it:

• Connected schedule - feed prompt_schedule (and optionally timeline_audio and prompt_envelopes) from the Fill Nodes sequencer. Each shot gets its own audio slice, reactive envelopes get sliced and rebased per planned render.
• No schedule - it plans one continuous render from the length input, which makes it a compact drop-in alternative to Prompt Timeline when you don't need per-section scheduling.

Inputs that matter

The encoder trio - clip, vae, audio_vae - plus global_prompt and width/height work like Prompt Timeline. Then there are the reference-strength controls that are genuinely H3-specific:

• visual_condition_fidelity - noise fidelity for visual latent references and motion context. Lower values replace more latent detail with seeded noise. It does not scale Qwen-VL, so don't use it as a generic prompt-weight knob.
• visual_reference_mode - which visual reference paths (images, videos) actually receive connected references. Audio and soundtracks stay active in every mode.
• reference_strength - blends from a prediction with no visual references (0) to the selected mode (1). The Beat KSampler's cfg multiplies this value, and the tooltip's warning matters: effective blends other than 0 or 1 require both predictions while sampling, so an in-between blend costs extra.

Gotchas and install

Grouped vs. independent renders is the thing to think about before you queue: grouped sections render together (better internal-cut coherence), but the sampler returns one editable nested latent per planned render regardless. And remember H3's 17k+5 frame grid - the planner snaps lengths, and the length input only matters when no schedule is connected.

Install is the pack-standard story: ComfyUI Manager → FL MiniMax H3, or cd ComfyUI/custom_nodes && git clone https://github.com/filliptm/ComfyUI-FL-MiniMaxH3.git, then restart. No extra Python deps, and the pack doesn't ship models - you need H3's diffusion model, Qwen3-VL text encoder, and both VAEs. The beat-schedule inputs come from Fill Nodes, which is recommended for this path but not required if you're just running a single continuous render. And as with every node here: don't run this pack against an old Fill Nodes build that still registers the MiniMax node IDs, or you'll get duplicate registrations.