Pulse Slate · MiniMax H3

The whole point is that you never type <Picture 3> again

MiniMax H3 is the 33B omni-modal video model that ComfyUI got day-zero support for, and it's genuinely good - native stereo audio, joint generation, the works. But its reference tags are a trap. <Picture N>, <Audio j> and friends are ordinals assigned by socket position in ComfyUI's source. Insert one image at the top of your reference bin and every <Picture N> after it silently shifts by one. The render still succeeds - it just describes the wrong pictures, with no error anywhere. A reference video's soundtrack even renumbers every standalone <Audio> in your prompt before you know what happened.

Pulse Slate is the compiler node from the Pulse Studio pack that kills that whole failure class. You write @Mimi or {{mimi}} instead of a number; the node computes every ordinal from live bin order at compile time, and hand-typed tags are reported as an error rather than silently trusted. It's the difference between a tool that fights you and one that checks your work.

What it actually does: compile, don't render

Pulse Slate takes your prompts, an asset bin and (optionally) a chain of PulseShot nodes, and turns them into a PULSE_TIMELINE. It does not sample. There are two paths depending on length:

• ≤ 15 s (one window) - it hands back positive and latent, and your own graph does sampler → decode → mux. This is PulseSlate_Single.json, the short path.
• > 15 s (many windows) - positive and latent are blocked; you take the timeline output to a PulseRender node, which does the window loop, disk cache and stitching.

That blocking is a fix, not a bug: pre-3.0.0 builds handed the last window's latent back, and a still-wired sampler quietly re-rendered that one window as the whole film - a 7-second file where there should have been fifteen, no error anywhere.

The inputs that matter

The node face carries two multiline boxes plus an asset panel - nothing hidden behind a gear.

• global_prompt - art style, lighting, camera rules, identity locks, score. Optional line labels: style: identity: retention: soundscape: music:; unlabelled text counts as style.
• shot_prompt - one shot per line, each starting with [Shot N] or a [MM:SS.mmm] timecode. Shots without a timecode spread evenly between the ones that have them. Quoted "text" becomes dialogue.
• The asset bin - drag images, video or audio straight onto the node body, no loader nodes. Each drop gets a short alias you rename inline. A description promotes a reference from a bare citation to a <Subject N> definition (that's what identity consistency is built on); retention is written verbatim into the prompt's retention section.

Beyond those: aspect_ratio (presets fitted to H3's 1,032,192 px budget - 16:9 is 1344×768), window_seconds (each H3 call, capped at ~15.08s), partition_strategy (balanced / fill / shot_aligned), and continuity - last_frame_carry and keyframe_pairs pin real frames, which is the fl2va checkpoint's job, so they need model_fl2va wired and fail at compile time without it.

cfg stays at 1.0 - seriously

H3's reference pipeline has no negative conditioning; it runs BasicGuider, a single-conditioning path. cfg = 1.0 is the native default. Above 1.0 it switches to CFGGuider with an empty negative - offered because people asked, not because it's good. Leave it.

Install and the model haul

cd ComfyUI/custom_nodes
git clone https://github.com/Addis-Pulse-Studio/comfyui-pulse-studio

Restart ComfyUI. There's nothing to pip install - zero dependencies, just torch, numpy, Pillow and av that a working ComfyUI already has. You can also find it in ComfyUI Manager by searching "comfyui-pulse-studio". Two real requirements:

1. ComfyUI 0.30.0+, carrying comfy_extras/nodes_minimax_h3.py and comfy/ldm/minimax/. Older builds simply don't have H3 to direct.
2. The weights, not redistributed here. From Comfy-Org/MiniMax-H3: the ref2va DiT (~20 GB) and the fl2va DiT (~20 GB, only if you use anchored continuity) into models/diffusion_models/minimax/, the qwen3vl text encoder (~15 GB) into models/text_encoders/minimax/, and the video + audio VAEs into models/vae/minimax/.

Two gotchas before you commit 40 GB of disk. On Linux and macOS the five loader widgets load red because the example graphs were authored on Windows with backslash paths - click each and re-pick the file. And the weights carry the MiniMax H3 Community License, which excludes the US, EU, UK and South Korea from its applicable territory - if you're in one of those, running the local weights isn't licensed, even though the hosted Hailuo API is.

Where people get burned

• Pre-2.0.0 workflows won't load. Widget order changed once, deliberately; those files are refused by name rather than guessed at. Rebuild from example_workflows/PulseSlate_LongForm.json. There is no migration and won't be one.
• The speed patches are wired upstream, not by this pack. The documented chain is UNETLoader → SpectrumApplyMiniMaxH3 (history_storage: system_ram) → PatchSageAttentionKJ → Sol-Attn → PulseSlate. The node inspects what it's handed and warns - on the node face and console - when it finds no attention or offload patch, but it never blocks and never patches for you. And system_ram isn't a speed setting: on a 32 GB card it's what makes a full-length window fit at all; set Spectrum to vram and expect an OOM.
• Run dry_run first. A wrong reference binding renders successfully and gives you a well-formed film of the wrong person; the per-shot ordinal map in the report is how you catch it before spending GPU hours.