IAMCCS H3 Settings PRO → CineLinX

The short version

Configuration panel with a cable. Nothing goes in - no image, no latent, no model - and one thing comes out: cine_linx. It's the single source of truth for MiniMax H3 generation settings in IAMCCS-nodes' Shotboard workflows, so the Shotboard stays a planner and your settings live in one place. If it looks like a cockpit, that's on purpose.

Why it exists at all

MiniMax H3 is the 33B omni-modal model that makes 4–15 second clips at 24fps with native synced stereo audio (sampled in the same pass, not bolted on) and day-one ComfyUI support. What it doesn't do is long video: its native sample window tops out at 362 frames, the 15-second-class ceiling the pack hard-codes. Anything longer is the Shotboard's job: plan a timeline, cut it into legal chunks, carry the AV tail between them.

All of those knobs inside the planner node would be miserable, so they live here instead - hence no inputs, one output. One caveat before you commit a weekend: the H3 Community License excludes the US, EU, UK and South Korea, outputs included.

How it works

On execute, every widget not on an internal omit list gets packed into a named dict under the schema iamccs.minimax_h3.settings_cine_linx, wrapped in a CineLinX stage payload, and returned as cine_linx (type IAMCCS_SUPERNODE_LINX). Two fields are dropped on purpose: duration_seconds and audio_mode. Those are authored Shotboard facts, and a stale settings payload must never flip native H3 audio back into a previously published soundtrack.

The planner reads the dict off the cable, ignores those two, and stamps its plan's settings source as IAMCCS_ShotboardH3Settings - versus r38_defaults when nothing's connected. That log string is proof the cable is live. Missing fields are deliberately not backfilled from the Shotboard's saved snapshot, so a field you never set stays unset.

PRO versus the plain Settings node

Same backend - INPUT_TYPES is a deep copy of the standard node and export just delegates. The separate class exists so the frontend can mount a different panel: a tabbed rail, a Mode Assistant that writes your chosen mode onto the Shotboard, one-click VRAM and accelerator recipes, a Queue Truth readout, and health lines that flag incoherent combos before you queue. Pick PRO for the UI; nothing here is exclusive to it.

The inputs that matter

• task_mode - 18 modes, auto_from_timeline by default: i2va, fl2va, ref2va, the longvid family, control video, object and face swap. PRO mirrors this plus duration_seconds, frame_rate and audio_mode from the Shotboard, so set the mode there and let this follow.
• performance_profile - default low_vram_balanced, ladder from rtx_xx60_safe to rtx_xx90_max. Pair it with motion_context_window_frames (max and default 362 - the trained window including the carried AV tail; smaller is an explicit memory trade-off) and h3_exact_chunk_rows (2048 is the conservative 8–12GB value). h3_exact_profile is only a label preset, 124 frames on a xx60 up to 362 on a xx90, and it never overrides the boxes below it.
• steps / sampler_name / scheduler / denoise - 16 steps, euler, simple by default. Alongside sit shift_video (12) and shift_audio (3): H3 shifts video and audio sigmas separately because the two are sampled jointly. Not decoration.
• acceleration - 21 entries: exact attention, low-VRAM fallbacks, PDD's 8-step recipe, FastH3 at 6, SLA at 4, a fused INT8 trunk. Each needs its LoRA or model actually installed.

Install

ComfyUI Manager, search IAMCCS (publisher iamccs), or:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

Restart ComfyUI. The pack wants ComfyUI ≥ 0.3.0, Python ≥ 3.12, Torch ≥ 2.8 (CUDA 12.6 or 12.8). No requirements.txt, so no separate pip step - it rides on ComfyUI's own deps and wires up an ffmpeg shim via imageio-ffmpeg at import. Nothing downloads itself either: the H3 trunk and text encoder, a 3D latent upscaler in latent_upscale_models, the Fun Union patch in model_patches, accelerator LoRAs in loras, SAM3 and BiRefNet for face swap are yours to source.

Where people get burned

The health line already tells you the usual one: connect Settings PRO to Cine H3 Input, then Cine H3 Input to the Shotboard. The cable doesn't reach the planner directly.

The rest fail before sampling rather than during it: Fused Fast without T2VA, with a secondary style LoRA, or alongside Turbo (they can't both run); Turbo or PDD enabled with its adapter missing ("H3 turbo: missing adapter" - the PDD picker rejects old aptech files and wants native *_Acc-8Step*_comfy LoRAs); a FL2V LoRA left on a Ref2VA mode; ControlNet enabled without the Fun Union checkpoint or its branch connected; Face Swap without BiRefNet in background_removal. And "VRAM preset is partially overridden" just means you edited one box out of the contract - reapply the recipe.

One annoyance: the README points at docs/MINIMAX_H3_WORKFLOW_REQUIREMENTS.md, which isn't in the 1.5.4 checkout. The tooltips are this node's real docs.