MiniMax H3 Shotboard

Most video nodes are a single job: load a model, condition, sample. The IAMCCS_MiniMaxH3ShotPlanner - displayed as "MiniMax H3 Shotboard" - is the opposite. It's the editorial brain of the IAMCCS MiniMax H3 pipeline, a single node that decides everything about how a shot gets made: which task mode, what resolution, which sampler, whether a turbo LoRA or LTX detailer gets applied, how the audio behaves, how reference images are used, and what the upscale/delivery path is. Every downstream node in the H3 workflow reads its decisions through one contract output.

If you're coming from a plain KSampler habit, this looks overwhelming - it has dozens of widgets, and most of them you should touch rarely. The design philosophy is that a shot plan is a persisted object, not a set of loose constants: you save a board once and it reproduces its VRAM strategy, its LoRA choices and its audio policy every time.

What to actually set

• global_prompt - your master shot description. It's the anchor that the local timeline rows refine.
• timeline_data - the shot's timeline (segments, timings), usually filled by the Shotboard UI rather than typed by hand.
• task_mode - the big one. auto_from_timeline is the default, but the useful choices are t2va (text → video + audio), i2va (image → video + audio), ref2va (reference → video + audio), v2va_object_swap, and ref2vid_lipsync. Getting this wrong sends your references to the wrong conditioning path.
• duration_seconds - shot length; the planner splits it into the chunked segments the renderer can actually produce.
• width / height - 960×544 is the proven low-VRAM default (roughly 0.5MP, 16:9). Raise it only if you have the VRAM to burn.
• performance_profile - low_vram_balanced is the safe start; the rtx3060_* and h3_turbo_* presets trade quality/VRAM for speed.
• seed, steps, sampler_name, scheduler - the sampling core, matching MiniMax H3's native expectation (res_multistep + simple are moved to the top of the lists for a reason).
• upscale_mode and the upscale_* block - whether the shot gets an LTX 2.3 or Wan 2.2 5B finishing pass; the Shotboard decides, the actual model branches stay lazy in the workflow.

There's also a full reference layer: reference_role_1..4 (subject identity / keyframe / composition / style) and reference_video_role / reference_audio_role, plus turbo_mode with turbo_lora_name, ltx_detailer_enabled, and the whole LTX temporal-looper block for long masters. You don't need all of it on day one.

The contract

The single output is cine_linx - a structured payload that bundles the shotplan, the timeline, and a report string. Everything downstream (IAMCCS_MiniMaxH3Backend, the generation nodes, the upscale router, even the Prompter) reads this instead of duplicating your settings. That's the whole trick of the pack: one authority, many readers.

Install and gotchas

Install via ComfyUI Manager (search "IAMCCS") or cd ComfyUI/custom_nodes && git clone https://github.com/IAMCCS/IAMCCS-nodes.git, then restart. The H3 graphs themselves need a current ComfyUI with native MiniMax H3 support, ComfyUI-GGUF if you load H3 or the Qwen text encoder from GGUF, and KJNodes for the Sage/low-VRAM patches. Keep a single copy of the pack - the docs are explicit that duplicate/backup copies in custom_nodes register stale classes and cause import failures. And note this planner only plans: it doesn't load or sample anything. Wire its cine_linx into the Shotboard backend nodes that do the actual work.