沐阳 H3 · 分段计划

Here's the wall every MiniMax H3 user hits eventually: the model clips out at 4–15 seconds, and your video is a minute. You need to split it into segments, and H3 is picky about the grid those segments land on. H3ScriptSplitter is the first step of that job - it takes one long script and slices it into a per-shot plan on H3's official 17k+5 frame grid, at the model's fixed 24fps. It's the planning head of the 沐阳 H3 pack, the "director's cut" of what was once a lot of fiddly per-segment prompt wrangling.

The mechanism is simple and worth understanding, because it drives the whole chain. You feed it a script (or the myang_prompt output of the Media Agent), and it asks an LLM to cut that script along the time axis into however many segments fit your duration. Each segment gets its own prompt, and all of it is embedded into a single plan_json string - one LLM call for the whole job, not one per shot. If the LLM times out, returns malformed JSON, or produces the wrong segment count, it silently falls back to a local timeline algorithm, so the plan still happens. Leave script empty and it skips the LLM entirely and just computes time and frame math - 0 tokens, pure arithmetic.

The inputs a beginner actually touches:

• script - your full prompt or screenplay. Wire the Agent node's myang_prompt here, or paste raw text.
• total_seconds / length_source - the total runtime. Switch length_source to "match reference video" and it reads the duration straight off a connected ref video, ignoring the number you typed.
• segment_seconds - how long each clip should be. H3's grid snaps these, so don't expect exact values back.
• overlap_frames - the big one. This MUST match the context_length on the long-video node, because every later segment carries the previous one's tail as temporal context. 22 is the recommended baseline; 5 is the "I'm just testing speed" anchor.
• max_segments, llm_enabled, use_cache - caps the shot count, controls whether the LLM runs at all, and reuses the last split when script and segment count haven't changed (no repeated token spend).

On the way out you get plan_json (feed this to H3SegmentPrompt), segment_count, segment_seconds, frames_per_segment, plan_preview (a readable dump for checking), and ref_frames_needed - how many context frames the chain will consume.

Where people get burned: the overlap_frames vs context_length mismatch is the classic silent killer - your seams drift and the audio drifts off the picture. And the fps input is locked at 24 for a reason: H3 models at 24fps, and the node will raise an error if you try anything else. Don't fight it.

Install - it ships with the whole pack, so the shared steps apply: ComfyUI Manager, search ComfyUI-MiniMaxH3-Myang (or the "沐阳 H3" display name), install, restart. Manual route is the usual

cd ComfyUI/custom_nodes
git clone https://github.com/civilcoco/ComfyUI-MiniMaxH3-Myang.git

then restart ComfyUI and hard-refresh the browser tab if the node list looks stale. It declares no extra Python dependencies beyond the standard ComfyUI environment, but it does assume you're on a recent ComfyUI build with the official MiniMax H3 nodes - those are the actual models behind this, and you'll need to supply the H3 diffusion model, Qwen text encoder, and both VAEs yourself. Also remember the MiniMax H3 Community License geofences the weights out of the US, EU, UK and Korea - worth reading before you download anything.

Start with two segments at 22 context frames before scaling up. If the seam holds at two, it'll hold at twelve.