沐阳 H3 · 长视频（原生多关键帧）

沐阳 H3 · 长视频（原生多关键帧） (H3LongVideo) is where the pack earns its "long-video workflow" headline. It's one node that, at runtime, expands into N sampling chains - one per segment - stitching them together with the anchor mechanism you'd otherwise wire by hand: carry the tail of each segment into the head of the next, sample, trim the overlap, move on. If 沐阳 H3 加载器 is the entry point and the Director is the all-in-one, this is the manually-wired graph's engine, and it's also what the Director delegates to internally.

The critical input is plan_json, and the tooltip hammers the point: it's a segment plan (how many segments, how many frames each, how much they overlap), not a prompt. You get it from a splitter like H3ScriptSplitter, and the number that matters most is the overlap, because context_length here must match the plan's overlap_frames exactly. Get those out of sync and your seams stop being seamless - that mismatch is probably the single most common LongVideo bug.

The settings that actually matter

• task_mode - 动作迁移 (motion transfer, default): each segment follows the corresponding slice of the reference video. 视频续写 (continuation): only the reference video's ending is used as the starting point, and picture and sound pick up from there. 纯生成 (pure generation): no reference video, prompt and images only.
• context_length - 22 is the stable default; 5 is the experimental speed anchor; 39/56 for stronger continuity at higher token cost. Must match plan overlap.
• ref_image_size - the reference budget. "匹配生成分辨率" is the memory-cheap default; the max-fidelity 2048 short-edge option makes every reference token more expensive every step, and the tooltip warns it's "several times slower."
• save_segments / segment_prefix / save_raw_segments - per-segment saving. Long runs die at hour three sometimes; saved segments mean you can resume and only re-run what broke. save_raw_segments additionally keeps the pre-second-pass segments (suffixed _原始) when you want to A/B against the refined output.
• prompt_mode / media_prefix - when the prompt box is empty, each segment uses its plan brief, prefixed with the media-reference line so @视频1 style mentions keep working.

Optional inputs do the important work: ref_video (motion transfer needs it long enough to slice one piece per segment; continuation only reads its ending), ref_audio (continuation's sound picks up from the reference's ending), media from the Media Agent (the whole-job media bundle), 二采设置 (the second-pass settings bundle), and context_video/context_audio for resume - a previous segment's finished clip whose tail only seeds the first segment's anchor, never the ref2va reference channel.

Outputs: images and audio, the assembled long take.

Install and ground rules

Install via ComfyUI Manager search "ComfyUI-MiniMaxH3-Myang", or git clone https://github.com/civilcoco/ComfyUI-MiniMaxH3-Myang into custom_nodes, restart, hard-refresh. No extra Python deps; supply the H3 model, text encoder and both VAEs yourself.

The README's procedure is the right one: render two segments with a 22-frame context first, inspect picture, motion, lip-sync and audio at the join, and only then crank up the segment count or add Turbo/second-pass. Long chains accumulate losses in detail, timbre, brightness and saturation - a clean seam doesn't guarantee a clean chain. And the pack's compatibility caveat: it's regression-tested against ComfyUI v0.33.2/v0.34.0 layouts, so after any ComfyUI update, validate a two-segment render before committing to a long job.