CivArchive
    MiniMax-H3 Seamless, lossless long-video extension | Noise-free, optimized audio | Handles complex, high-speed, high-dynamic-range content | Workflow - v1.0
    NSFW

    MiniMax-H3 Infinite Continuation Workflow — lossless long-video extension for complex, high-speed, high-dynamic action.

    🚀 Run It on RunningHub

    https://www.runninghub.ai/zh-cn/post/2115906018629775362/?inviteCode=rh-v1429/?inviteCode=rh-v1679

    一句话上手:在提示词框里用 & 9 & 标注每段秒数、用 === 分隔段落,点一次运行,工作流自动循环生成全部段落,并拼成一条无接缝、无爆音、无色差、不发油的影视级长镜头。

    One-liner: Write one prompt with & 9 & per segment and === between them, hit Run once — the workflow loops through every segment automatically and stitches them into a single seamless, noise-free, color-matched cinematic long take.


    ✨ 四个核心机制 / Four Core Mechanisms

    1️⃣ 二采无加噪续航 —— 解决「二次加噪爆音」

    一采从 output 直接输出 0.690 处的带噪潜变量,二采挂 DisableNoise 沿用一采剩余噪点继续退噪至 0。没有第二次加噪,就不会破坏音频相位,也就不会有跨段爆音。

    Pass 1 hands over the still-noised latent at 0.690; pass 2 continues from that exact noise state via DisableNoise. No re-noising → no broken audio phase → no crackle at the seam.

    2️⃣ 绕开有损压缩的「20 帧无损音频 + 2 帧隔离带」

    不走 CreateVideo 的 H.264/AAC 有损链路,直接从 32kHz fp32 无损音轨截取 20 帧送入下一段音画引导;引导在第 16~20 帧做 7 ticks 梯度衰减,画面引导持续到第 22 帧,中间留 2 帧(0.08 秒)物理隔离缓冲带——续段裁头时连同缓冲带一起切掉,爆音 100% 留在正片之外。

    20 frames of lossless 32kHz fp32 audio (bypassing H.264/AAC), faded out over the last 7 ticks, plus a 2-frame (0.08 s) isolation buffer that gets trimmed away with the segment head. Character timbre is preserved; the crackle never reaches the final cut.

    3️⃣ 22 帧锥形彩噪 —— 抗「长视频发油、形变、糊」

    对上段末尾 22 帧注入 0.55 强度的像素级块状彩噪,且末尾 3 帧递减到 0.0,在多轮重编码之间打断累积的涂抹感与面部油光。这是条件输入端的抗劣化,不会出现在成片画面里。

    A tapering chroma-noise injection (0.55 → 0.0 over the last 3 frames) into the tail conditioning breaks up the "greasy / smoothed-out" accumulation across re-encode rounds. It lives in the conditioning path only — never in your final frames.

    4️⃣ 前 12 帧三阶渐隐消色差

    提取上段最后 1 帧作参考,仅对续段前 12 帧做 0.75 ➔ 0.45 ➔ 0.18 的渐隐色彩校准,第 13 帧起 100% 原画直通,既抹平色差又不伤画质。

    The previous segment's last frame becomes the color reference; only the first 12 frames get 0.75 / 0.45 / 0.18 graded falloff. From frame 13 onward the image passes through untouched.

    ➕ 两个额外保险 / Two Extra Safeguards

    • 零裁尾防丢台词:每段有效帧的尾部 100% 保留,句尾台词、动作收势不再被吞;只在续段裁掉前 22 帧重叠区(自动对齐 17n+5 帧网格)。 Zero tail-trim: the tail of every segment is kept intact — no swallowed last words. Only the 22-frame head overlap of continuations is trimmed, auto-aligned to the 17n+5 frame grid.

    • 文戏 / 武戏一键切换:开关打开(True)时惰性旁路 6 个高速动态 / 倍速 LoRA,不加载、不占显存,人物神态、口型、慢镜头更稳;关闭(False)即满血武戏。 One switch lazily bypasses the 6 high-speed / motion LoRAs (never loaded, no VRAM cost) for dialogue-driven scenes; turn it off for full combat mode.


    🧩 安装依赖 / Required Custom Nodes

    在 ComfyUI-Manager 搜索安装,或按仓库说明放入 ComfyUI/custom_nodes 后重启后端并刷新网页。 Install via ComfyUI-Manager, or clone into ComfyUI/custom_nodes, then restart ComfyUI and refresh the browser.

    关键节点 / Nodes

    1 ComfyUI-KJNodes 信号总线、TAEH3 实时预览、H3 显存优化

    2 rgthree-comfy 单个输入的启用 / 忽略开关

    3 ComfyUI-VideoHelperSuite

    4 ComfyUI-Easy-Use

    5 孤海工具箱 GoohaiTools

    6 MiniMax H3 Audio T8

    另需 MiniMax-H3 官方节点与 taeh3.safetensors(实时预览用)。 MiniMax-H3 core nodes and taeh3.safetensors (for live preview) are also required.


    🚀 快速上手 / Quick Start

    1. Load the workflow. Drag the .json onto the canvas. Missing nodes show as red — install the six packs above first.

    1. Write the prompt (top-left 📝 node). Each segment opens with & 9 & to declare its duration in seconds (omit it to use the default 8 s). Separate segments with === alone on its own line.

    1. Load references (G4). 9 image slots, 2 video slots, 3 audio slots. Use the Fast Bypasser above each slot to enable or ignore it — unused slots can stay empty.

    1. Set the director controls (G2): resolution, default segment duration, motion slider -10 ~ +10, seed, and start / end segment numbers (0 = auto-detect every segment).

    1. Pick a mode. 🎭 → True for dialogue and slow narrative; keep False for full combat. Turn on 📉 Reduce VRAM 2 if you're tight on memory.

    1. Queue and walk away. The loop runs every segment automatically.

    Pro tip: run segment 1 → 1 first to check framing, identity and action read before committing to the full batch. Each segment is roughly one complete two-pass inference, so total time scales linearly with segment count.


    📝 提示词写法 / Prompt Format

    & 9 &
    subject_definitions:
    <Subject 1> (Picture 1) is ...   # lock the face from the close-up only;
                                     # full-body views supply proportions + wardrobe
    <Subject 2> is ...               # set / environment
    <Subject 3> is ...               # style: camera language, action physics, grade
    
    summary:
    [reference generation] A continuous 9-second single-take ...
    
    retention_analysis:
    <Subject 1>: fully_preserved - ...
    <Subject 2>: fully_preserved - ...
    
    ===
    & 9 &
    (segment 2 — continue the action and camera from the previous beat,
     repeat the full subject and style definitions)
    ===
    & 9 &
    (segment 3 ...)
    1. & n & at the segment head

      Generation length in seconds. Optional — omit and the default duration applies.

    2. === on its own line That is the separator. Not ---, not a blank line, not ***.


    FAQ

    Why is the final cut shorter than segments × duration?

    Each continuation generates 22 extra head frames for overlap, then trims them. So 6 × 9 s ≈ 49 s, not 54 s. That's the cost of a seamless join — no frames are lost from your actual picture.

    The first segment flashes or looks odd at the start?

    On the first iteration there's no previous segment to guide from, so the workflow substitutes 22 blank frames to avoid a crash. That branch's result is discarded downstream. Normal placeholder behaviour — it does not affect the final film.

    Red "missing node" boxes?

    Install all six packs from section 03 and restart the backend, not just the browser tab.

    Out of VRAM / OOM?

    ① Drop resolution to 0.6–0.9 MP

    ② enable 📉 Reduce VRAM 2

    ③ enable MiniMax H3 ChunkFeedForward (slightly slower)

    ④ keep Low VRAM Attention on permanently.

    Action always over-driven when I want a quiet dialogue scene?

    Set 🎭 Dialogue mode to True. The six motion / speed LoRAs are bypassed by lazy evaluation — not loaded, no VRAM cost — and lip sync plus micro-expressions settle immediately.

    Still a slight color pop at a seam?

    Reuse the same <Subject 3> style definition and the same lighting description across segments. The color match only covers 12 frames — source consistency does the heavier lifting.

    How long can it run?

    Segment count is unbounded in principle (loop-counter driven); the real limits are disk space and your patience. Batches of 6–10 segments work well — watch the running concatenation output as you go.

    Does it work for non-action content?

    Yes. Dialogue mode plus a slower motion setting handles conversation, and the continuation logic is content-agnostic. The bundled demo just happens to be a fight scene.


    0Credits

    Workflow by Acheng 琢影 (Mr.CheerFul) Bilibili: https://space.bilibili.com/152078897

    任何使用问题请加 QQ 群咨询:1098238945 Skill 请看群文件

    [🐰兔子的 AI 资源中心](https://x.gd/fourbunny)

    If you have any questions about using this, please join the discord https://discord.gg/XUEZNkFkJu

    For the skill, please check discord

    BUNNY Resource hub: https://x.gd/fourbunny Discord: https://discord.gg/XUEZNkFkJu

    Description

    FAQ

    Workflows
    MiniMax H3

    Details

    Downloads
    260
    Platform
    CivitAI
    Platform Status
    Available
    Created
    11/10/2026
    Updated
    11/18/2026
    Deleted
    -

    Files

    MiniMax-H3 Infinite Long-Take — Seamless Long-Video Extension for High-Speed Action.json