凤希AI - MiniMax音频分段加载

FxAiMiniMaxAudioSegmentLoad (凤希AI - MiniMax音频分段加载) is the node that makes "one long audio file, many MiniMax H3 segments" actually work. MiniMax H3 generates video in short clips - the model's legal frame counts sit on a 17k+5 grid (5, 22, 39, 56, … up to roughly 5–15 seconds) - so a 3-minute render has to be chopped into a loop of smaller segments. The hard part isn't the chopping, it's keeping the audio cut and the video frame count exactly in sync so each segment's picture lines up with its sound. That's this node's entire job: given your full audio and a list of segment durations, cut the audio for scene N, and hand back the exact number of H3 frames that segment should generate.

Mechanism: it takes the 分段时长列表 (a LIST of per-segment durations, in seconds), converts each duration to a frame count, and rounds each down to the H3 grid - then the last segment absorbs whatever remainder is left so the sum of all segments covers the whole audio. For the current segment (当前索引), it computes how many frames came before it, slices the 原始音频 waveform from that point, and adds the 过渡帧数 overlap onto the end. Outputs: 剪切音频 (AUDIO) for this segment, and 生成帧数 - the frame count (segment frames plus transition frames) you feed into the sampler so the video and audio stay locked.

The inputs that matter:

• 原始音频 - your full track (AUDIO).
• 分段时长列表 - per-scene durations. Produce this from your scene timings (the pack's audio segmenter/manager nodes build it).
• 当前索引 - which scene you're rendering this loop pass.
• 过渡帧数 - the overlap frames carried into the next segment; multiples of 17 up to 17.

Where this fits: it's the companion to FxAiMiniMaxFrameCalculate (which computes a single segment's frame count from an audio length) - the segment loader is the same idea generalized to a whole scene list, so a loop over 当前索引 drives a full long-video render scene by scene. That "segment and loop" architecture is the pack's core pitch: keep each pass tiny enough that a low-VRAM card never holds a whole clip at once.

Install: ComfyUI Manager → "fxai-toolkit", or cd ComfyUI/custom_nodes && git clone https://github.com/fxai666/fxai-toolkit, restart. Auto-installs soundfile/psutil, FFmpeg on PATH, and note the MiniMax nodes additionally need the H3 weights and a ComfyUI core with MiniMaxH3 support - the pack patches the core H3 implementation at import (fxai_minimax_core_patch) to fix an official bug where references overwrite first/last-frame keyframes, so keep the pack updated with your ComfyUI. All labels are Chinese (凤希AI, QQ group 775649071, Bilibili).

Failure modes worth knowing: an index past the segment list doesn't crash - it returns the raw audio and a frame count of 5, which will quietly render a 5-frame segment of your whole audio if you're not watching. And since the last segment absorbs all rounding remainder, one segment will always be slightly longer than its listed duration; don't fight it, it's how the math closes.