凤希AI - MiniMax视频保存

FxAiMiniMaxVideoSave (凤希AI - MiniMax视频保存) is the output end of the fxai-toolkit MiniMax H3 pipeline - and unlike a plain "save video" node, it's built for the loop. You feed it the decoded 图片序列 and the segment's 音频, and it writes an mp4, then hands back the tail of the render: the last frames and the trailing audio, pre-cut for you to feed into the next segment's H3 encode. That "pass the tail forward" behavior is the entire reason a 3-minute H3 render can be built from a chain of 5-second segments without visible seams.

Mechanically it splits your image sequence three ways, and the tooltip explains it exactly: the video takes the first 生成帧数 frames (grid-aligned, 17k+5); the 过渡帧列表 output is fixed at the last 22 frames of the sequence; and 过渡帧 is always just the single very last frame. The transition frames are your continuity handoff - feed 过渡帧列表 into the next segment's 过渡帧列表 input on the encoder (v1) and the next clip starts from where this one ended instead of from noise. 过渡帧音频 is the matching tail of the audio, cut to the transition length, so the sound carries over too.

The inputs:

• 图片序列 - decoded H3 frames (IMAGE).
• 生成帧数 - how many frames actually enter the video (first N; defaults 227).
• 帧率FPS - default 24 (H3's native rate).
• 目录 - output folder under your videos directory (default sucai, which is "material/asset" in Chinese).
• 视频序号 - filename suffix; -1 auto-numbers.
• 音频 - the segment's audio, muxed into the mp4.
• 音频去噪 - optional FFmpeg afftdn suppression on the output audio (off by default).

Outputs: 过渡帧, 过渡帧列表, 视频文件路径, 保存目录, 实际帧数, 过渡帧音频. The path outputs are handy for downstream merge/stitch nodes.

The pack also frees memory aggressively after saving - it deletes the image tensors, runs garbage collection, and empties the CUDA cache before returning. That's not cosmetic; it's the low-VRAM strategy that lets a 12GB card chain segments without the VRAM climbing every pass. It also registers the saved file with the pack's task store so its manager UI can find results.

Install: ComfyUI Manager → "fxai-toolkit", or cd ComfyUI/custom_nodes && git clone https://github.com/fxai666/fxai-toolkit, restart. Auto-installs soundfile/psutil; needs FFmpeg on PATH (it's doing the muxing). MiniMax family needs the H3 weights and ComfyUI core with MiniMaxH3 support. Labels Chinese (凤希AI; QQ 775649071, Bilibili).

The gotcha to remember: the transition outputs assume a 24fps, ~22-frame tail. If you run a nonstandard FPS, the 过渡帧音频 cut uses your FPS value, but the frame-count assumptions (22) are fixed - so keep FPS at 24 unless you've got a reason not to, and let the next segment consume the tail as-is rather than re-trimming it.