凤希AI - MiniMax图生视频

FxAiMiniMaxImageToVideo (凤希AI - MiniMax图生视频) is the front end of the fxai-toolkit's local MiniMax H3 pipeline. It's not a sampler and it's not a saver - it's the encode-and-condition stage: you give it a prompt, a CLIP, a video VAE, dimensions, frame count, and as much reference material as you have (first frame, last frame, reference images, reference audio, even a reference video), and it returns the two things your sampler needs: a 正向条件 (positive conditioning) and an AV潜变量 - the joint audio-video latent that H3's omni-modal architecture consumes.

If you've never touched H3: it's MiniMax's 33B open-weight omni-modal model that treats text, image, video, and audio as one context and outputs clips with native audio - no bolted-on lipsync pass. That unified design is exactly why this node has audio inputs at all. H3 generates on a 17k+5 frame grid at 24fps; the 帧数 field (default 124 ≈ 5s) auto-aligns to it, and the training range is roughly 124–362 frames.

The inputs that matter (the ones the source comments explain):

• 提示词, CLIP模型, 视频VAE - your text encoder, and the video VAE that builds the latent.
• 宽度 / 高度 / 帧数 - canvas and clip length. Step-32 dimensions, step-17 frames.
• 首帧图片 / 尾帧图片 - the keyframe anchors. This is where the pack's core patch matters: the pack fixes an official H3 bug where reference images overwrote keyframe conditioning, so both coexist here.
• 参考图片列表 / 参考视频列表 / 参考音频列表 - reference material injected by number. You reference them from the prompt as <Picture N>, <Audio N>, etc., per H3's official convention.
• 外置音频 - drives the audio track; 音频VAE needed when external audio is used.
• 过渡帧列表 + 过渡羽化 - carry the previous segment's tail frames into this latent and soft-lock them. 过渡羽化 (-1 auto, 0 hard lock, 1–5 feather steps) controls the smooth width from locked to free region so segment boundaries don't flash. This is the long-video continuity magic.
• 音频模式 - 音色参考 (external audio only informs timbre, model generates content), 原音频 (external audio is locked into the output channel - for singing/digital-person lip-sync), or 系统生成 (model does its own audio).
• 音频去噪 - FFmpeg afftdn noise suppression on the reference, off by default.

Outputs: 正向条件 and AV潜变量 (a nested tensor of video [1,24,T,h,w] plus audio [1,32,2,T]). Both go to a sampler (paired with FxAiMiniMaxBlockCache for speed) and then to FxAiMiniMaxVideoSave.

Hard requirements the source enforces: 原音频 mode requires both 音频VAE and 外置音频 or it errors; 音色参考 with external audio also requires the 音频VAE. So connect the audio VAE before you get clever with modes.

Install: ComfyUI Manager → "fxai-toolkit", or cd ComfyUI/custom_nodes && git clone https://github.com/fxai666/fxai-toolkit, restart. The pack auto-installs soundfile/psutil and needs FFmpeg; the MiniMax family additionally needs the H3 weights and a ComfyUI core with MiniMaxH3 - the pack patches core at import, so keep both updated together. Labels are Chinese (凤希AI; QQ 775649071, Bilibili), and there's no English Reddit community around the pack - H3 itself is very well received, but this specific wrapper's help is Chinese-language.

Where people get burned: forgetting the audio VAE in an audio mode (hard error, but the message is Chinese), and expecting the node to sample - it doesn't, it only conditions and encodes; if nothing comes out, check what's downstream. Start with 首帧图片 + 系统生成 audio, get one clip working, then add refs and the 原音频 lock.