凤希AI - MiniMax图生视频V2

FxAiMiniMaxImageToVideoV2 (凤希AI - MiniMax图生视频V2) is the same idea as the v1 H3 encoder with the sharp edges sanded off: prompt, CLIP, video VAE, canvas, frame count, first/last frame keyframes, reference images, reference audio, and an audio mode - but without the v1 extras (reference video input, transition-frame feathering, and the FFmpeg audio-denoise step). If v1 is the full control room, V2 is the streamlined panel for the workflow you use every day: a clean image-to-video encode that still returns the 正向条件 and AV潜变量 pair your H3 sampler consumes.

The core behavior is identical to v1, and it's worth repeating because it's the reason this family exists: H3 is MiniMax's 33B omni-modal video model - text, image, video, and audio share one context, so a "video" node legitimately takes audio inputs - and it generates on a fixed 17k+5 frame grid at 24fps. The 帧数 field (default 124 ≈ 5s) auto-aligns to that grid. The pack's core patch also applies here: it fixes the official H3 bug where reference images clobbered the first/last-frame keyframe conditioning, so your keyframes and refs coexist.

The inputs that matter:

• 提示词, CLIP模型, 视频VAE - the conditioning stack.
• 宽度 / 高度 / 帧数 - 32-step dimensions, 17-step frames, grid-aligned for you.
• 首帧图片 / 尾帧图片 - keyframe anchors (frame 0 and the last frame).
• 参考图片列表 (up to 9, per the source's slicing) and 参考音频列表 - reference material, addressed from the prompt as <Picture N> / <Audio N>.
• 外置音频 + 音频VAE - the audio track.
• 音频模式 - 音色参考 (external audio informs timbre only), 原音频 (external audio locked into the output - singing/digital-person), or 系统生成 (model improvises audio).

Outputs: 正向条件 (CONDITIONING) and AV潜变量 (LATENT) - the nested video+audio latent for your sampler. The same hard rules as v1: 原音频 mode demands both 音频VAE and 外置音频; 音色参考 with external audio demands the 音频VAE. The source raises Chinese-language errors when you forget, so "it broke" usually means "you skipped the audio VAE."

Which one should you use? If you don't need reference-video injection, per-segment transition-frame feathering, or audio denoising, V2 is the less fiddly pick and the one I'd reach for on a routine clip. If you're doing long multi-segment renders with carried-over transition frames, v1's 过渡帧列表/过渡羽化 and its denoise step are the tools that keep segment boundaries clean - go v1 there. Both output the same contract, so swapping is one node replacement.

Install: ComfyUI Manager → search "fxai-toolkit", or cd ComfyUI/custom_nodes && git clone https://github.com/fxai666/fxai-toolkit, restart. Auto-installs soundfile/psutil, FFmpeg on PATH, and the H3 weights plus a ComfyUI core with MiniMaxH3 support (pack patches core at import - keep both updated). All labels Chinese (凤希AI; QQ 775649071, Bilibili; no English Reddit presence for the pack itself).