comfy-minimax-videogen

Use this skill for local MiniMax H3 text-to-video, image-to-video, and reference-to-video generation through the comfy-videogen CLI. The pipeline is included in comfy-diffusion 2.6.0 and generates an MP4 with synchronized audio. All MiniMax modes use the MiniMax-H3 Turbo LoRA and its four-step audio/video sampler. The LoRA is expected at loras/minimax/minimax_h3_turbo_4step_ckpt850.safetensors under the configured models directory. The custom node directory defaults to /home/victor/AI/Comfy/ComfyUI/custom_nodes/ComfyUI-MiniMax-H3-Turbo; use --turbo-node when ComfyUI is installed elsewhere.

For MiniMax H3 generations, enable both accelerators by default by passing --sage-attention --easycache. Only omit either flag when the user explicitly requests it or prioritizes quality/compatibility over speed. EasyCache remains an opt-in CLI default internally; this skill makes it the agent's default choice.