V2 更新为 LightX2V MiniMax H3 Turbo LoRA,新增两组一键切换:① 4步/8步 LoRA 与对应采样步数同步切换;② 循环视频/首尾帧视频切换。循环模式会把第一张图同时作为首帧与尾帧,首尾帧模式则分别使用两张图。

采样与解码改用 comfyui-minimax-h3-audio-T8 插件的 MiniMaxH3DualClockSamplerT8 与 MiniMaxH3AVDecodeT8 节点,改善原生音画联合生成的音频表现。

RTX 5090 单次实测:5秒、约560p档的视频约47秒完成。实际速度会受分辨率与环境影响。请将 CUDA 环境升级到13.0,否则无法达到上述速度。

LightX2V LoRA:
https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main

V2 adds two one-click switches: (1) synchronized 4-step/8-step LoRA and sampling-step selection; and (2) Loop Video / First-and-Last-Frame Video. Loop mode reuses the first image as both endpoints, while first/last-frame mode uses two separate images.

The workflow now uses the comfyui-minimax-h3-audio-T8 dual-clock sampler and AV decoder for improved native audio output. Tested once on an RTX 5090: a 5-second video at approximately 560p-class completed in about 47 seconds. Performance varies by resolution and environment. Use CUDA 13.0 to reproduce this speed.

LightX2V LoRA:
https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main