🎬 MiniMax-H3 — Video + Audio Generation (FP8 · Turbo)

Text→Video and Reference→Video with synchronized native audio, 4–15 s, up to 1080p-class output, 24 fps. Runs on 1× NVIDIA RTX PRO 6000 Blackwell (96 GB) via Hugging Face ZeroGPU (xlarge), FP8 weights + official 8-step PDD acceleration LoRA pre-filled. 中文: 文生视频 / 参考生成,带原生音频。默认已填好阿里 PAI 官方 8 步加速 LoRA(Steps=8,速度快);想要最高质量:清空 LoRA 输入框,Steps 调到 20–28。

MiniMax H3 — Text to Video (+ Audio)

💡 Tip: Type a prompt and generate video with synchronized native audio. ⚡ The 4-step turbo acceleration LoRA (pruned-compatible build) is pre-filled in LoRA Settings (expand it) — default Steps = 6 (4 is fastest, 8 is safer; strength 0.8–1.8 also works). For max quality: clear the LoRA field and use 20–28 steps. With the default 120s ZeroGPU window, 544p / 5s tasks pass comfortably; 768p is recommended for quality.

Resolution
Aspect Ratio
Steps(步数)
0.2 15

💡 Tip: When downloading from Civitai, please use the Version ID, not the Model ID. You can find the Version ID in the URL (e.g., civitai.com/models/123?modelVersionId=456) or under the model's download button. When downloading from Hugging Face, please use the format: repo_id/filename.extension or repo_id/folder_path/filename.extension (e.g., Comfy-Org/MiniMax-H3/loras/minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors or Comfy-Org/MiniMax-H3/loras/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors).

LoRA Source 1
0 2
Demo by e5ey · Framework: RioShiina (AGPL) · Weights: MiniMax-H3 / Comfy-Org FP8 / rzgar FP8 encoder / alibaba-pai Acc LoRAs