MiniMax H3: Text to Video
Generate video with native stereo audio directly from a text prompt using MiniMax H3, an omni-modal model that synthesizes voice, sound effects, and music in a single pass. Output is up to 2K resolution, 24fps, and roughly 15 seconds long, with no file inputs required — just a prompt to one video output. Ideal for rapid concept visualization, social media content creation, and audio-synced storytelling.
Try an example
About this workflow
This ComfyUI tutorial workflow turns a single text prompt into a finished video with native stereo audio using MiniMax H3, an omni‑modal model that synthesizes visuals, voice, sound effects, and music in one pass. There are no file inputs to manage—just describe what you want, and the graph returns a roughly 15‑second clip at 24 fps, up to 2K resolution, with stereo sound baked in.
Under the hood, a MiniMax H3 subgraph (node 79dd8a95-ce9d-4c14-b264-2162e8bec5ce) handles generation and decoding. It uses the MiniMax H3 video and audio VAEs (minimax_h3_video_vae_fp16.safetensors and minimax_h3_audio_vae_fp32.safetensors) to reconstruct frames and stereo audio from the model’s latents. ResolutionSelector maps convenient presets to valid dimensions in multiples of 32 (e.g., 960×544, 1152×640, 1216×672), and SaveVideo multiplexes the frames and stereo audio stream into a single MP4. A MarkdownNote in the canvas provides quick tips and prompt examples. The result is a compact, prompt‑only pipeline ideal for rapid concept visualization, social posts, and audio‑synced storytelling.
Questions & answers
Include explicit audio guidance in your prompt. Specify narration style (voice gender/tone/accent), music genre and mood (e.g., "lo‑fi, soft piano, no drums"), and key sound effects ("gentle wind, distant traffic, minimal footsteps"). Placing audio cues near the relevant visual beats often improves alignment.
The graph targets ~15 seconds at 24 fps and supports up to ~2K resolution. Use ResolutionSelector to choose valid 16:9 sizes in multiples of 32 (e.g., 608×352, 736×416, 864×480, 960×544, 1056×608, 1152×640, 1216×672, 1280×736). Larger sizes increase generation and decode time.
This tutorial workflow is prompt‑only and outputs the model’s native stereo audio. To replace audio with your own track, extend the graph: disable or ignore the generated audio stream and mux your external WAV/MP3 before SaveVideo (or post‑replace audio in your NLE).
Yes, the subgraph expects the MiniMax H3 VAEs: minimax_h3_video_vae_fp16.safetensors and minimax_h3_audio_vae_fp32.safetensors. If ComfyUI reports missing files on first run, download them from the provided links and place them in your models directory so the node can decode video and stereo audio correctly.