<div data-youtube-video="https://youtu.be/mw3JTMK0ZmA"><iframe src="https://www.youtube.com/embed/mw3JTMK0ZmA" width="853" height="480" allowfullscreen></iframe></div><br /><br />
<p>RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2085223972794355714?inviteCode=rh-v1111</p>
<h2>MiniMax H3 Digital Human - Two-Character Dialogue</h2>
<p>Purpose: a two-character (and by extension multi-character) digital-human graph. Each character gets an image and an audio clip; the two clips are trimmed, padded and concatenated into one timeline, that combined audio is locked, and MiniMax H3 generates the matching two-person performance.</p>
<h2>Graph (as saved)</h2>
<p>32 nodes / 38 links / 5 groups.</p>
<ul>
<li>Core: <code>MiniMaxH3ReferenceToVideo</code> at 1344x768, length 124, <code>ref_image_size=match</code>; <code>VRGDG_MiniMaxH3AudioDrive</code> set to lock the concatenated original audio and generate picture only.</li>
<li>Models: UNETLoader <code>minimax_h3_ref2va_int8_convrot.safetensors</code>; CLIPLoader <code>qwen3vl_32b_minimax_h3_int8_convrot.safetensors</code> (type minimax); VAELoader video <code>minimax_h3_video_vae_fp16.safetensors</code> + audio <code>minimax_h3_audio_vae_fp32.safetensors</code>.</li>
<li>Sampling: <code>SamplerCustomAdvanced</code> + <code>KSamplerSelect</code> <code>res_multistep</code> + <code>BasicScheduler</code> <code>simple</code>, 20 steps, denoise 1.</li>
<li>Inputs: three LoadImage nodes (character 1, character 2, plus a third reference slot), two LoadAudio clips with <code>AudioCrop</code> at 0:00-0:05 and 0:00-0:06, two <code>AudioConcat</code> nodes, <code>EmptyAudio</code> and <code>PreviewAudio</code>. Prompt addresses subjects as <code>&lt;Picture 1&gt;</code> / <code>&lt;Picture 2&gt;</code> and the audio as <code>&lt;Audio 1&gt;</code>.</li>
<li>Output: <code>VHS_VideoCombine</code> (h264 mp4, crf 19, yuv420p). Note the saved filename prefix is still <code>AnimateDiff</code> — a historical leftover in the export, not a different model path.</li>
</ul>
<h2>How to use</h2>
<p>Feed character 1 and character 2 with their own images and audio, then write the dialogue prompt in the order the graph expects. The audio crops define who speaks when: if the two clips do not line up with the planned exchange, fix the crop times first, not the prompt. The third image slot is present in the save state; check whether you need it before queueing.</p>
<h2>Limits and not-verified items</h2>
<p>No inference was run for this listing: no speed, VRAM, lip-sync, voice or multi-character consistency claim is made. The saved prompts, images and audio are the author's demo content. VoxCPM and the H3 checkpoints are not bundled; check their licenses separately. The cover image is a promotional still, not a frame produced by this graph.</p>
<p>Tutorial: <a href="https://www.youtube.com/watch?v=mw3JTMK0ZmA" rel="ugc">YouTube</a> · <a href="https://www.bilibili.com/video/BV1kUuW6FEVs" rel="ugc">Bilibili</a>. Workflows supplied through AI-KSK; verify workflow, node, model, audio, likeness and cover rights separately. No blanket commercial or redistribution license is asserted.</p>

MiniMax H3 Digital Human - Two-Character Dialogue

<div data-youtube-video="https://youtu.be/mw3JTMK0ZmA"></div><br /><br />
<p>RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2085223972794355714?inviteCode=rh-v1111</p>
MiniMax H3 Digital Human - Two-Character Dialogue
<p>Purpose: a two-character (and by extension multi-character) digital-human graph. Each character gets an image and an audio clip; the two clips are trimmed, padded and concatenated into one timeline, that combined audio is locked, and MiniMax H3 generates the matching two-person performance.</p>
Graph (as saved)
<p>32 nodes / 38 links / 5 groups.</p>
<ul>
<li>Core: <code>MiniMaxH3ReferenceToVideo</code> at 1344x768, length 124, <code>ref_image_size=match</code>; <code>VRGDG_MiniMaxH3AudioDrive</code> set to lock the concatenated original audio and generate picture only.</li>
<li>Models: UNETLoader <code>minimax_h3_ref2va_int8_convrot.safetensors</code>; CLIPLoader <code>qwen3vl_32b_minimax_h3_int8_convrot.safetensors</code> (type minimax); VAELoader video <code>minimax_h3_video_vae_fp16.safetensors</code> + audio <code>minimax_h3_audio_vae_fp32.safetensors</code>.</li>
<li>Sampling: <code>SamplerCustomAdvanced</code> + <code>KSamplerSelect</code> <code>res_multistep</code> + <code>BasicScheduler</code> <code>simple</code>, 20 steps, denoise 1.</li>
<li>Inputs: three LoadImage nodes (character 1, character 2, plus a third reference slot), two LoadAudio clips with <code>AudioCrop</code> at 0:00-0:05 and 0:00-0:06, two <code>AudioConcat</code> nodes, <code>EmptyAudio</code> and <code>PreviewAudio</code>. Prompt addresses subjects as <code>&lt;Picture 1&gt;</code> / <code>&lt;Picture 2&gt;</code> and the audio as <code>&lt;Audio 1&gt;</code>.</li>
<li>Output: <code>VHS_VideoCombine</code> (h264 mp4, crf 19, yuv420p). Note the saved filename prefix is still <code>AnimateDiff</code> — a historical leftover in the export, not a different model path.</li>
</ul>
How to use
<p>Feed character 1 and character 2 with their own images and audio, then write the dialogue prompt in the order the graph expects. The audio crops define who speaks when: if the two clips do not line up with the planned exchange, fix the crop times first, not the prompt. The third image slot is present in the save state; check whether you need it before queueing.</p>
Limits and not-verified items
<p>No inference was run for this listing: no speed, VRAM, lip-sync, voice or multi-character consistency claim is made. The saved prompts, images and audio are the author's demo content. VoxCPM and the H3 checkpoints are not bundled; check their licenses separately. The cover image is a promotional still, not a frame produced by this graph.</p>
<p>Tutorial: <a href="https://www.youtube.com/watch?v=mw3JTMK0ZmA" rel="ugc">YouTube</a> · <a href="https://www.bilibili.com/video/BV1kUuW6FEVs" rel="ugc">Bilibili</a>. Workflows supplied through AI-KSK; verify workflow, node, model, audio, likeness and cover rights separately. No blanket commercial or redistribution license is asserted.</p>

Files with SHA256: cb7381d908ef50276031ccd4ddd2cdd954b04df7de39aa50b54a1b5301e56dde