<div data-youtube-video="https://youtu.be/WdAaPq2Xa0I"><iframe src="https://www.youtube.com/embed/WdAaPq2Xa0I" width="853" height="480" allowfullscreen></iframe></div> 
 
MiniMax H3 audio-driven digital human workflow (VDN 12-step HD edition). Turn one reference portrait plus a user-supplied audio track (singing or talking) into an approximately 1080P, 24fps lip-synced digital human video. 
 
Main features: 
- Manual audio input (MP3) with start_time / duration; AudioSeparation isolates the vocals, LTXV audio VAE encodes them into H3 audio conditioning 
- Reference-to-video from a single portrait (3:4 portrait mode, 544x736 start) 
- OpenVDN DMD8 two-pass pipeline: 8NFE base pass, learned latent x2 upscale to 1920x1088, 4NFE refine (12-step total) 
- Two-pass audio audit, AV decode and output trim, safe AV save at approx 1080P / 24fps 
 
Suggested workflow: 
1. Load your portrait into the first-frame slot and select the audio file with start_time / duration 
2. Run the graph as-is; the low pass, latent upscale, high refine and audio audit chain runs in one pass 
3. Export the safe AV save (approx 1080P, 24fps) 
 
Dependencies: 
- minimax_h3_fl2va_pruned_int8_convrot.safetensors (H3 base UNET) 
- qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (CLIP) 
- minimax_h3_video_vae_fp16.safetensors and minimax_h3_audio_vae_fp32.safetensors 
- minimax_h3_latent_upscaler_3d_fp16.safetensors (learned latent x2) 
- OpenVDN/vdn-minimax-h3 (DMD8 stage), LTXV audio VAE encoder, KSK T8 custom nodes 
 
RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2108580258148929538?inviteCode=rh-v1111 
Bilibili tutorial video: https://www.bilibili.com/video/BV1pTph6EEUv/ 
RunningHub registration with invite code rh-v1111 (1000 RH coins): https://www.runninghub.ai/?inviteCode=rh-v1111

MiniMax H3 Audio-Driven Digital Human VDN

<div data-youtube-video="https://youtu.be/WdAaPq2Xa0I"></div> 
 
MiniMax H3 audio-driven digital human workflow (VDN 12-step HD edition). Turn one reference portrait plus a user-supplied audio track (singing or talking) into an approximately 1080P, 24fps lip-synced digital human video. 
 
Main features: 
- Manual audio input (MP3) with start_time / duration; AudioSeparation isolates the vocals, LTXV audio VAE encodes them into H3 audio conditioning 
- Reference-to-video from a single portrait (3:4 portrait mode, 544x736 start) 
- OpenVDN DMD8 two-pass pipeline: 8NFE base pass, learned latent x2 upscale to 1920x1088, 4NFE refine (12-step total) 
- Two-pass audio audit, AV decode and output trim, safe AV save at approx 1080P / 24fps 
 
Suggested workflow: 
1. Load your portrait into the first-frame slot and select the audio file with start_time / duration 
2. Run the graph as-is; the low pass, latent upscale, high refine and audio audit chain runs in one pass 
3. Export the safe AV save (approx 1080P, 24fps) 
 
Dependencies: 
- minimax_h3_fl2va_pruned_int8_convrot.safetensors (H3 base UNET) 
- qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (CLIP) 
- minimax_h3_video_vae_fp16.safetensors and minimax_h3_audio_vae_fp32.safetensors 
- minimax_h3_latent_upscaler_3d_fp16.safetensors (learned latent x2) 
- OpenVDN/vdn-minimax-h3 (DMD8 stage), LTXV audio VAE encoder, KSK T8 custom nodes 
 
RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2108580258148929538?inviteCode=rh-v1111 
Bilibili tutorial video: https://www.bilibili.com/video/BV1pTph6EEUv/ 
RunningHub registration with invite code rh-v1111 (1000 RH coins): https://www.runninghub.ai/?inviteCode=rh-v1111

01_minimax-h3-audio-driven-digital-human-vdn.json

v1.0

<div data-youtube-video="https://youtu.be/WdAaPq2Xa0I"><iframe src="https://www.youtube.com/embed/WdAaPq2Xa0I" width="853" height="480" allowfullscreen></iframe></div> 
 
RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2108580258148929538?inviteCode=rh-v1111Bilibili tutorial video: https://www.bilibili.com/video/BV1SEJm6PE22/
# MiniMax H3 — Audio Latents and Two Output Routes## Purpose and audienceCombine a reference image with manually encoded vocal latents while separately routing selected source audio to output. Intended for users who can resolve matching custom nodes, inspect media permissions, and verify actual outputs. This is a workflow resource, not a tested model release.## Required asset references`qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors`, `minimax_h3_video_vae_fp16.safetensors`, `minimax_h3_audio_vae_fp32.safetensors`, `minimax_h3_latent_upscaler_3d_fp16.safetensors`, `minimax_h3_fl2va_pruned_int8_convrot.safetensors`.The VDN composers reference OpenVDN/vdn-minimax-h3 and stage_dmd_8nfe; refinement plans store4. These describe two-pass intent, not measured evaluation counts, a twelve-step ordinary sampler, or a verified distilled-LoRA filename. Reconcile/audio-audit settings are requested policies, not passed tests. Resolve exact node versions and assets before queueing; no installation, GPU, speed, lip-sync, voice-preservation, or license verification is included.## Saved configuration44 nodes/80 links. Uploaded audio feeds a window whose duration comes from the loader-reported length. Separated vocals are encoded and given a zero noise mask. The LOW template is split before concatenating its video latent with the injected audio.## How to prepareReplace the image/audio selections and inspect the actual window. Selector47 supplies544×736, overriding stale960/544 labels. The ×2 upscale suggests portrait HIGH controls around1088×1472, not a fixed landscape canvas. Inspect both LOW and HIGH decode/trim/save chains; both saves are active and share a prefix.## Known limitationsDecoded generated-audio outputs are unconnected; selected original audio supplies both trims. This routing does not prove bit-exact passthrough, lip sync, voice preservation, or successful temporal fitting. Check actual dimensions, duration, and output filenames.Inspect actual output dimensions, duration, sound, input alignment, and unintended edits. Do not turn a node title, saved preview, or audit setting into a claim of a successful render.## Files and version note- 02: [Workflow JSON]() · [Selected cover]()Historical VHS video-preview metadata is omitted from distribution copies. Other data is retained, including point-editor images where present: inspect their contents and rights before any publication. Source originals are unchanged. Selected covers are not verified inference results. Input media, model weights, and VDN artifacts are not bundled.This is new local descriptive copy, not a remote publication or an announcement of a model upgrade. Source discrepancies are documented, not silently corrected.## Tutorial, attribution, and rights[Source-provided tutorial](https://www.youtube.com/watch?v=WdAaPq2Xa0I), not live-verified here. The workflows are supplied through AI-KSK. Verify workflow, node, model, source-video, image, voice, and artwork rights separately; no universal commercial permission is asserted.<h2>Source-listed resources</h2>Additional entry points listed by the episode source. Some contain referral/affiliate codes; no bonus, price, quota, or speed is asserted here.<ul><li>RunningHub registration (international, invite code rh-v1111): <a href="https://www.runninghub.ai/?inviteCode=rh-v1111" rel="ugc">https://www.runninghub.ai/?inviteCode=rh-v1111</a></li><li>RunningHub registration (China-site mirror, same invite code): <a href="https://www.runninghub.cn/?inviteCode=rh-v1111" rel="ugc">https://www.runninghub.cn/?inviteCode=rh-v1111</a></li><li>AI-KSK workflow catalog on the international site: <a href="https://www.runninghub.ai/zh-cn/user-center/1846601088051884033/userPost?inviteCode=rh-v1111" rel="ugc">https://www.runninghub.ai/zh-cn/user-center/1846601088051884033/userPost?inviteCode=rh-v1111</a></li><li>Quick-create entry (international site): <a href="https://www.runninghub.ai/ai-generator/2046514150500524034?inviteCode=rh-v1111" rel="ugc">https://www.runninghub.ai/ai-generator/2046514150500524034?inviteCode=rh-v1111</a></li><li>VIP rights entry (international site): <a href="https://www.runninghub.ai/vip-rights/3?inviteCode=rh-v1111" rel="ugc">https://www.runninghub.ai/vip-rights/3?inviteCode=rh-v1111</a></li><li>Workflow model files (Quark netdisk, China): <a href="https://pan.quark.cn/s/9ec095b95838" rel="ugc">https://pan.quark.cn/s/9ec095b95838</a></li><li>Cloud GPU signup (Compshare; referral code in the link): <a href="https://passport.compshare.cn/register?referral_code=A7fTQUYY9lCDFJwA6MMvoq" rel="ugc">https://passport.compshare.cn/register?referral_code=A7fTQUYY9lCDFJwA6MMvoq</a></li><li>Cloud image index for Compshare (Notion): <a href="https://www.notion.so/2b72459a3ac6808f8282f2c687ed8f52" rel="ugc">https://www.notion.so/2b72459a3ac6808f8282f2c687ed8f52</a></li><li>Support the channel (Ko-fi): <a href="https://ko-fi.com/aiksk" rel="ugc">https://ko-fi.com/aiksk</a></li></ul>

MiniMax H3 General Video Generation with Audio-Driven Avatar (VDN 12-Step)

<div data-youtube-video="https://youtu.be/WdAaPq2Xa0I"></div> 
 
RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2108580258148929538?inviteCode=rh-v1111Bilibili tutorial video: https://www.bilibili.com/video/BV1SEJm6PE22/
# MiniMax H3 — Audio Latents and Two Output Routes## Purpose and audienceCombine a reference image with manually encoded vocal latents while separately routing selected source audio to output. Intended for users who can resolve matching custom nodes, inspect media permissions, and verify actual outputs. This is a workflow resource, not a tested model release.## Required asset references`qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors`, `minimax_h3_video_vae_fp16.safetensors`, `minimax_h3_audio_vae_fp32.safetensors`, `minimax_h3_latent_upscaler_3d_fp16.safetensors`, `minimax_h3_fl2va_pruned_int8_convrot.safetensors`.The VDN composers reference OpenVDN/vdn-minimax-h3 and stage_dmd_8nfe; refinement plans store4. These describe two-pass intent, not measured evaluation counts, a twelve-step ordinary sampler, or a verified distilled-LoRA filename. Reconcile/audio-audit settings are requested policies, not passed tests. Resolve exact node versions and assets before queueing; no installation, GPU, speed, lip-sync, voice-preservation, or license verification is included.## Saved configuration44 nodes/80 links. Uploaded audio feeds a window whose duration comes from the loader-reported length. Separated vocals are encoded and given a zero noise mask. The LOW template is split before concatenating its video latent with the injected audio.## How to prepareReplace the image/audio selections and inspect the actual window. Selector47 supplies544×736, overriding stale960/544 labels. The ×2 upscale suggests portrait HIGH controls around1088×1472, not a fixed landscape canvas. Inspect both LOW and HIGH decode/trim/save chains; both saves are active and share a prefix.## Known limitationsDecoded generated-audio outputs are unconnected; selected original audio supplies both trims. This routing does not prove bit-exact passthrough, lip sync, voice preservation, or successful temporal fitting. Check actual dimensions, duration, and output filenames.Inspect actual output dimensions, duration, sound, input alignment, and unintended edits. Do not turn a node title, saved preview, or audit setting into a claim of a successful render.## Files and version note- 02: [Workflow JSON]() · [Selected cover]()Historical VHS video-preview metadata is omitted from distribution copies. Other data is retained, including point-editor images where present: inspect their contents and rights before any publication. Source originals are unchanged. Selected covers are not verified inference results. Input media, model weights, and VDN artifacts are not bundled.This is new local descriptive copy, not a remote publication or an announcement of a model upgrade. Source discrepancies are documented, not silently corrected.## Tutorial, attribution, and rights[Source-provided tutorial](https://www.youtube.com/watch?v=WdAaPq2Xa0I), not live-verified here. The workflows are supplied through AI-KSK. Verify workflow, node, model, source-video, image, voice, and artwork rights separately; no universal commercial permission is asserted.Source-listed resourcesAdditional entry points listed by the episode source. Some contain referral/affiliate codes; no bonus, price, quota, or speed is asserted here.<ul><li>RunningHub registration (international, invite code rh-v1111): <a href="https://www.runninghub.ai/?inviteCode=rh-v1111" rel="ugc">https://www.runninghub.ai/?inviteCode=rh-v1111</a></li><li>RunningHub registration (China-site mirror, same invite code): <a href="https://www.runninghub.cn/?inviteCode=rh-v1111" rel="ugc">https://www.runninghub.cn/?inviteCode=rh-v1111</a></li><li>AI-KSK workflow catalog on the international site: <a href="https://www.runninghub.ai/zh-cn/user-center/1846601088051884033/userPost?inviteCode=rh-v1111" rel="ugc">https://www.runninghub.ai/zh-cn/user-center/1846601088051884033/userPost?inviteCode=rh-v1111</a></li><li>Quick-create entry (international site): <a href="https://www.runninghub.ai/ai-generator/2046514150500524034?inviteCode=rh-v1111" rel="ugc">https://www.runninghub.ai/ai-generator/2046514150500524034?inviteCode=rh-v1111</a></li><li>VIP rights entry (international site): <a href="https://www.runninghub.ai/vip-rights/3?inviteCode=rh-v1111" rel="ugc">https://www.runninghub.ai/vip-rights/3?inviteCode=rh-v1111</a></li><li>Workflow model files (Quark netdisk, China): <a href="https://pan.quark.cn/s/9ec095b95838" rel="ugc">https://pan.quark.cn/s/9ec095b95838</a></li><li>Cloud GPU signup (Compshare; referral code in the link): <a href="https://passport.compshare.cn/register?referral_code=A7fTQUYY9lCDFJwA6MMvoq" rel="ugc">https://passport.compshare.cn/register?referral_code=A7fTQUYY9lCDFJwA6MMvoq</a></li><li>Cloud image index for Compshare (Notion): <a href="https://www.notion.so/2b72459a3ac6808f8282f2c687ed8f52" rel="ugc">https://www.notion.so/2b72459a3ac6808f8282f2c687ed8f52</a></li><li>Support the channel (Ko-fi): <a href="https://ko-fi.com/aiksk" rel="ugc">https://ko-fi.com/aiksk</a></li></ul>

Files with SHA256: dc20c8933030193bfe68f9140e2888ba733daea60512753740a564c6ea9f15b4