A talking-head video normally means a camera, a person, a quiet room and a retake or three. This episode of AI-KSK does it from one still image plus an audio track, and gets to a 2K result without hand-animating a single frame.
Four ComfyUI workflows make up the chain, and all four run on RunningHub (cloud ComfyUI), so nothing heavy is installed locally.
The four workflows
| # | Workflow | What it does |
|---|---|---|
| 1 | Voice clone + TTS (CosyVoice3) | Clone a voice from a short reference clip, then generate the narration from text |
| 2 | Digital human, OmniRef 8-step | Image + audio drive the mouth, head motion and eye line; the distilled 8-step sampler keeps it fast |
| 3 | Digital human, Pika 6-step | The faster six-step route for the same task |
| 4 | Restoration + 2K upscale (SeedVR2) | Restore and upscale the finished clip to 2K |
- CosyVoice3: https://www.runninghub.ai/zh-cn/post/2119390483754557441?inviteCode=rh-v1111
- OmniRef 8-step: https://www.runninghub.ai/zh-cn/post/2119390479195639809?inviteCode=rh-v1111
- Pika 6-step: https://www.runninghub.ai/zh-cn/post/2119390488365547522?inviteCode=rh-v1111
- SeedVR2 2K: https://www.runninghub.ai/zh-cn/post/2119389472021323777?inviteCode=rh-v1111
All four pages returned HTTP 200 when verified on 2026-11-13.
Why two digital-human routes
OmniRef and Pika are not interchangeable settings — they are two different chains that happen to do the same job. OmniRef trades a little speed for stability; Pika is the six-step sprint. Run both on the same image and audio before you commit to one, because the failure modes differ: one drifts on head rotation, the other on the eye line.
That comparison is the reason the episode exists. Picking the "better" route from a spec sheet is guesswork; running both on your own material takes ten minutes.
What the pipeline looks like end to end
- Voice first. Clone from a short reference, generate the narration text. The audio track is the driver for everything downstream, so it decides the timing.
- Then the digital human. One portrait plus that audio. No keyframes, no masks, no lip-sync pass.
- Then the finish. SeedVR2 restoration and 2K upscale, so the talking head is not the softest thing in the edit.
The order matters: changing the narration later means regenerating the head, because the mouth is driven by the audio timing.
Workflows and files
The four workflow JSONs, the covers, and a manifest with SHA-256 hashes are in the companion repository:
https://huggingface.co/AI-KSK/minimax-h3-digital-human-329
They are ComfyUI exports that reference models present in the cloud instance. On RunningHub those models are already installed; locally you would need to match them yourself.
Five language editions
The episode released in five languages on the AI-KSK channel:
- English — https://www.youtube.com/watch?v=4qGqI3iaupg
- Japanese — https://www.youtube.com/watch?v=UkalemrsMdw
- Korean — https://www.youtube.com/watch?v=L3LjcQdIJvM
- Brazilian Portuguese — https://www.youtube.com/watch?v=hMDlfJcam8w
- Spanish — https://www.youtube.com/watch?v=x0g5o0QeQIA
Chinese original on Bilibili: https://www.bilibili.com/video/BV1qTWJ6kEw3/
Honest notes
- This is a tutorial release, not a product. Node parameters are the ones used in the episode.
- The narration in each language edition is AI-assisted translation plus AI voice synthesis, not a human studio recording. That is stated in every video description too.
- These workflows are not affiliated with or endorsed by MiniMax, Alibaba, ByteDance or RunningHub.
If you run both digital-human routes on the same clip, the differences show up immediately — that is the part worth watching.
Optional API platform - 88API (affiliate link)
Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.
https://88api.ai/sign-up?aff=usx3
Disclosure: I may earn a commission from eligible purchases after you register through this referral link.
Top comments (0)