# OmniRef 8-Step: Image + Audio Driven Digital Human
One portrait plus one audio track drives the mouth, the head motion and the eye line. No keyframes,
no masks, no manual lip-sync pass. The eight-step distilled sampler keeps generation time short.
## The chain
- **Input A**: a single portrait image (front-facing works best, but three-quarter views are fine).
- **Input B**: a speech audio track — from CosyVoice3 or your own recording.
- **Output**: a talking-head clip where the mouth, head and gaze follow the audio timing.
## Why the 8-step sampler
The full sampling chain is slow. The distilled eight-step version gives you a usable shot in a
fraction of the passes, which matters when you are iterating on a portrait or an audio take. If a
particular shot needs more fidelity, that is the case for the full chain — but start here.
## Running next to Pika
Episode 329 ships two digital-human routes on purpose: this OmniRef eight-step chain and a Pika
six-step chain. They fail differently — one drifts on head rotation, the other on the eye line.
Run both on the same image and audio before committing to one; the difference shows up in ten
minutes and is worth the comparison.
## Tips from the episode
- Crop the portrait to a similar aspect ratio as the output, otherwise the model spends capacity on
filling.
- The first and last frame are where identity drift shows; check those frames specifically.
- If the mouth opens too wide, the audio track is peaking — normalise it first.
- Keep clips short (under ~30 s) while you are dialling in settings.
---
**About this workflow**
Part of AI-KSK episode 329, which builds a talking-head digital human from a single image plus an audio track and finishes it at 2K. This is a ComfyUI workflow export: it references models that must be present in your ComfyUI environment. On RunningHub the cloud instance already has them installed.
It is a tutorial workflow, not a packaged product. Node parameters are the ones used in the episode; adjust frame count, sample steps and resolution for your own hardware.
**License / credit:** independent AI-KSK tutorial release, not affiliated with or endorsed by MiniMax, Alibaba, ByteDance or RunningHub. Model licences remain with their respective authors — check each model's own licence before commercial use.
One portrait plus one audio track drives the mouth, the head motion and the eye line. No keyframes,
no masks, no manual lip-sync pass. The eight-step distilled sampler keeps generation time short.
## The chain
- **Input A**: a single portrait image (front-facing works best, but three-quarter views are fine).
- **Input B**: a speech audio track — from CosyVoice3 or your own recording.
- **Output**: a talking-head clip where the mouth, head and gaze follow the audio timing.
## Why the 8-step sampler
The full sampling chain is slow. The distilled eight-step version gives you a usable shot in a
fraction of the passes, which matters when you are iterating on a portrait or an audio take. If a
particular shot needs more fidelity, that is the case for the full chain — but start here.
## Running next to Pika
Episode 329 ships two digital-human routes on purpose: this OmniRef eight-step chain and a Pika
six-step chain. They fail differently — one drifts on head rotation, the other on the eye line.
Run both on the same image and audio before committing to one; the difference shows up in ten
minutes and is worth the comparison.
## Tips from the episode
- Crop the portrait to a similar aspect ratio as the output, otherwise the model spends capacity on
filling.
- The first and last frame are where identity drift shows; check those frames specifically.
- If the mouth opens too wide, the audio track is peaking — normalise it first.
- Keep clips short (under ~30 s) while you are dialling in settings.
---
**About this workflow**
Part of AI-KSK episode 329, which builds a talking-head digital human from a single image plus an audio track and finishes it at 2K. This is a ComfyUI workflow export: it references models that must be present in your ComfyUI environment. On RunningHub the cloud instance already has them installed.
It is a tutorial workflow, not a packaged product. Node parameters are the ones used in the episode; adjust frame count, sample steps and resolution for your own hardware.
**License / credit:** independent AI-KSK tutorial release, not affiliated with or endorsed by MiniMax, Alibaba, ByteDance or RunningHub. Model licences remain with their respective authors — check each model's own licence before commercial use.
Description
# OmniRef 8-Step: Image + Audio Driven Digital Human
One portrait plus one audio track drives the mouth, the head motion and the eye line. No keyframes,
no masks, no manual lip-sync pass. The eight-step distilled sampler keeps generation time short.
## The chain
- **Input A**: a single portrait image (front-facing works best, but three-quarter views are fine).
- **Input B**: a speech audio track — from CosyVoice3 or your own recording.
- **Output**: a talking-head clip where the mouth, head and gaze follow the audio timing.
## Why the 8-step sampler
The full sampling chain is slow. The distilled eight-step version gives you a usable shot in a
fraction of the passes, which matters when you are iterating on a portrait or an audio take. If a
particular shot needs more fidelity, that is the case for the full chain — but start here.
## Running next to Pika
Episode 329 ships two digital-human routes on purpose: this OmniRef eight-step chain and a Pika
six-step chain. They fail differently — one drifts on head rotation, the other on the eye line.
Run both on the same image and audio before committing to one; the difference shows up in ten
minutes and is worth the comparison.
## Tips from the episode
- Crop the portrait to a similar aspect ratio as the output, otherwise the model spends capacity on
filling.
- The first and last frame are where identity drift shows; check those frames specifically.
- If the mouth opens too wide, the audio track is peaking — normalise it first.
- Keep clips short (under ~30 s) while you are dialling in settings.
---
**About this workflow**
Part of AI-KSK episode 329, which builds a talking-head digital human from a single image plus an audio track and finishes it at 2K. This is a ComfyUI workflow export: it references models that must be present in your ComfyUI environment. On RunningHub the cloud instance already has them installed.
It is a tutorial workflow, not a packaged product. Node parameters are the ones used in the episode; adjust frame count, sample steps and resolution for your own hardware.
**License / credit:** independent AI-KSK tutorial release, not affiliated with or endorsed by MiniMax, Alibaba, ByteDance or RunningHub. Model licences remain with their respective authors — check each model's own licence before commercial use.
One portrait plus one audio track drives the mouth, the head motion and the eye line. No keyframes,
no masks, no manual lip-sync pass. The eight-step distilled sampler keeps generation time short.
## The chain
- **Input A**: a single portrait image (front-facing works best, but three-quarter views are fine).
- **Input B**: a speech audio track — from CosyVoice3 or your own recording.
- **Output**: a talking-head clip where the mouth, head and gaze follow the audio timing.
## Why the 8-step sampler
The full sampling chain is slow. The distilled eight-step version gives you a usable shot in a
fraction of the passes, which matters when you are iterating on a portrait or an audio take. If a
particular shot needs more fidelity, that is the case for the full chain — but start here.
## Running next to Pika
Episode 329 ships two digital-human routes on purpose: this OmniRef eight-step chain and a Pika
six-step chain. They fail differently — one drifts on head rotation, the other on the eye line.
Run both on the same image and audio before committing to one; the difference shows up in ten
minutes and is worth the comparison.
## Tips from the episode
- Crop the portrait to a similar aspect ratio as the output, otherwise the model spends capacity on
filling.
- The first and last frame are where identity drift shows; check those frames specifically.
- If the mouth opens too wide, the audio track is peaking — normalise it first.
- Keep clips short (under ~30 s) while you are dialling in settings.
---
**About this workflow**
Part of AI-KSK episode 329, which builds a talking-head digital human from a single image plus an audio track and finishes it at 2K. This is a ComfyUI workflow export: it references models that must be present in your ComfyUI environment. On RunningHub the cloud instance already has them installed.
It is a tutorial workflow, not a packaged product. Node parameters are the ones used in the episode; adjust frame count, sample steps and resolution for your own hardware.
**License / credit:** independent AI-KSK tutorial release, not affiliated with or endorsed by MiniMax, Alibaba, ByteDance or RunningHub. Model licences remain with their respective authors — check each model's own licence before commercial use.