How to Create AI Talking Head Videos with OmniHuman v1.5

Your first OmniHuman v1.5 talking head video takes about ten minutes end to end — and it costs exactly $9.60. This tutorial shows you every step, every input decision, and every quality trick we have learned from generating thousands of talking head videos on the platform.

What You Need Before You Start

Three inputs, no exceptions:

1. A portrait photo — head and shoulders, well-lit, minimum 512x512 pixels
2. An audio file — MP3, WAV, or M4A, clean speech, up to 60 seconds
3. A scene prompt — a short description of the background and lighting

Plus a Seedance account with at least 960 credits. New accounts get 50 free credits on signup, but a talking head costs 960, so you will need at least a $10 Starter pack to run your first video.

Step 1: Pick or Create the Right Photo

The photo is the single biggest quality lever in the entire workflow. Feed the model a great photo and it will return a great video. Feed it a rushed phone snap and the output will reflect that.

What Makes a Great Reference Photo

• Even lighting. Diffused natural light or soft studio light. No harsh shadows across the face.
• Clear face. Eyes open, no glasses glare, no hair covering features, no hands near the face.
• Right framing. Head and shoulders visible. Face occupies at least 30% of the frame.
• Neutral expression. A relaxed or mildly pleasant face gives the model the most flexibility.
• Clean background. High contrast between subject and background helps the model isolate identity.
• Sharp resolution. 1024x1024 or larger. No motion blur.

Don't Have a Photo? Generate One

If you need a reference photo you do not already have — for a virtual spokesperson, a persona, or a character — use Seedream to generate one. Prompt for "professional headshot of [description], soft studio lighting, neutral background, looking at camera" and pick the cleanest result.

Formats

JPEG and PNG both work. Transparent backgrounds are acceptable. The platform accepts photos up to several megabytes.

Step 2: Prepare Clean Audio

Your audio determines the lip sync, the facial expression, and the gesture pattern. The cleaner the audio, the more the model has to work with.

Recording Options

Record yourself. A quiet room and a decent USB microphone will produce audio clean enough for OmniHuman. Speak at a natural pace, with natural pauses. Do not over-enunciate.

Use a TTS service. Any quality text-to-speech tool works — ElevenLabs, Play.ht, Google, Amazon Polly. Export at 44.1kHz or 48kHz mono or stereo.

Extract from existing audio. A podcast segment, a webinar clip, or a voiceover recording all work as long as speech is isolated.

Audio Dos and Don'ts

• Do trim leading and trailing silence
• Do normalize audio levels to prevent clipping
• Don't leave music or heavy ambient sounds in the mix
• Don't use reverb-heavy room recordings
• Don't exceed the duration cap: 60s for 720p, 30s for 1080p

Step 3: Write a Scene Prompt That Helps

The scene prompt describes the visual environment around your presenter. It does not describe the person — the model gets that from the photo.

Good Prompt Structure

A strong prompt includes four elements:

1. Background — where is the person?
2. Lighting — how is the scene lit?
3. Framing — how close is the camera?
4. Style — any notes on tone or finish?

Example Prompts That Work

Modern corporate office with large windows, soft natural lighting from the left, medium close-up framing head and shoulders, professional broadcast style.

Minimalist studio setting with a soft gradient backdrop, even studio lighting, medium shot, clean and contemporary look.

Warm home office with bookshelves in the background, golden afternoon sunlight, medium close-up, natural and approachable tone.

What to Avoid in Prompts

• Do not describe the person (hair color, age, outfit) — the photo handles that
• Do not contradict the photo (no "white background" if the photo has a black background, unless you really want the model to replace it)
• Do not use vague phrases like "good lighting" — pick a specific light source
• Do not overload the prompt with more than five or six details

Step 4: Upload and Configure

Open seedance.it.com and select OmniHuman v1.5, or go straight to the model page.

Upload Order

1. Reference photo — drag the portrait into the image slot
2. Audio file — drop the speech file into the audio slot
3. Scene prompt — paste your scene description

Configure the Settings

• Resolution: 720p for drafts or clips longer than 30s, 1080p for professional final renders
• Turbo mode: on for iteration, off for final quality
• Review credit cost: the UI will confirm 960 credits before you generate

Step 5: Generate and Review

Click generate. Standard mode takes several minutes. Turbo mode is faster. You will get a notification when the video is ready.

Reviewing the Output

Play the video at full size and check these four things:

1. Lip sync — do mouth shapes match the phonemes?
2. Identity — does the face look like the reference throughout?
3. Gesture naturalism — does motion feel organic, not robotic?
4. Background consistency — does the scene match your prompt?

If any of these feel off, the fix is almost always in the inputs, not a re-roll. Swap photo, clean audio, or tighten the prompt.

Ready to try OmniHuman v1.5? Start creating free →

Want a presenter like this? Try OmniHuman free →

Advanced Tips from Real Production Workflows

Use Turbo for Exploration, Standard for Finals

Turbo mode still costs 960 credits per generation, but it trims the wait. Use it to test different prompts or audio takes quickly, then render the final version in standard mode.

Match Audio Emotion to Reference Expression

If your photo shows a neutral face but your audio is highly animated, the model will generate lots of motion. If the audio is calm and the photo is smiling, expect subtle variation. Match them for predictable results.

Split Long Content into Segments

Need a 90-second video? Split the audio into two segments (e.g., 45s each), generate two 720p clips, and stitch them in any video editor. Identity stays consistent because you are using the same reference photo.

Prep Multiple Photos for the Same Person

Having three or four reference photos of the same person — different outfits, different lighting, different expressions — lets you match the photo to the content tone. A warm anecdote gets a warmer photo; a corporate update gets the headshot.

Keep a Prompt Library

Save scene prompts that worked. "Minimalist studio gradient" or "modern office window light" can be reused across projects for visual consistency.

Common Mistakes and How to Fix Them

Lip sync feels slightly off

• Check audio quality. Noise, compression artifacts, or low bitrate hurt phoneme extraction
• Check duration. Clips at exactly the cap can get clipped on the last frame — aim for a second under
• Try a cleaner recording or re-export at higher bitrate

Face looks "plasticky" or too smooth

• Use a higher-resolution photo. Sub-512px input produces soft output
• Adjust prompt lighting to "natural" instead of "studio" for more texture

Gestures look too subtle

• Use more expressive audio. Monotone speech produces minimal gesture variation
• Select a photo with slightly open posture rather than stiff frontal framing

Background doesn't match the prompt

• Be more specific. "Office" is vague; "modern office with a bookshelf behind the subject and a large window on the left" is actionable
• Match photo context. The model weighs the photo background as a reference

Cost Math for Different Projects

Every OmniHuman v1.5 video costs 960 credits (~$9.60). Here is what that looks like in practice:

| Project | Videos Needed | Total Cost | |---|---|---| | Single LinkedIn post | 1 | $9.60 | | Five-video welcome series | 5 | $48 | | Ten-lesson course | 10 | $96 | | Weekly updates for a year | 52 | $499.20 |

Compare that to HeyGen's $24-$48 per month (even in months you create nothing) or Synthesia's $30-$90 per month. At low and moderate volumes, OmniHuman saves hundreds per year. See the full pricing breakdown for every tier.

Your First Video Checklist

• [ ] Portrait photo, 1024x1024 or larger, well-lit, neutral expression
• [ ] Clean audio file, under 60 seconds, no music or reverb
• [ ] Scene prompt with background, lighting, framing, style
• [ ] Seedance account with at least 960 credits
• [ ] Resolution choice (720p or 1080p)
• [ ] Turbo mode decision
• [ ] Open OmniHuman v1.5 and generate

Once you have shipped your first talking head, explore the lip sync technical guide, the multilingual workflow guide, and use-case specific tutorials like news anchors and corporate training.

Ready to try OmniHuman v1.5? Start creating free →