Most AI video looks fake because the face keeps changing between shots. The fix is not a longer prompt. It is a reference image the model treats as a hard rule. Here is the five-step Seedance 2.0 workflow, the tag system, and three prompts you can copy.
You want video for your brand. Product shots, a talking founder, a scene that looks like it cost money to shoot. So you open an AI video tool, type a long prompt, and hope.
The first frame looks great. Then the face shifts. The lighting drifts. By the end the person on screen is not the person you started with. You cannot post it. You cannot put your name on it.
That problem has a name. It is called drift, and it is why most AI video still looks off. The good news is the fix is boring and repeatable. You anchor the model to a reference image, and the character holds steady across every frame.
What a reference image actually does
A reference image is a consistency anchor. You hand the model a photo of the face, the style, or the motion you want, and it holds that constant across the whole clip. The model reads the reference as a hard constraint, not a suggestion.
That single shift is the whole game. Consistency is not something you coax out of a prompt with better wording. It is a job the reference image does. Lock the reference once, and your character stays the same across every video you make.
Why this matters for a business
Think about what a drifting face costs you. Every clip looks like a different company made it. The founder in Monday's video is not quite the founder in Friday's. The product changes shape between the ad and the landing page. Viewers do not consciously notice it. They just feel that something is off, and off does not sell.
Brand video runs on recognition. The same face, the same look, the same treatment, over and over, until people know the work is yours before they read the name. A locked reference is what gives you that. You set the anchor once and reuse it across a whole content calendar.
The payoff is blunt. You make a week of video from your desk with no shoot to book, no crew to schedule, and no location to rent, and every clip still looks like it came from the same brand. That is the trade the setup below buys you.
Step one: Open Seedance 2.0
Go to higgsfield.ai. Open the AI Video section. Select Seedance 2.0.
That is the room you work in. Everything below happens here, so get comfortable with where the upload button and the prompt field sit before you start.
Step two: Upload your anchor photo
This is the most important file you will pick, so pick it well.
Use a sharp, front-facing, well-lit photo of the face. This is your identity anchor. The model builds every frame off it, which means a clean photo gives clean output and a bad one poisons the whole batch.
Your first real run: grab one good headshot of yourself or your product and upload it. Nothing else. One clean anchor is enough to see the effect.
The mistake that makes it fail: feeding it a blurry or side-lit photo. A soft, shadowed source gives soft, inconsistent results, and no amount of prompt wording fixes a weak anchor.
Step three: Tag every reference
Seedance reads tags as hard rules. Each tag does one job, so you tell the model exactly what a given file is for. A photo tagged as a face gets locked as a face. The same photo with no tag is just a hint.
Here is what each tag locks.
| Tag | What it locks |
|---|---|
| @character | A face or person. Holds the face shape, skin tone, and look steady across the clip. |
| @style | A film still or color reference. Sets the lighting, palette, and mood. |
| @motion | A short clip. Copies the camera movement. |
| @audio | A voice or music clip. Shapes the rhythm and sound. |
Keep the roles clean. Text defines the world. Images lock identity. Video guides movement. Audio shapes sound. Each input has one job. Do not ask a prompt to do a reference image's work.
Your first real run: tag your anchor photo @character and leave the other tags empty for now. One labeled reference is enough to prove the lock works before you layer in style and motion.
The mistake that makes it fail: uploading a reference and forgetting to tag it. An untagged file is a hint, not a rule, so the model is free to ignore it. If your face is not holding, check the tags first.
Step four: Add angles to lock the character
One front photo works. If you want a tighter lock, feed it more than one view.
A front view, a side profile, and a three-quarter angle together build a character the model can turn without drifting. It can then rotate the head, walk past camera, or glance away, and still hold the same face. You can combine up to nine images, three video clips, and three audio clips in a single generation.
Your first real run: once your single-photo test looks right, add a second angle and generate the same shot again. Watch how much steadier the turn gets.
The mistake that makes it fail: loading all nine references before you know what any one of them does. Start with one clean @character photo. Add a second angle only when the output needs a tighter lock. Fewer, better references beat more, every time.
Step five: Write the prompt and generate
Now the prompt. There is one order, and you use it every time.
Seedance reads the opening of your prompt to lock the subject and the main action before it reads the rest. So lead with who is in the frame and what they are doing. Then build outward.
| Order | What it does |
|---|---|
| Subject | Who or what is in the frame. Name it first. |
| Action | What they are doing. One clear movement. |
| Environment | Where it happens. The setting. |
| Camera | Shot type and move. "Slow push-in," "medium close-up." |
| Lighting | The light setup. "Soft side light," "high contrast." |
| Style | The visual treatment. "Cinematic, film grain, teal and warm tones." |
| Constraints | What to keep out. "No text, no logos, no extra people." |
The template you fill in every time:
Copy this.
@character [your reference], [action], in [environment]. [Camera move]. [Lighting]. [Style]. [Constraints].
Start with one or two references. Add more only once you see how each one changes the output. Then generate.
Three prompts to copy
Swap in your own reference and go. Each of these maps to a job a business actually needs done.
Personal-brand cinematic, for a founder clip or a talking-head open:
Copy this.
@character stands in a dark modern studio, arms crossed, looking straight into the lens. Slow push-in on a medium close-up. Soft key light from the left, deep shadows on the right. Cinematic, shallow depth of field, teal and warm skin tones. No text, no logos.
Product hero, for an ad or a landing-page loop:
Copy this.
@product sits on a black stone pedestal in a dark studio. A single drop of water rolls down the surface. Medium close-up with a slow circular dolly. Soft side lighting, high-contrast reflections, luxury ad look. No people, no text.
Location establishing, for the scene that sets up the story:
Copy this.
@character walks through a rain-soaked city street at night, neon signs glowing behind. Tracking shot from the side, steady pace. Wet reflections, cool blue light with warm neon accents. Cinematic, filmic grain. No on-screen text.
The pre-generation checklist
Run this before you hit generate. Five checks, in order.
- Your anchor photo is sharp, front-facing, and well-lit.
- Each reference has a clear role tag: @character, @style, or @motion.
- You started with fewer, better references, not nine at once.
- Your prompt leads with the subject and the action.
- You ran one short test before committing to the full set.
That last one saves the most time. A short test clip tells you whether the anchor holds before you spend a generation on the whole scene.
When the face still drifts
Sometimes you run all five steps and the face still slides. Do not argue with a bad generation and do not pile on more prompt words to force it back. The problem is almost always an input, not the wording.
Fix it by subtraction. Strip the generation back to a single @character photo and run it again. If the face holds now, you know one of the references you removed was the culprit. Add them back one at a time, generating between each, until the one that breaks the lock shows itself. Remove that file or swap it for a cleaner version.
If a single clean @character photo still will not hold, the anchor itself is the problem. Go back to step two and pick a sharper, more front-facing photo. The model can only be as consistent as the face you gave it.
The one mistake that makes it fail
Piling on nine references at once.
It feels productive. More inputs, more control. In practice it does the opposite. The model gets pulled in nine directions, the tags start fighting each other, and you cannot tell which file caused the mess. Start with one clean @character photo. Add a second angle only when you need a tighter lock.
Two honest limits before you build a week of content on this.
First, the anchor is a ceiling, not a floor. A weak source photo caps how good the output can get, and no prompt rescues it. Fix the photo before you touch the wording.
Second, this is a paid tool. A full batch of generations on Higgsfield has a real cost, so test small before you commit to a whole content run. If you want to try the reference-image idea before you pay for anything, start on a free path first. I walk through a free, open-source video generator here, then bring the same reference-image discipline back to Seedance once the workflow clicks.
Where to start
If you only do one thing this week: upload one clean headshot, tag it @character, and generate a single short clip using the personal-brand prompt above. Nothing else. One anchor, one shot, one test.
That run teaches you more than reading this twice. You will see the face hold, you will see where a soft photo hurt you, and you will know exactly what to fix on the next pass.
The order that works is always the same. Clean anchor first. One tag per file. Subject and action at the front of the prompt. Test small, then scale. Get those four right and the drift that makes AI video look fake is gone. What is left is video of yourself, on brand, made from your desk, with no camera in the room.