What is the Personal Model?
When you train a Personal Model, HeyGen builds a dedicated image model based on your unique appearance. That's why we ask you to upload multiple photos: it allows the model to learn your features in much more depth than a single image ever could.
The result is significantly higher identity accuracy and consistency across every look you generate. The prompt tells HeyGen what to create, and the Personal Model tells HeyGen who should be in it.
If you want your generated looks to consistently look like you across different outfits, backgrounds, and scenes, the Personal Model is the feature you need to know.
How to train your Personal Model
Training your Personal Model takes just a few steps and around 10 to 15 minutes to complete. Training costs 60 credits and is available on paid plans only.
Navigate to your avatar and open the Generate Looks panel. You'll see a Personal Model toggle in the prompt bar. Click it to get started.
The "Train your personal model" panel will open. From here you can select from your existing uploaded looks, your AI generated looks, or upload new photos directly. Use the tabs at the top to filter between them.
As you select photos, a counter on the right tracks your progress. You need a minimum of 10 photos to train, and 30 or more for the best results. The quality indicator will update as you add more, letting you know if your current selection is likely to produce strong results.
Once you're happy with your selection, click Train. You'll be able to see how many credits you will be spending for this training. Once complete, your Personal Model will be active and ready to use when generating new looks.
What makes a good training photo?
The Personal Model is designed to work from still photos. This is intentional: photos give the model the cleanest possible data about your appearance, allowing it to learn your bone structure, skin texture, expressions, and the details that make you recognizably you.
A good rule of thumb: include photos that cover the types of looks you might want to prompt for. If you plan to prompt for side-angle shots, include side-angle photos. If you want to prompt for specific expressions like surprised or serious, include real photos of you with those expressions. The more real data the model has to work with, the less it needs to make up.
Not all photos are equal though. For the best results, include:
A mix of close-ups and wider shots so the model can see your face across different framings
Different angles and expressions, such as straight-on, slight turn, smiling, neutral, and serious
A variety of outfits to help the model separate you from what you happen to be wearing
High-resolution, well-lit images that reflect how you currently look
Avoid group photos, hats, sunglasses, heavy filters, low-resolution images, or old photos that don't represent your current appearance.
Aim for 30 or more photos for the strongest results. The in-app quality indicator will let you know if you need to add more.
New to creating a Photo Avatar? Check out our Photo Avatar guide to get started.
What if my results still don't feel like me?
If you've trained your model but your generated looks are still off, the most effective fix is to retrain with a stronger photo set. Ask yourself:
Were the photos high resolution and well lit?
Did they show a variety of angles and expressions?
Were there any blurry, filtered, or inconsistent images in the mix?
You can retrain your model at any time from the base look panel. Upload better photos, hit Train, and you'll often see a noticeable improvement.
For more tips on prompting and refining your looks, check out our Generate Looks guide and the Prompt Like a Pro guide.
Pro tip: training a Personal Model on virtual characters
If you're working with a virtual character rather than a real person, here's a workflow to build strong consistency from scratch, even if you're starting with just one photo.
Generate 50 to 100 looks of your virtual character using HeyGen's default look generation, without a Personal Model trained yet.
Curate the best 10 to 30 images that show the strongest character consistency, and train your Personal Model on those selected photos.
Optionally, repeat the process to nail consistency even further. Generate another 50 to 100 looks with your Personal Model turned on, curate the top 30, and retrain on those.
Each round gives the model better data to work with, and the results improve with every iteration.