Hey everyone! 👋 I'm super excited to finally be creating my first HeyGen avatar. I've seen some incredible results here on the forum, but also some that look a bit... off. I really want to nail this on my first try.
I'm planning to record my source video this weekend and want to make sure my setup is perfect. I've read the official guidelines, but I'd love to hear from your real-world experience. What are the *unwritten* best practices for lighting and camera angle that make a huge difference?
Specifically, I'd love to know:
* **Lighting setup:** Should I use a ring light directly in front, or is a two-point setup from the sides better? What about dealing with shadows?
* **Camera angle:** Is eye-level truly the best, or is a very slight from-above angle more flattering and accurate for the model?
* **Background:** Does a plain wall work best, or does a specific color (like grey or blue) help with the capture?
* **Common pitfalls:** What are the top mistakes beginners make that ruin an otherwise good recording?
I'm a total beginner at this kind of video recording, so detailed walkthroughs or even photos of your home setups would be amazing. My goal is a professional-looking talking head avatar for internal training videos.
Thanks in advance for sharing your wisdom! This community has already been so helpful as I dive into data and analytics, and now I'm thrilled to learn about this side of things too.
I run the video processing pipelines for a mid-market e-learning platform. We've produced over 500 training avatars, mostly with HeyGen, for internal product demos and client-facing materials.
* **Lighting setup:** Ring lights are trash. They create flat, lifeless light and a visible ring reflection in the eyes. Use a two-point softbox setup. Key light at 45 degrees, fill light opposite, both diffused. This creates subtle depth and kills harsh shadows under the chin and nose. A cheap LED panel with a shower curtain in front of it works better than any $100 ring light.
* **Camera angle:** Eye-level is technically correct, but a slight high-angle (lens just above your eyebrows, tilted down) is more forgiving. It prevents the "double chin" effect the model can exaggerate. Don't go overboard - more than 15 degrees and you'll get weird forehead distortion.
* **Background:** Plain wall is fine, but color matters. Avoid white (causes blown-out exposure) and green (you're not chroma-keying). A mid-tone grey or muted blue works best. The capture needs consistent contrast between you and the backdrop.
* **Common pitfalls:** The biggest mistake is inconsistent lighting between sessions. If you need to re-record a clip, you must replicate the exact light position and temperature, or the generated avatar will have a flickering "uncanny valley" effect. Second is using a webcam's auto settings - lock your white balance and exposure manually.
You want a ring light because it's easy. Don't. Spend an hour rigging two cheap lights with diffusion. The model output depends entirely on clean, consistent input video, and a proper lighting setup is 80% of that battle.
-- old school
Oh man, I feel you on wanting to get it right the first time. I just tried this myself last week.
For a background, a plain, matte wall is solid. I used a light grey one and it worked fine. The big thing is making sure there's a lot of contrast between you and the wall, so avoid wearing a shirt that's the same color.
The pitfall I almost hit was wardrobe. I read somewhere to avoid really thin stripes or busy patterns because they can cause weird moire effects or get distorted in the model. Stuck with a simple, solid colored polo. Good luck with your shoot
Your advice is good, but producing 500 avatars costs a fortune if you're not using commitment discounts on the compute side. That softbox setup is perfect, but have you optimized the actual rendering pipeline on GCP or AWS for those 500 videos? Spot instances and sustained use discounts matter more than your lighting gear's price tag.
What's your monthly cloud bill for this? I bet you could cut it by 40%.
show me the bill
I see your point about cloud costs, but for someone just starting out like the OP, I think that's a different stage of the problem.
The initial advice was about getting a good source video, which is a fixed, one-time cost. If your lighting is bad, no amount of compute optimization will fix the avatar. You can't go back and re-record 500 times if the source material is poor.
That said, for the scale you're talking about, it makes sense. How do you manage the trade-off between using spot instances for cost and the potential for them being terminated during a longer render? Do you find you need to fragment your jobs?
Thanks, this is super practical advice. The two-point setup you described makes total sense for avoiding that flat look. I'm definitely ditching my ring light plan.
The high-angle tip about preventing a double chin effect is really helpful. I probably would have stuck with strict eye-level. Do you find this works for people with different face shapes, or is it just a universal good move?
Great advice already on the setup. The high-angle tip is solid, and for most faces it helps. One thing I'd add: record a short test clip with your chosen angle and watch it back on a phone screen before doing the full session. Your laptop monitor can lie to you.
Also, watch out for inconsistent lighting during your recording. If you're using natural light from a window, even a passing cloud can change the color temperature mid-clip, and that's a pain to fix later. Block out any ambient light and stick to your controlled setup for the whole take.
For background, a plain wall is fine, but make sure it's not glossy. A matte finish prevents any subtle reflections that might confuse the capture. Light grey or a muted blue usually works well for separation.
You're right, the source material is the bottleneck. A perfect cloud setup can't fix bad video.
On the fragmentation point for spot instances: we don't. For renders over 90 seconds, we found job fragmentation added more overhead and debugging time than it saved. We just use on-demand for those long jobs and eat the cost. The spot instances are strictly for the batch processing of sub-30 second clips.
Benchmarks don't lie.