Skip to content
Notifications
Clear all

Has anyone successfully trained an avatar on a low-quality webcam video?

1 Posts
1 Users
0 Reactions
0 Views
(@crusty_pipeline_redux)
Estimable Member
Joined: 4 months ago
Posts: 124
Topic starter   [#8126]

Tried it. It's predictably terrible. Everyone wants to use their laptop webcam from 2018 and expects movie-grade results. The marketing hype doesn't mention the hard requirements.

You need:
* Studio-level lighting. Your dim office with a window behind you is useless.
* A proper microphone. The webcam's built-in mic adds noise that messes with the training.
* A locked-off, stable shot. No hand-holding the laptop.
* A high bitrate source. Most conferencing software butchers quality before you even save the file.

Even if you meet all that, the output avatar often has these artifacts:
* Weird, dead-eye stare because the training data lacked enough eye movement variation.
* Uncanny valley mouth movements on plosives (p, b sounds).
* A persistent, blurry quality that no "enhancement" toggle fixes.

The real test? Run the output through a basic PSNR/SSIM check against a proper source. The numbers don't lie. My quick bash one-liner for comparing a frame:

```bash
convert input_frame.png reference_frame.png -metric PSNR -compare -format "%[distortion]" info:
```

You'll see the delta. It's big.

Save your credits. Garbage in, gospel out is still a myth.

-- old school


-- old school


   
Quote