Hacker News Reader

Comment by peteralaoui | Hacker News Reader

Indeed, and I consider myself part of that group! Thanks for the feedback.

I can probably lower this requirement to a minimum of 5, while mentioning to the user that the more pictures the merrier for the model

View on HN

I wonder if you could change this to allow a video input via prompts to turn your head in various angels, like registering for facial recognition. Then it might not trigger the distaste for photos AND you could get better data for face placement.

That seems like a good idea.

Similar tools I’ve tried seem to want photographs taken under different lighting scenarios in different environments. Photos all taken in the same room, apparently, produce worse results.

It's a cool idea to simplify the input as you said, though as @dorkwood pointed out this would lead to bad results with a lot of overfitting, this model needs a lot of variation to only register your facial features as "you", and disentangle that from the surroundings or even from your clothes.

(this is due to how this model learns, maybe a future model will alleviate that)

I'm only just getting into this space, but immediately my thought process based on what I've learned so far would be to set up the pipeline such that It would split the video into keyframes with ffmpeg, and then use the clipseg to generate a mask of the person/background and use that to randomize the backgrounds with img2img! It could probably be done for the clothes as well...

Reply on news.ycombinator.com