Image to Human Avatar

This project aims to automatically process photos of humans into animatable human avatars.

Specifically, we aim to detect the human in the image, position it to our standard positioning, and mask it.

Source Image	Area of Interest	Cropped	Masked
High quality image of a human. Shoulder width $x$	Area of interest in $2.5x$ shoulder width where the center is the neck	Image after cropping	Image after masking

Real Examples

Source Image	Avatar	Masked Avatar

Conditions

• Input can be of any aspect ratio.
• There must be a human in the image (real or fictional).
• The human must be standing, with their arms down, facing the camera (TODO).
• The image must be safe-for-work.
• The creator of the avatar must have the right to use the image.

Usage

pip install git+https://github.com/sign/image-to-human-avatar

To then process an image:

from human_avatar.image_to_avatar import image_to_avatar
from PIL import Image

image = Image.open("example.jpg")
cropped, masked, pose = image_to_avatar(image)

masked.save("masked.png")

Pass include_pose=False to skip the full-body pose extraction (much faster).

To only crop a person out of an image of any size, without the avatar quality conditions:

from human_avatar.image_to_avatar import crop_image
from PIL import Image

image = Image.open("example.jpg")
crop = crop_image(image, resolution=256)  # 256x256 shoulder-centered crop

Web Server on Docker

docker build -t human-avatar .

docker run --rm -p 9874:8080 -e PORT=8080 \
  -v $(pwd)/output:/mnt/output \
  human-avatar

curl -X POST http://localhost:9874/ \
  -F "output=gs://output/flux" \
  -F "file=@assets/examples/flux/source.jpg"

Pass -F "pose=false" to skip writing the full-body pose.pose file (much faster).