AdaFace-Animate: Zero-Shot Human Subject-Driven Video Generation

Official demo for our working paper AdaFace: A Versatile Text-space Face Encoder for Face Synthesis and Processing.

❗️NOTE❗️

Support switching between three model styles: Realistic, Photorealistic and Anime. Realistic is less realistic than Photorealistic but has better motions.
If you change the model style, please wait for 20~30 seconds for loading new model weight before the model begins to generate images/videos.

❗️Tips❗️

You can upload one or more subject images for generating ID-specific video.
If the face loses focus, try enabling "Highlight face".
If the motion is weird, e.g., the prompt is "... running", try increasing the number of sampling steps.
Usage explanations and demos: Readme.
AdaFace Text-to-Image: AdaFace

Drag / Select 1 or more photos of a person's face

[Optional] Generate 4 images and select 1 image

Prompt

Try something like 'walking on the beach'.

Enhance the facial features by prepending 'face portrait' to the prompt

Highlight face

Init Image Strength

How much the init image should influence each frame. 0: no influence (scenes are more dynamic), 3: strongest influence (scenes are more static).

0 3

Final Strength of the Init Image

How much the init image should influence the end of the video

0 2

Base Model Style Type

Switching the base model type will take 10~20 seconds to reload the model

Guidance scale

If > 10, there may be artifacts.

1 12

Seed

0 10000

Uncheck for reproducible results

Randomize seed

Number of sampling steps. More steps for better composition, but longer time.

30 70

Generated Animation