Generating Images with AI Diffusion Models
What's behind the magic?
How do diffusion models associate images with words?
How can diffusion models generate hybrids of different images?
Associating images with words
Uploading images
Viewing images
Captions captured in LAION/CLIP dataset
Generating hybrid images
If generative AI creates by "averaging," can we just average the relevant images in our dataset to create new hybrids?
martian_daisy_10
martian_daisy_20
AI can't generate hybrid images just by averaging pixels ☹️
A better approach to generating hybrid images
The solution turns out to be related to the problem of
removing noise from a photo
🤔
The reason is that the best de-noising algorithms work by
segmenting the image into shapes
.
denoise_10
denoise_20
denoise_30
denoise_40
denoise_50
denoise_60
denoise_70
Finding noise means finding shapes
A de-noising algorithm can find shapes in noise 💪
💡 So instead of averaging the original photos, we could
Add noise to both images
Average the noisy results
De-noise to find new shapes
martian_daisy_30
martian_daisy_40
martian_daisy_50
The transporter analogy
Star Trek Voyager,
Tuvix
An infamous Star Trek episode about a transporter malfunction offers an analogy:
Creating a hybrid of two people is
easier while their molecules are scrambled
than in their original form.
tuvix_10
tuvix_20
tuvix_30
tuvix_40
Results of averaging noisy images
/
Go to slide: