| Summary: | Diffusion models, which utilize a multi-step denoising sampling procedure and
leverage extensive image-text pair datasets for training, have emerged as an innovative
option among deep generative models. These models exhibit superior
performance across various applications, including image synthesis and video generation.
In this thesis, we further explore applications of pre-trained diffusion
models other than text-to-image generation applications in a tuning-free manner.
In Chapter 1, we discuss image morphing between two real images via diffusion
models. Our approach, FreeMorph, is based on key insights regarding attention
interpolation and layout similarity in latent noise, which are critical for enhancing
morphing quality.
In Chapter 2, we discuss attention interpolation in diffusion models. This work
introduces a novel training-free technique named Attention Interpolation via
Diffusion (AID). AID has two key contributions: 1) a fused inner/outer interpolated
attention layer to boost image consistency and fidelity; and 2) selection of
interpolation coefficients via a beta distribution to increase smoothness.
|
|---|