Ethan Follow @Ethan_smith_20 1 INTRODUCTION In image synthesis, diffusion models have adVanced to ing quality and mode coverage since introduction of 2020). They disrupt прабез adding noise thr. and g

Ethan Follow
@Ethan_smith_20
1 INTRODUCTION
In image synthesis, diffusion models have adVanced to
ing quality and mode coverage since introduction of
2020). They disrupt прабез adding noise thr.
and generate samples progr€ssive denoising through
process. Since their version provides nearly
(Song suitable for image editing,
images. Howev€r, simply latent variables
graded гези (Kim 2021). In\stead, they require comp
in the reverse\process or finetuning/models for an attribute.
Figure the existing approaches. Imd
of the guiding image with variables
Meng 2021). Though it provides it is
reflect among the ones in the guide and the unconditional restr, ana
magnitude of change. Classifier guidance manipulates images imposing gradients of a classifier
on the latent variables in the reverse process to match the target class (Dhariwal Nichol, 2021;
Avrahami 2022; Liu 2021). It requires training an extra classifier for the latent variables,
noisy images. Furthermore, computing gradients through the classifier during sampling is costly.
Finetuning the whole model can steer the resulting images to the target attribute without the above
problems (Kim 2021). Still, it requires multiple models to reflect multiple descriptions.
On the other hand, generative adversarial networks (Goodfellow 2020) inherently provide
straightforward image editing in their latent space. Given a latent vector for an original image, we
can find the direction in the latent space that maximizes the similarity of the resulting image with
a target description in CLIP embedding (Patashnik 2021). The latent direction found on one
author
23:42 2024/07/02 From Earth Views