xaddwell / Stable-Edit

Text-based real image editing with stable diffusion models

Stable-Edit

This is the PyTorch implementation for image editing framework as described in:

Stable-Edit: Text-based real image editing with stable diffusion models

We address the consistency image editing by inversing both image and text embedding.

1. Model Structure

_{Overview of the proposed stable editing framework.}

Specifically, provide with the input image $X$ and target text $Y$, we first learn the inversed text embedding $e_{opt}$ of image $X$. Then, we combine the target text embedding $e_{tgt}$ from text encoder and inversed text embedding $e_{opt}$ with cross attention. Next, we learn the inversed image latents according to image latents $h_{in}$ of image $X$ from VAE according to DDIM scheduler. Finally, we forward the standard text-to-image generation.

2. Cases

_{Cases for the image editing.}

This repository is based on diffusers.