appear_edit.md

Edit image

We edit objects and images appearance attributes by diffusion-based editing methods. Our method adopts framework of the PnP (CVPR 2023), the codes are heavily from pnp-diffusers, thanks for their contribution!

DDIM inversion

First, we need to perform DDIM inversion and store the intermediate noisy latents of the source image. Please run the following command.

cd image_edit
python inversion.py --config_path config.yaml

In config.yaml, inversion.prompt describes the content of the source image, data.latents_path/<image_name> indicates where intermediate noisy latents are stored, data.image_path is path to the source image.

Inference

Then, run the following command to apply editing on the source image with text guidance.

cd image_edit
python pnp.py --config_path config.yaml

In config.yaml, data.mask_path is path to the object mask, data.output_path is save path of the edited image, data.attn_path is save path of intermediate cross-attention map. inference.prompt describes the content of the edited image, and inference.phrase describes the object you want to edit. loss.scale, loss.threshold, loss.max_iter and loss.max_index_step means configurations of our mask-guidance energy function.

For additional hyperparameters details, please see config.yaml.

Dataset editing

Above is the instruction of editing individual image, evaluation dataset editing need iteractively editing. You should write your own scripts to generate new evaluate datasets. You may need object masks, source images, and edited text prompts accquired from previous steps(dataset preparation and text edit).