DiffEdit: Diffusion-based Semantic Image Editing with Mask Guidance

[jnhwkim]

DiffEdit: Diffusion-based Semantic Image Editing with Mask Guidance

Couairon et al., arXiv 2022 (ICLR'23 under reviewing)

Image generation has recently seen tremendous advances, with diffusion models allowing to synthesize convincing images for a large variety of text prompts. In this article, we propose DiffEdit, a method to take advantage of text-conditioned diffusion models for the task of semantic image editing, where the goal is to edit an image based on a text query. Semantic image editing is an extension of image generation, with the additional constraint that the generated image should be as similar as possible to a given input image. Current editing methods based on diffusion models usually require to provide a mask, making the task much easier by treating it as a conditional inpainting task. In contrast, our main contribution is able to automatically generate a mask highlighting regions of the input image that need to be edited, by contrasting predictions of a diffusion model conditioned on different text prompts. Moreover, we rely on latent inference to preserve content in those regions of interest and show excellent synergies with mask-based diffusion. DiffEdit achieves state-of-the-art editing performance on ImageNet. In addition, we evaluate semantic image editing in more challenging settings, using images from the COCO dataset as well as text-based generated images.

🔑 Key idea:

• Semantic image editing via a text query maintaining other parts of the original image.

💪 Strength:

• Intuitive solution to a practical problem using two-query denoising and differentiation.

😵 Weakness:

• Ad-hoc masking. Binarization of the difference of conditionally estimated noises is used as the retaining mask thresholding by 0.5 (Sec. 3.2). Assume that the mask generally overshoots the region to cover the region of interest.
• The insight from Proposition 1 is shallow. The tighter bound of DiffEdit comes from using a noised version of the input image.

🤔 Confidence:

• Medium