Diffusion Handles:
Enabling 3D Edits for Diffusion Models by Lifting Activations to 3D

Karran Pandey1, Paul Guerrero2, Matheus Gadelha2, Yannick Hold-Geoffroy2, Karan Singh1, Niloy Mitra2,3
1University of Toronto, 2Adobe Research, 3University College London
CVPR 2024 - Highlight
Arxiv Video Code
3D-aware object edits for diffusion images. 3D-aware object edits for diffusion images.

Diffusion Handles enables 3D-aware object edits on images generated by diffusion models. The 3D edit captures complex light and shading effects, as well as a changes in perspective, which is hard to achieve by enforcing object pixel intensity invariance used in image-space identity control.

Abstract

Diffusion Handles is a novel approach to enabling 3D object edits on diffusion images. We accomplish these edits using existing pre-trained diffusion models, and 2D image depth estimation, without any fine-tuning or 3D object retrieval. The edited results remain plausible, photo-real, and preserve object identity.

Diffusion Handles addresses a critically missing facet of generative image based creative design, and significantly advance the state-of-the-art in generative image editing. Our key insight is to lift diffusion activations for an object to 3D using a proxy depth, 3D-transform the depth and associated activations, and project them back to image space. The diffusion process applied to the manipulated activations with identity control, produces plausible edited images showing complex 3D occlusion and lighting effects.

We evaluate Diffusion Handles: quantitatively, on a large synthetic data benchmark; and qualitatively by a user study, showing our output to be more plausible, and better than prior art at both, 3D editing and identity control.

3D-Aware Editing Results


Comparison to the State-of-the-Art

Input Target Edit Ours Obj. Stitch Zero123 3DIT

(Comparison to the state-of-the-art on a much larger set of images are available here for images with estimated depth & synthetic depth.)


ObjectStitch is a generative object compositing method which we adapt for object-level editing on scenes.
Zero-1-2-3 is a diffusion-based novel-view synthesis method which we use in conjunction with inpainting to perform object-level 3D-aware image editing.
Object 3DIT is a language-guided model for 3D-aware image editing.



Frame-by-Frame Interpolation

Interpolate start reference image.

Start Frame

Loading...
Interpolation end reference image.

End Frame


Interpolate start reference image.

Start Frame

Loading...
Interpolation end reference image.

End Frame

Method Overview

Method overview. 3D Edit.
  • The input image is first reconstructed with a depth-to-image diffusion model. Intermediate activations are recorded.
  • Depth is estimated using a monocular depth estimator and the intermediate activations from the last step are lifted to the 3D depth surface.
  • A user-supplied 3D transform is applied to the depth surface and the lifted activations.
  • The 3D-transformed depth and activations are used to guide the diffusion model to generate an edited image.