EDICT: Exact Diffusion Inversion via Coupled Transformations

Arxiv

Blog

Code

Demo

Bram Wallace, Akash Gokul, and Nikhil Naik

Reversing the generative diffusion process allows for editing of real images, but existing methods can only compute this process approximately. We develop an algorithm to perform mathematically exact inversion without any auxiliary optimization needed and demonstrate the impressive power of this method in image editing where large-scale changes can be made to specific objects while other details are near-perfectly preserved. We will be releasing follow-up work this winter/spring.

 

Activation Regression for Continuous Domain Generalization with Applications to Crop Classification

Arxiv

Samar Khanna, Bram Wallace, Kavita Bala, Bharath Hariharan

Application of deep learning to remote sensing imagery for crop classification with domain generalization. We find that in this setting it’s better to have the features be informative about the climate variables rather than mask them as in other DG approaches! Also, there’s a lot of work with EarthEngine in this paper for dataset collection which is basically programmatic Google Earth with tons of datasets.

Ensembling Self-Supervised Feature Extractors

 

 

 

 

 

Arxiv

Bram Wallace, Devansh Arpit, Huan Wang, Ciaming Xiong

Are you back in 2021 with a ton of self-supervised models and vision-language models just starting to look promising? Do you want even better features from your plethora of self-supervised models? Good news! This paper has a solution to your era-specific problem (the technique is actually pretty cool, auto-decoders are fun!)

Can We Characterize Tasks Without Labels or Features?

CVPR 2021

Bram Wallace*, Ziyang Wu* and Bharath Hariharan

We extend task characterization (see Task2Vec) to operate without task labels and pretrained feature extractors respectively.

 

Extending and Analyzing Self-Supervised Learning Across Domains

ECCV 2020

Bram Wallace and Bharath Hariharan

In this work we experiment with a variety of self-supervised learning techniques on a broad set of domains to better understand these methods outside of the CIFAR/ImageNet setting where the abundance of labels partially obviates the use of self-supervision.

 

 

 

 

 

Few-Shot Generalization for Single-Image 3D Reconstruction via Priors

ICCV 2019

Bram Wallace & Bharath Hariharan

My initial work with my advisor, Dr. Hariharan. In this paper, we present a single-view 3D reconstruction network that can leverage shape information for novel categories without retraining.

 

Förster resonance energy transfer: Role of diffusion of fluorophore orientation and separation in observed shifts of FRET efficiency

PLOS One

Bram Wallace & Paul Atzberger

My post-undergraduate research, completed partially at UC Santa Barbara. This work uses Monte Carlo simulations to investigate the effect of solution viscosity on FRET transfer efficiency.