Export
Eli Shechtman
· 6 previous affiliations
- Adobe Systems (United States)Since 2007
- Senior Principal Scientist · Since 2020
- Research Scientist · Since 2007
- Current
- Principal Scientist · 2014–2020
- Senior Research Scientist · 2010–2014
- University of Washington2007–2009
- Postdoc · 2007–2009
- Elbit Systems (Israel)2001–2007
- Computer Vision consultant · 2001–2007
- Weizmann Institute of Science2001–2007
- PhD student · 2003–2007
- M.Sc., Computer Science and Applied Mathematics · 2001–2003
- IDF1996–2001
- Electrical Engineer and Team Leader · 1996–2001
- Tel Aviv University1992–1996
- Undergrad student · 1992–1996
Research Overview
Eli Shechtman focuses on computer vision, computational photography, and generative modeling, with an emphasis on controllable image and video synthesis. His trajectory spans early work in space-time action recognition and image morphing to the development of advanced latent space manipulation techniques for generative adversarial networks and diffusion models. Recently, his research addresses real-time interactive video generation and fast generative distillation.
Research Trajectory
His early research addressed fundamental problems in computer vision and image processing, focusing on motion analysis and image synthesis. In Actions as Space-Time Shapes, he and his collaborators modeled human actions as three-dimensional shapes within a space-time volume, utilizing properties of the Poisson equation to extract features for action recognition and clustering. He later developed Regenerative morphing, an optimization-based approach that generates morphing sequences by achieving bidirectional similarity between the frames and the source images, eliminating the need for manual correspondence even across disparate objects.
As deep learning transformed computer vision, his focus shifted toward generative adversarial networks (GANs) and neural image manipulation. He contributed to Generative Visual Manipulation on the Natural Image Manifold, which constrained user edits to a learned manifold to maintain photorealism during real-time shape and color transformations. Expanding on neural synthesis, he explored the research direction of Structural Preservation in Generative Image Stylization through Deep Photo Style Transfer. This work introduced locally affine constraints in colorspace to prevent the painterly distortions typical of earlier style transfer methods, enabling faithful photorealistic stylization. Concurrently, he investigated conditional generative modeling in Toward Multimodal Image-to-Image Translation, proposing an invertible connection between the output and a low-dimensional latent code to prevent mode collapse and model a diverse distribution of possible outputs. This contributed significantly to the broader area of Disentangled Control in Generative Latent Spaces.
A critical component of evaluating and training generative models is measuring visual fidelity. In The Unreasonable Effectiveness of Deep Features as a Perceptual Metric, he helped introduce a dataset of human perceptual judgments to demonstrate that deep features from classification networks serve as highly effective perceptual losses. This work became foundational in the area of The Perceptual Misalignment of Image Quality Metrics.
His subsequent work extensively analyzed and manipulated the latent spaces of state-of-the-art GANs. In StyleSpace Analysis: Disentangled Controls for StyleGAN Image Generation, he identified highly disentangled channel-wise style parameters in StyleGAN2, allowing for localized control over distinct visual attributes. Building on this capability, he co-developed StyleCLIP: Text-Driven Manipulation of StyleGAN Imagery, which integrated Contrastive Language-Image Pre-training (CLIP) to enable intuitive, text-based image editing without requiring manual annotation or exhaustive latent space exploration.
Continuing to refine generative synthesis and alignment, he contributed to Interactive Sketch & Fill: Multiclass Sketch-to-Image Translation, introducing a gating mechanism that allows a single model to generate diverse object classes from sparse user sketches. He also addressed the limitations of fixed-resolution generative models in Any-resolution Training for High-resolution Image Synthesis. By sampling patches at random scales, this continuous-scale training approach advanced the research direction of Resolution Agnostic Generative Models for Arbitrary Scale Synthesis. In parallel, he explored joint learning of discriminative models and generated data in GAN-Supervised Dense Visual Alignment, utilizing a Spatial Transformer to discover dense correspondences across unaligned data. To improve real-world image composition, he developed Semi-supervised Parametric Real-world Image Harmonization, a parametric model trained on unpaired real composites to handle complex local appearance mismatches like illumination and object boundaries.
More recently, his research has focused on accelerating generative models and extending them to video. He contributed to the area of Preserving Diversity in Fast Generative Distillation with Improved Distribution Matching Distillation for Fast Image Synthesis, which refines the DMD2 framework to transform iterative diffusion models into fast, few-step generators using GAN-based supervision. In the video domain, he addressed exposure bias in autoregressive generation with Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion, mitigating the accumulation of errors during inference.
His latest work explores novel training paradigms and real-time video architectures. In Learning an Image Editing Model without Image Editing Pairs, he participated in developing NP-Edit, a framework that utilizes differentiable feedback from Vision-Language Models to train image editors without paired before-and-after data. He also co-authored MotionStream: Real-Time Video Generation with Interactive Motion Controls, which distills a teacher model into a causal student architecture to enable high-framerate, interactive video generation guided by user trajectories. Finally, in What matters for Representation Alignment: Global Information or Spatial Structure?, his research demonstrated that spatial structure and patch relationships are more critical than global semantic accuracy when selecting teacher models for diffusion training, introducing the iREPA method to accelerate generative model convergence.
Current Directions
Recent work investigates the acceleration and real-time deployment of generative models, particularly in the video domain. A major focus is overcoming the latency and error-accumulation challenges inherent in autoregressive and diffusion-based video generation. By designing causal architectures and addressing exposure bias, this research enables interactive, high-framerate video synthesis that responds instantly to user-guided motion controls. Additionally, current efforts explore novel training paradigms that bypass the need for exhaustively curated datasets. This includes leveraging differentiable feedback from Vision-Language Models for unpaired image editing and prioritizing local spatial structure over global semantics to optimize representation alignment in diffusion models. Together, these directions aim to make high-fidelity image and video synthesis faster, more controllable, and less reliant on paired supervision.
Profile updated
Loading new papers…
Other researchers whose papers cluster into the same research directions — a proxy for overlapping technical interests and shared intellectual neighborhoods.
See something wrong? Report a mistake