This week, Adobe researchers are helping shape the future of AI at the world’s premiere computer vision event, CVPR 2026, the IEEE/CVR Computer Vision and Pattern Recognition Conference. Adobe is presenting over 75 research papers, including several Oral Papers and award finalists. Many of these papers share a common direction: allowing users to have more control of the images they create.
In addition, numerous Adobe Researchers are serving as Area Chairs and several have been recognized as Outstanding Area Chairs.
Adobe CVPR papers: Intelligent object removal, sliders for more editing control, generative light movement control, and more
The papers from Adobe Research cover image editing breakthroughs, from intelligent object removal to new tools for precise AI edits (check out some highlights in our video above). They also offer advances in Adobe’s work to build trust and transparency through content attribution, making it possible to trace which images and artists have influenced a generated image.
Among the papers from Adobe Research and collaborators is Object-WIPER: Training-Free Object and Associated Effect Removal in Videos. This experimental research project includes a tool for automatically removing unwanted objects and their shadows, reflections, and other visual effects from videos without the need for frame-by-frame editing. Users simply create a rough mask around an object and enter a simple text instruction about what to remove, such as “the car and its shadow.” The research includes a pre-trained text-to-video diffusion transformer with deep understanding of how light, materials, and objects interact. The team also developed WIPER-Bench, a benchmarking tool for evaluating object removal, temporal consistency, and scene coherence. See the project page and the full paper here.
Adobe Researchers also published SliderEdit: Continuous Image Editing with Fine-Grained Instruction Control. While most instruction-based image editors offer users a single generated option that can’t be easily fine-tuned, the SliderEdit framework turns each editing instruction into a continuously controllable, adjustable slider. For example, if a user wants to edit lighting, a slider allows them to gradually adjust to the exact intensity of lighting they prefer. Or if they want to change a person’s hair, a slider can adjust the hair’s curliness or color up or down, all while preserving the subject identity. The technology provides an intuitive way to refine images, and it adapts across diverse edits and instructions, creating sliders without retraining for each attribute. In evaluations, SliderEdit achieved strong continuity and identity preservation compared to other methods. The technology will be presented as an Oral Paper at CVPR. See the project page and the full paper here.
In the paper LightMover: Generative Light Movement with Color and Intensity Controls, researchers explore a new AI framework for editing the lighting in a single image. The technology formulates light editing as a sequence-to-sequence prediction problem in visual token space, allowing users to adjust a light’s position, color, and intensity—together with the shadows, falloff, and reflections—all without re-rendering the scene. An adaptive token-pruning mechanism reduces control sequence length by 41% while maintaining editing fidelity. LightMover provides precise control of lighting with strong semantic consistency. Read the full paper here.
Additional papers from Adobe Research explore ideas across the spectrum of AI and computer vision, from the concept of a spatial scratchpad, to a novel approach for generating photorealistic physical-based rendering (PBR) materials that integrate semi-supervised learning with Latent Diffusion Models (LDMs), a two-stage reinforcement learning framework that explicitly trains anticipatory reasoning by forecasting short-horizon trajectories before execution, and the first comprehensive benchmark for automatic album organization. Read about these ideas and more here.
Reimagining image editing in the age of AI
Adobe Research’s contributions at CVPR offer a glimpse of the team’s work at the forefront of computer vision and AI, and they represent our deep collaborations with academic colleagues who are also pushing the boundaries of what’s possible. Working alongside this community, we are redefining how we’ll create, edit, and attribute images in the age of AI.
Wondering what else is happening in Adobe Research? Check out our latest news here.
