Two Undergraduates Earn Best Student Paper Nomination at CVPR 2026
At CVPR 2026, two undergraduate researchers from Guangdong University of Technology secured a Best Student Paper nomination for their ChordEdit work, which introduces a low‑energy optimal‑transport framework for one‑step image editing and outperforms existing methods in speed, memory usage, and user preference.
CVPR 2026 was held in Denver, receiving 16,092 submissions and accepting 4,090 papers (acceptance rate 25.42%). The program committee announced all award winners, including best paper, best student paper nominations, and test‑of‑time awards.
Best Student Paper Nomination: ChordEdit
The nominated paper, ChordEdit: One‑Step Low‑Energy Transport for Image Editing , lists Liangsi Lu (first author) and Yang Shi as corresponding authors, both undergraduate students at Guangdong University of Technology. Their research focuses on representation learning and visual generation.
Current one‑step text‑to‑image models (e.g., SD‑Turbo, SwiftBrush) achieve high generation speed but fail at image editing because they rely on multi‑step, training‑free methods (FlowEdit, Direct Inversion) that average over many inference steps to stabilize trajectories. Compressing these methods into a single step leads to noisy results, object distortion, and loss of consistency between edited and non‑edited regions—a structural mathematical flaw rather than a tunable hyper‑parameter issue.
ChordEdit reframes image editing as an optimal‑transport problem between the source distribution (original image + prompt) and the target distribution (image + edited prompt). It builds on the dynamic optimal‑transport formulation introduced by Benamou and Brenier (2000), which treats transport as a fluid‑mechanics flow. Instead of directly applying the raw drift‑field difference, ChordEdit computes a weighted average of drift fields at two adjacent time points, yielding a low‑energy Chord Control Field . Jensen’s inequality guarantees energy contraction, reducing variance and discretization error, making a single‑step integration stable.
The method requires no training, no inversion, no extra mask network, and no modifications to the underlying diffusion model, making it model‑agnostic (compatible with SD‑Turbo, SwiftBrush‑v2, etc.). Experiments ran on a 2018 NVIDIA Titan 24 GB GPU, using only 7 GB of VRAM. Compared to FlowEdit, ChordEdit is 19× faster; compared to Direct Inversion, it is 208× faster. A user study showed 42.5 % of participants preferred ChordEdit for semantic accuracy and 48.3 % for background preservation.
Other CVPR 2026 Highlights
Best Paper went to DeepMind’s D4RT: Efficiently Reconstructing Dynamic Scenes One D4RT at a Time , which unifies depth estimation, camera pose, 3D point tracking, and 4D point‑cloud inference in a lightweight, training‑free framework, surpassing prior SOTA on multiple 4D reconstruction benchmarks.
Best Student Paper (other than ChordEdit) was awarded to TRELLIS.2: Native and Compact Structured Latents for 3D Generation from Tsinghua University and Microsoft Research, introducing the O‑Voxel “all‑purpose voxel” that sparsely encodes geometry and appearance, coupled with a Sparse Compression VAE and a 4‑billion‑parameter flow‑matching generator.
The Longuet‑Higgins Test‑of‑Time Award recognized two seminal CVPR 2016 papers—ResNet and YOLO—highlighting their lasting impact on deep learning and real‑time object detection.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
Machine Learning Algorithms & Natural Language Processing
Focused on frontier AI technologies, empowering AI researchers' progress.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
