Undergrad Wins CVPR Best Student Paper Nomination Using an Old NVIDIA Titan GPU

At the CVPR 2026 award ceremony, a paper called ChordEdit: One-Step Low-Energy Transport for Image Editing was accepted for an oral presentation and received an Honorable Mention for Best Student Paper. The work was led by Liangsi Lu, a third‑year undergraduate at Guangdong University of Technology, and the experiments were run on a Turing‑based NVIDIA Titan 24 GB GPU released in 2018.

ChordEdit addresses the “image collapse” problem that arises when one‑step text‑to‑image (T2I) models are forced to perform editing in a single inference step. Direct vector operations on the model’s latent field produce high‑energy, unstable trajectories, leading to severe object deformation and loss of consistency in non‑edited regions.

To solve this, the authors reformulate image editing as a transport problem between the source distribution (defined by the source prompt) and the target distribution (defined by the target prompt). Using dynamic optimal transport theory, they derive a principled low‑energy control strategy that yields smoother, lower‑variance editing fields. This strategy enables a large integration step that traverses the transport path in a single pass.

Experiments show that ChordEdit requires only two network calls (one transport step and an optional post‑process) and runs in 0.38 seconds on the NVIDIA Titan 24 GB GPU, consuming 6988 MB of VRAM—substantially less than competing methods such as SwiftEdit (≈15 GB). On the PIE‑bench dataset, ChordEdit outperforms baselines in background consistency (higher PSNR) and semantic alignment (higher CLIP score). The method is model‑agnostic and works with popular one‑step generators including SD‑Turbo, InstaFlow, and SwiftBrush‑v2.

The paper’s technical and theoretical contributions demonstrate how optimal‑transport tools can solve practical engineering challenges in real‑time image editing, and the authors anticipate further research building on this framework.