Liangsi Lu

Vision Generation

[Liangsi Lu @LiangsiLu]

🎉 ChordEdit has been accepted to #CVPR2026

It's a training-free & inversion-free text-guided image editing framework. We accelerate editing by 270x vs baselines, taking just 0.2s! ⚡️ Preserves backgrounds & layouts without masks. 👇 (1/n)

[Liangsi Lu @LiangsiLu]

🎊 Our paper has been selected for an Oral presentation at #CVPR2026

[Liangsi Lu @LiangsiLu]

🎉 ChordEdit has been accepted to #CVPR2026

It's a training-free & inversion-free text-guided image editing framework. We accelerate editing by 270x vs baselines, taking just 0.2s! ⚡️ Preserves backgrounds & layouts without masks. 👇 (1/n)

[Liangsi Lu @LiangsiLu]

Check out more technical details and results on our project page! 👇

🌐 Page: https://t.co/OhKyHwBl2t 
📄 Preprint: https://t.co/89olhred12 
💻 Code: https://t.co/8zaE9inh7b

[Liangsi Lu @LiangsiLu]

🎉 ChordEdit has been accepted to #CVPR2026

It's a training-free & inversion-free text-guided image editing framework. We accelerate editing by 270x vs baselines, taking just 0.2s! ⚡️ Preserves backgrounds & layouts without masks. 👇 (1/n)

[Liangsi Lu @LiangsiLu]

When approaching the single-step limit, the energy of naive vector difference shows a catastrophic surge, causing a collapse in fidelity (PSNR).

Our Chord Control Field firmly locks the energy at a low level, maintaining extremely high visual fidelity. (6/n)