Two Multimodal Generation Projects Accepted to CVPR 2026
Practical advances in multimodal generation
Two 360 AI Research projects have been accepted to CVPR 2026: RefTON for virtual try-on and NAMI for efficient high-resolution generation. CVPR 2026 received 16,092 submissions and accepted 4,090 papers, an acceptance rate of about 25 percent.
RefTON: reference-based virtual try-on
RefTON uses an on-person reference image to transfer clothing while preserving identity and garment details. Its two-stage design avoids several complex preprocessing requirements found in earlier pipelines and improves robustness for real-world inputs.
NAMI: faster high-resolution generation
NAMI introduces a progressive bridge for Rectified Flow models. It reduces the cost of high-resolution synthesis while retaining image quality, cutting inference time by 64 percent in the reported 1024 × 1024 setting.
The two projects share an engineering goal: make high-quality multimodal generation easier to control and more practical to deploy.