Second Short-Form Video Quality Assessment and Enhancement Challenge (CVPR NTIRE 2025)

CVPR NTIRE Workshop is a leading international computer‑vision symposium that includes low‑level vision tasks such as enhancement, restoration, and quality assessment. Building on the successful first short‑video quality assessment competition, the organizers are launching the second edition to establish benchmarks for short‑video quality evaluation and enhancement.

Competition Background

Short videos are ubiquitous due to low creation cost and diverse formats, but their quality varies widely because of complex processing pipelines and non‑professional creation environments. Accurately measuring and improving subjective quality is a key challenge for platforms.

Track 1: Efficient Short‑form UGC Video Quality Assessment

This track uses the KVQ dataset (4,200 short videos across nine content scenes) split 70%/20%/10% for training, validation, and testing. Participants train models on the provided training set, submit predictions via CodaLab, and are evaluated on the hidden test set. Model inference cost is limited to 120 GFlops per video, encouraging techniques such as distillation or pruning.

Track URL: https://codalab.lisn.upsaclay.fr/competitions/21335

Track 2: Diffusion‑based Image Super‑Resolution for Short‑form UGC Images

This track provides the KwaiSR dataset (1,800 paired synthetic distortions and 1,900 real low‑quality short‑video frames). Participants are encouraged to use diffusion models to improve perceptual quality while preserving content fidelity; GAN‑based methods are also accepted. Data splits follow an 80%/10%/10% ratio for training, validation, and testing.

Track URL: https://codalab.lisn.upsaclay.fr/competitions/21346

Awards

1st place: $1,000 + certificate

2nd place: $600 + certificate

3rd place: $400 + certificate

Participation Rules

The challenge is open to individuals, academia, research institutions, and companies. Teams may have up to 10 members, and each participant may join only one team. Only one final submission per team is allowed; any external data used must be disclosed.

Organizers

Co‑organized by Kuaishou Audio‑Video Technology Team and the Intelligent Media Computing Lab (IMCL). Track‑specific organizers are listed for both tracks.

Additional Information

Participants may optionally submit a workshop paper; top teams will be invited to contribute to CVPR 2025 workshop proceedings. Project homepage: https://github.com/lixinustc/KVQE-Challenge-CVPR-NTIRE2025 . CVPR NTIRE 2025 website: https://cvlai.net/ntire/2025/ .