MagicArena adopts a multi-level evaluation system: the base level collects two-choice votes from mass users, the expert level invites more than 200 CG artists to give professional ratings, and the system level automatically scores through CLIP and other evaluation models. This triple data is weighted and calculated, and then updated in real-time in the three dimensions of daily/weekly/monthly list.
The ranking algorithm has the following features: 1) a 'newbie protection period' is set, and the data of the first 48 hours of newly launched models is calculated separately; 2) the Elo competitive scoring system is used, and the points gained by winning models are positively correlated with the rank of their opponents; 3) there is a 'theme-specific ranking' to track the performance differences of models in different fields. The current ranking data shows that the commercial model maintains its dominance in the realistic tasks, while the open source community model overtakes 23% in the abstract creation category.
The ranking has become an important industry wind vane, and a head cloud service provider's latest model procurement program directly references the platform's 6-month trend data as a basis for decision-making. The recently added 'Enterprise API Access' function allows organizations to directly access standardized evaluation reports.
This answer comes from the articleMagicArena: Visual Model Matchmaking Ranking Platform Launched by HeadlineThe
