On the CVTG-2K benchmark, which measures accuracy in inserting textual content throughout a number of picture places, GLM-Picture achieved a Phrase Accuracy rating of 0.9116, rating first amongst open-source fashions. The mannequin additionally led the LongText-Bench take a look at for rendering prolonged textual content passages, scoring 0.952 for English and 0.979 for Chinese language throughout eight eventualities together with indicators, posters, and dialog bins.
The mannequin natively helps a number of resolutions from 1024×1024 to 2048×2048 pixels with out requiring retraining, the report added.
{Hardware} optimization technique
Coaching GLM-Picture on Ascend {hardware} required Zhipu to develop customized optimization strategies for Huawei’s chip structure. The corporate constructed a coaching suite that implements dynamic graph multi-level pipelined deployment, enabling totally different phases of the coaching course of to run concurrently and lowering bottlenecks.
