Inception Labs' Mercury 2 AI Beats Google's DiffusionGemma at Its Own Game

Inception Labs' Mercury 2 AI model outperformed Google's DiffusionGemma on image generation tasks this week, according to Inception Labs' technical blog post. Mercury 2 achieved a FID score of 3.1, while DiffusionGemma scored 4.5 on the same benchmark dataset. Both models utilize a parallel denoising approach, a departure from traditional word-by-word generation, to accelerate image synthesis. However, Inception Labs claims Mercury 2 maintains a higher level of semantic understanding and coherence in its generated images, a common challenge with parallel denoising methods. The company stated that Mercury 2 was trained on a dataset of 1 billion image-text pairs, with a training budget of 10,000 GPU hours. DiffusionGemma, released by Google DeepMind in February 2024, was designed to be an open-source, efficient diffusion model for research purposes. Inception Labs has not yet announced a release date for Mercury 2, but indicated it would be available through their API in Q4 2024.