Training Text-to-Image Models 3.6× Faster

Linum AI has developed a new technique called JIT (Just-In-Time) DDT (Differentiable Dynamic Tape), which accelerates the training of text-to-image models by 3.6 times. The technique uses a combination of compiler and runtime techniques to optimize the computational graph of the model, reducing the number of computations and memory accesses required. This results in a significant speedup in training time, from 86 hours to 24 hours for a given model. The technique has been implemented in the Linum AI library and is available for use by other researchers.