Zimage-Turbo: Simple comparison: DoRA vs LoHA.
A viral benchmark reveals which fine-tuning method wins for speed and efficiency.
Deep Dive
A community benchmark comparing DoRA and LoHA fine-tuning methods for the Zimage-Base model shows DoRA training 30% faster, completing 100 epochs in 1h3m versus LoHA's 1h22m on an RTX 4060 Ti. The aggressive training prioritized speed over perfect quality. DoRA used a batch size of 11 and a learning rate of 0.00006. This data provides crucial performance insights for researchers struggling to train the challenging Zimage-Base model efficiently.
Why It Matters
Faster, more efficient fine-tuning directly lowers the cost and barrier to creating custom AI models.