One interesting challenge in training image & video models: There are many dimensions across which you can measure quality. You likely won't have one checkpoint that is the best at EVERYTHING. Insights from the nano-banana team on how they decide what to prioritize ⬇️