fine-tuning sana with lance dataset

[AyushExel]

Sana is a hybrid DiT architecture with <3B params that turns one image and a camera trajectory into 720p, minute-long, controllable video on a single GPU.

A few open questions

• which dataset should be used to tune
• What skill could be teach the model ( example - train on generating specific scenes better)
• What features could be added offline to improve GPU util when training