Explendid work, Licon Studio!
It's really awesome that someone finally applied the VBVR dataset in LTX 2.3, turning the theory into reality.
Given that, when can we expect 240k to be released? ✨
It's really awesome that someone finally applied the VBVR dataset in LTX 2.3, turning the theory into reality.
Given that, when can we expect 240k to be released? ✨
Actually, I had nearly finished the 240k-step training, but the results didn't show the expected improvements. As a result, I've deleted the checkpoints from the past few days and rolled back to step 6,500 to adjust the strategy and restart.
Have you considered using DPO?
Have you considered using DPO?
Nope, the data prep cost for DPO is beyond my capacity. I can only refine the training plan using the provided official datasets.