The provided training configs have max_steps=1,000,000 which seems like a lot. Where the models trained for that long or where they stopped early? I'm trying to replicate the reported results by training TRELLIS.2 from scratch using by own dataset of ~700k meshes. The loss curves seem to be stagnating after training for ~100k iterations but the results aren't yet close to the ones from the released models. Additional details would be very helpful here.
The provided training configs have
max_steps=1,000,000which seems like a lot. Where the models trained for that long or where they stopped early? I'm trying to replicate the reported results by training TRELLIS.2 from scratch using by own dataset of ~700k meshes. The loss curves seem to be stagnating after training for ~100k iterations but the results aren't yet close to the ones from the released models. Additional details would be very helpful here.