I'm trying use ip-adapter to pixart-sigma.
But it seems like not convergence. (lr=1e-5, batch_size=10*8). after 4k steps, it does not convergence, and the generated image is collapsed at any training time.
I noticed that you can get a good result when training step is 200k steps. What does your results looked like in the begin of training .
(ps: the smaller lr can convergence more ( stop at about 8k steps), but the result is still collapsed too.)
I'm trying use ip-adapter to pixart-sigma. But it seems like not convergence. (lr=1e-5, batch_size=10*8). after 4k steps, it does not convergence, and the generated image is collapsed at any training time. I noticed that you can get a good result when training step is 200k steps. What does your results looked like in the begin of training . (ps: the smaller lr can convergence more ( stop at about 8k steps), but the result is still collapsed too.)