Questions about Training loss curve

[bollossom]

The training loss for the generation task seems to be converging very slowly. After 80,000 steps, the loss is still around 1.4. To help me verify if my training is on the right track, would it be possible for you to share the loss curve from your own experiments about model SphereAR-B?

[hasaki321]

Hello, have you recently reproduced the relevant experiments in the paper?

[guolinke]

this is expected, the convergence speed of diffusion loss is usually slow.

Refer to my loss curve.

[bollossom]

Has the SphereAR-B model’s loss converged to around 1 by the end of training? If not, can the model still generate images properly? Also, would it be possible for you to share the loss curves or final loss values of models with other scales?

Here is my training loss with more steps (up to 700k)

[guolinke]

the final losses:

model	loss
base	~1.36
large	~1.32
huge	~1.28

[bollossom]

Thanks for your quick reply.