Slow convergence on training on toy dataset that you provided

Nobody has claimed this yet.

First of all, thanks for your amazing work! @lllyasviel

However, I met the problem that as I followed your steps to retrain the model to fill the circle with your toy dataset, the model converges really slow.

Below is the sampling result of 3900 steps with batch size of 4 (all the parameters remain unchanged as your tutorial_train.py).

G.T. [reconstruction_gs-003900_e-000000_b-003900]:


Output [samples_cfg_scale_9 00_gs-003900_e-000000_b-003900]:


The model seems to be able to understand the color, but cannot understand the position of the circle.

\

While continuing the training, the sampling result of 11875 steps:

G.T. [reconstruction_gs-011875_e-000001_b-003000]:


Output [samples_cfg_scale_9 00_gs-011875_e-000001_b-003000]:


The model finally learns the position of the circle, but seems no longer understand the color anymore.

Do you have any insights on this problem, or do you have any instructions on helping me solve this? Thank you in advance!

Start with the tutorial_train.py procedure and the reported batch size of 4, then compare the sampling results at 3,900 and 11,875 steps. Check whether the slow convergence and changing color or position behavior can be reproduced with the toy dataset. Done means identifying a likely cause or documenting concrete training guidance backed by reproducible observations.

Written by the indexing model from the issue text.