Bagheera
Bagheera
not stale, just waiting on some pytorch improvements
the SDXL training scripts are broken this way as well @patrickvonplaten @sayakpaul which means the models that are trained with diffusers are likely all broken.
the prompt embed code also pools incorrectly. as tested by @AmericanPresidentJimmyCarter it looks like the correct method would be to pool the final layer while sampling hidden states from the...
still not stale :D
@sayakpaul i opened the pull request for this. but the code in question only runs when prompt embeds are None. do we want to mix and match provided pooled embeds...
good candidate for #7761
For me, brighter images make it more noticeable. # DDIM  # Euler (Patched) 