About text2img generation finetuning

[o0t1ng0o]

Hi @wangwei1237 @haooooooqi @helio9cn @TsuTikgiau @Luodian

Thank you for releasing such an impactful work.

I try to fine-tune my text2img data on Bagel.
But I found that the generated image suffering from strange color and its shape seems correct.
I have checked the input data which is valid. I also checked the inference code that is correct.

What could be causing this result?

Looking forward to your reply.

[o0t1ng0o]

In addition, the MSE loss cannot converge.