维度匹配不上啊

[f-chen165]

File "/vepfs/home/fanchen/fc_test_ID/DreamRunner/MotionDirector_SR3AI/diffusers/models/attention_processor.py", line 1927, in call
hidden_states = F.scaled_dot_product_attention(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: The size of tensor a (9004) must match the size of tensor b (17104) at non-singleton dimension 3

[wz0919]

17104 = 226 x 4 + 30 x 45 x 12
9004 = 226 x 4 + 30 x 45 x 6

It looks like the number of latent frames you used for computing the mask is different from the number of latent frames in your current hidden states. (one is 6 and the other is 12) The number of text conditions is similar (both are 4 conditions)
did you use the right frame size (48 in our case) and provide the right mask with a total of 12 latent frames?

[f-chen165]

17104 = 226 x 4 + 30 x 45 x 12 9004 = 226 x 4 + 30 x 45 x 6

It looks like the number of latent frames you used for computing the mask is different from the number of latent frames in your current hidden states. (one is 6 and the other is 12) The number of text conditions is similar (both are 4 conditions) did you use the right frame size (48 in our case) and provide the right mask with a total of 12 latent frames?

十分感谢您的回复，我确实使用的是24帧的生成情况，主要是48帧时80GB的A100会out of memory