Sizes of tensors must match except in dimension 1. Expected size 8 but got size 4 for tensor number 1 in the list.

[canrly]

[Debug] Generate image using aspect ratio [Instagram (1:1)] => 1024 x 1024
Start inference...
[Debug] Prompt: instagram photo, portrait photo of a woman img, colorful, perfect face, natural skin, hard shadows, film grain,
[Debug] Neg Prompt: (asymmetry, worst quality, low quality, illustration, 3d, 2d, painting, cartoons, sketch), open mouth
10

Traceback (most recent call last):
/photomaker/model.py", line 49, in fuse_fn
stacked_id_embeds = torch.cat([prompt_embeds, id_embeds], dim=-1)
last line
Sizes of tensors must match except in dimension 1. Expected size 8 but got size 4 for tensor number 1 in the list.

[rudy2steiner]

met the same issue

[channyi]

+1

[codewritz-yuri]

+1

[hotpot-killer]

same issue

[YIYANGCAI]

I found out why prompt_embeds's dim zero is always 2 times of id_embeds. This is because the num_tokens = 2. Could anyone give a hint of this parameter's correspondence in the original paper?

I think according its original paper's stacking strategy, shouldn't the prompt_embeds be calculated out of the expansion of the embedding of token of "man" or "woman" ([1x2048]) to [id_num x 2048], and be concated with id_embeds to be [id_num, 4096] then be processed by MLPs? However, in the code's implementation, prompt_embeds are sliced from the original text_embedding with the length of (id_num * num_token).

[Sooplex]

The image is from update Photomaker v2
V2 seems different from the paper(which may be the so-called 'V1')，that makes the updated code implementation incompatible with V1.

@YIYANGCAI I found out why prompt_embeds's dim zero is always 2 times of id_embeds. This is because the num_tokens = 2.