PIDM

PIDM copied to clipboard

Updated it to PyTorch2.0 and replaced Attention with PyTorch's Multihead attention

1

I've upgraded the code to be compatible with PyTorch2.0 and also replaced the attention/crossattention node to use PyTorch's build-in Multihead Attention which also of course supports flash attention out of...

Mut1nyJD

how to generate the reference_pose_0.npy from image ?

chenbolin-master

how to train and test with a 11 G gpu ?

chenbolin-master

About the implementation on multi-scale condition.

1

Thanks for sharing this great work. In the paper, you mentioned that "transfer rich multi-scale texture patterns from the source image distribution to the noise prediction" How ever, in the...

XiaoqiangZhou

hi, can you provide the model checkpoint of 512x352 ?

2

jokerlc

Hi, Current pipeline seems to start from complete noise, is it possible to have a sample code snippet where the generation starts from latents generated from another image like in...

aravind-h-v

About cond_scale

1

Have you ever conducted a CFG-deactivating ablation experiment? I'm curious as to whether deactivating CFG will significantly affect the results.

noway2beatme

is there any ablation study of TDB reference feature size?

1

 Thanks for your great work. did you do any ablation study of the TDB reference feature size, for example with 64x64, 32x32, 16x16, 8x8?

jokerlc

About the model structure

1

Incredible work! However, the code of the model structure is quite hard to read for me. Is there any chance to post a model structure figure or anything that helps...

noway2beatme

can you teach me how the"frozen_out" work? thanks!

2

frozen_out = th.cat([model_output.detach(), model_var_values], dim=1) terms["vb"] = self._vb_terms_bpd( model=lambda *args, r=frozen_out: r, x_start=x_start, x_t=x_t, t=t, clip_denoised=False, )["output"]

gouchaonijiao