output of dense_motion about thin-plate-spline-motion-model

[yoyo-nb]

out_dict['contribution_maps'] represents the contribution region of each TPS to optical flow

something like this:

out_dict['deformation'] is the optical flow
out_dict['occlusion_map'] is the Multi-resolution Occlusion Masks mentioned in the paper

out_dict['deformed_source'] is the deformed source_ image before using the inpaint module

[instant-high]

Ok. I had tried that but trying to convert it to display or write it into video file I've got error message. Something like "out_dict has no deformed source image" if I remember correctly.
If it was not that thing, how to transpose?

Maybe you can help me?
I would like to try if my face swap mod is still working then but waý faster.
Using face parsing for masking I get ~ 2.it/sec now

[yoyo-nb]

sorry i said wrong

out_dict['deformed_source'] is not the deformed source_ image before using the inpaint module, it's just the source_image deformed by each TPS.

What you need is here

[instant-high]

Yes, that is working.
So the deformed source image is not output of the dense motion script? I mean, I can not skip the whole inpaint module?
Line 91 in demo.py
out = inpainting_network(source, dense_motion)

Sorry for asking again

[yoyo-nb]

you can try this:

def deform_input(inp, deformation):
    _, h_old, w_old, _ = deformation.shape
    _, _, h, w = inp.shape
    if h_old != h or w_old != w:
        deformation = deformation.permute(0, 3, 1, 2)
        deformation = F.interpolate(deformation, size=(h, w), mode='bilinear', align_corners=True)
        deformation = deformation.permute(0, 2, 3, 1)
    return F.grid_sample(inp, deformation,align_corners=True)

# ............
# demo.py, Line 88
dense_motion = dense_motion_network(source_image=source, kp_driving=kp_norm,
                                                    kp_source=kp_source, bg_param = None, 
                                                    dropout_flag = False)
deformation = dense_motion['deformation']
deformed_source = deform_input(source, deformation)  # deformed source_ image before using the inpaint module

[instant-high]

that is working so far but giving some strange/funny results. But faster....

I think I keep my hands off from changing this part.
The results of your original demo code are very good.

Thanks a lot