I trained on UBC Fashion Dataset，but the outcome is pretty bad.

johannakarras / DreamPose

Official implementation of "DreamPose: Fashion Image-to-Video Synthesis via Stable Diffusion"

MIT License

951 stars 73 forks source link

I trained on UBC Fashion Dataset，but the outcome is pretty bad. #43

Open zhangtao22 opened 1 year ago

zhangtao22 commented 1 year ago

The first few frames did not change according to the pose. Overall, the clarity and coherence are poor, unable to achieve the unet 20 effect given by the author. I changed the code to 5 consecutive positions which couldn't lead to a better performance.

hello451 commented 1 year ago

Just find a good repo for similar task. They got much better results compared to DreamPose and quantitavely compare with DreamPose in their paper.

zhangtao22 commented 1 year ago

@hello451 十分感谢