Open rfeinman opened 2 months ago
Hi, this is a historical issue. The two correspond to different ckpts (aligned-vae and w/o aligned). The ckpt in hf is aligned-vae with True zero_uncond_embeds
and we will unified it in the future. Set zero_uncond_embeds
maybe better in my practice.
Moreover, we will release a dit-based version soon, which will greatly improve performance. Please stay tuned. (The input is the 1st col, followed with 4 rendered normals with predicted mesh)
Hi @wyysf-98
FYI the huggingface repo contains two separate configs:
According to these configs, zero_uncond_embeds=True
in both variants. Is this a mistake?
Hi,
What is the correct value for the
zero_uncond_embeds
hyperparameter of your released diffusion model checkpoint clip-mvrgb-modln-l256-e64-ne8-nd16-nl6? In the config file included with this github repo, the value is set toFalse
. However in the huggingface repo, it is set toTrue
. Which is correct?config file in this github repo:
https://github.com/wyysf-98/CraftsMan/blob/9c836e201a122b1f89fc51b3152de870b2ef6ba6/configs/image-to-shape-diffusion/clip-mvrgb-modln-l256-e64-ne8-nd16-nl6.yaml#L63
config file in huggingface repo:
https://huggingface.co/wyysf/CraftsMan/blob/main/image-to-shape-diffusion/clip-mvrgb-modln-l256-e64-ne8-nd16-nl6/config.yaml#L68