600M-512px for comfyui not working

[Pancat009]

I download the Sana_600M_512px model in /checkpoint,
and download the whole Sana_600M_512px_diffusers in /models/diffusers
and copy the vae from /Sana_600M_512px_diffusers/vae to models/vae
but it repots error

Error(s) in loading state_dict for SanaMS:
size mismatch for pos_embed: copying a param with shape torch.Size([1, 256, 1152]) from checkpoint, the shape in current model is torch.Size([1, 1024, 1152]).

[lawrence-cj]

This bug should be fixed if you update the source code here: https://github.com/Efficient-Large-Model/ComfyUI_ExtraModels

Refer: #123

[Pancat009]

Thank you.
The image can be generated now, but not sure what mistake i've made, the image looks like this

[lawrence-cj]

Try to specify the precision when you load the weight. Like FP16 or BF16

Refer to: https://github.com/NVlabs/Sana/blob/main/asset/docs/model_zoo.md#-1-we-provide-all-the-links-of-sana-pth-and-diffusers-safetensor-below

[Pancat009]

I've tried to change resolution, gemma loader, vae type, vae dtype, arrange combination with each other, but still get the blur image.
changing the paramters in Ksampler has no effect neither