No description provided.
darkstorm2150 changed pull request status to open

@patrickvonplaten :

The conversion scripts doesn't seem to work on unpruned ckpt files and produces and empty pull request... as demonstrated here.

@johnslegers I'm adding pruned fp16 ckpt files, maybe this will fix it

Hey sorry to answer so late here!

It seems like the format of the checkpoint is slightly different - when converting this manually I'm getting a:

RuntimeError: Error(s) in loading state_dict for UNet2DConditionModel:
        Missing key(s) in state_dict: "up_blocks.0.upsamplers.0.conv.weight", "up_blocks.0.upsamplers.0.conv.bias",
"up_blocks.1.upsamplers.0.conv.weight", "up_blocks.1.upsamplers.0.conv.bias", "up_blocks.2.upsamplers.0.conv.weight",
"up_blocks.2.upsamplers.0.conv.bias".
        Unexpected key(s) in state_dict: "up_blocks.0.attentions.2.conv.bias", "up_blocks.0.attentions.2.conv.weight".

error message. @darkstorm2150 is this model also a v1 version? How is it different from: https://huggingface.co/darkstorm2150/Protogen_v5.3_Official_Release exactly?
Did the architecture slightly change? 😊

I found the problem; there was a subtle bug in the conversion script! All good now - sorry for the confusion.

The weights uploaded to this PR should work just fine.

@johnslegers , yes we recently added a couple of bug fixes that are not updated in the conversion space, such as: https://github.com/huggingface/diffusers/pull/1959 .

I'm trying to make a new conversion space here: https://huggingface.co/spaces/diffusers/sd-to-diffusers that should work with all checkpoints.

Tested the weights of this PR and they should work as expected!

This conversion script should work with all checkpoints: https://huggingface.co/spaces/diffusers/sd-to-diffusers

This conversion script should work with all checkpoints: https://huggingface.co/spaces/diffusers/sd-to-diffusers

Just tried the diffusers update, wouldn't convert this correctly. I get jibberish
(tried the repo one)

@patrickvonplaten no worries, x5.8 is rebuilt using dreamlikePhotoRealV2.ckpt as a core, adding small amounts during merge checkpoints. I'm looking into (researching) on building the diffusers weights from what ever I can find online, don't want to burden your time with 50+ models

darkstorm2150 changed pull request status to merged

No burden at all ;-) Super cool that you're open-sourcing all those weights for the community!
I think this space should work well going forward: https://huggingface.co/spaces/diffusers/sd-to-diffusers .

This comment has been hidden

@patrickvonplaten :

I just tried https://huggingface.co/spaces/diffusers/sd-to-diffusers and I can confirm it works on larger CKPT files the original script didn't work for

* excited *

Sign up or log in to comment