Definitely not working for new models which is a big deal.

#10
by alx-ai - opened

When converting new 2.1 models, the controlnet model fails to run at all while the standard finetuned model converts but when I run it, it throws this error:

The config attributes {'class_embeddings_concat': False, 'cross_attention_norm': None, 'encoder_hid_dim': None, 'mid_block_only_cross_attention': None, 'resnet_out_scale_factor': 1.0, 'resnet_skip_time_act': False, 'time_embedding_act_fn': None} were passed to UNet2DConditionModel, but are not expected and will be ignored. Please verify your config.json configuration file.

It also then returns totally blown out images
image.png

Any help would be extremely appreciated because this seems to be the only way to convert these models for use with diffusers.

The only hint I can see in the logs is:
"global_step key not found in model" not sure if related

Another hint is that the unet/config.json for this 2.X model now has:
"use_linear_projection": true
while a working 2.X model I trained has it set to false but if I try and repeat that now I get:
RuntimeError: weight should have at least three dimensions

Hey @patrickvonplaten really appreciate the help, I've been pulling my hair out the past week trying to figure this out.

Trying to convert this model (sun) https://huggingface.co/ItsPerpetual/perpetual-diffusion/tree/main to diffusers
You can see my efforts converting in this repo (https://huggingface.co/alxdfy/perpetualdiffusionsun) but when I try using it with the controlnet pipeline it's totally blown out while other models work great.
My current working notebook is here https://colab.research.google.com/drive/1MJjU5thLGkcclobd8kT2fUPet_KJ9M3B?usp=sharing

Once again thank you :)

Cool - seems to work now: https://huggingface.co/alxdfy/perpetualdiffusionsun?text=Generate+something (I've also enabled an inference API for you, you just need to add a text-to-image :-) )

alx-ai changed discussion status to closed

Sign up or log in to comment