Error loading Voxel Art model

[Venj-ADL]

Is there an existing issue for this?

• I have searched the existing issues

OS

Windows

GPU

cuda

VRAM

6GB

What happened?

When trying to load the Voxel Art model on a fresh install I get the following error:

Loading voxel_art-1.0 from C:\Users\venjs\invokeai\models\ldm\stable-diffusion-v1\VoxelArt_v1.ckpt
| LatentDiffusion: Running in eps-prediction mode
| DiffusionWrapper has 859.52 M params.
| Making attention of type 'vanilla' with 512 in_channels
| Working with z of shape (1, 4, 32, 32) = 4096 dimensions.
| Making attention of type 'vanilla' with 512 in_channels
** model voxel_art-1.0 could not be loaded: Error(s) in loading state_dict for LatentDiffusion:
size mismatch for embedding_manager.string_to_param_dict.: copying a param with shape torch.Size([1, 768]) from checkpoint, the shape in current model is torch.Size([8, 768]).
size mismatch for embedding_manager.initial_embeddings.: copying a param with shape torch.Size([1, 768]) from checkpoint, the shape in current model is torch.Size([8, 768]).
Traceback (most recent call last):
File "c:\users\venjs\documents\github\invokeai\ldm\invoke\model_cache.py", line 80, in get_model
requested_model, width, height, hash = self._load_model(model_name)
File "c:\users\venjs\documents\github\invokeai\ldm\invoke\model_cache.py", line 232, in _load_model
model.load_state_dict(sd, strict=False)
File "C:\Users\venjs\Documents\GitHub\invokeAI\installer_files\env\envs\invokeai\lib\site-packages\torch\nn\modules\module.py", line 1604, in load_state_dict
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for LatentDiffusion:
size mismatch for embedding_manager.string_to_param_dict.: copying a param with shape torch.Size([1, 768]) from checkpoint, the shape in current model is torch.Size([8, 768]).
size mismatch for embedding_manager.initial_embeddings.: copying a param with shape torch.Size([1, 768]) from checkpoint, the shape in current model is torch.Size([8, 768]).

Screenshots

Additional context

No response

Contact Details

venj.systems@gmail.com

[damian0815]

hmm. can you try something a bit hacky? can you go into configs\stable-diffusion\v1-inference.yaml, find personalization_config: and a few lines below that, change num_vectors_per_token: from 8 to 1? then quit and try to load the model again.

[Venj-ADL]

Yep I will do that as soon as the 1000 image batch i'm running finishes in 3 hours :P

[Venj-ADL]

Ok, yep that seems to have done the trick. Does this need to be set back to "8" for the standard models?

Scanning Model: voxel_art-1.0
Model Scanned. OK!!
Loading voxel_art-1.0 from C:\Users\venjs\invokeai\models\ldm\stable-diffusion-v1\VoxelArt_v1.ckpt
| LatentDiffusion: Running in eps-prediction mode
| DiffusionWrapper has 859.52 M params.
| Making attention of type 'vanilla' with 512 in_channels
| Working with z of shape (1, 4, 32, 32) = 4096 dimensions.
| Making attention of type 'vanilla' with 512 in_channels
| Using faster float16 precision
| Loading VAE weights from: C:\Users\venjs\invokeai\models\ldm\stable-diffusion-v1\vae-ft-mse-840000-ema-pruned.ckpt

[damian0815]

it should be fine to leave it at 1, this code is only used when there’s a textual inversion embedding being loaded or created, so it’s a bit weird that the voxel model uses it. i’ll have to investigate.

[Venj-ADL]

Excellent! thanks so much for the quick fix :)

[damian0815]

fixed by #1730