SDXL & SDXL-refiner models convert correctly

invokeai/backend/model_management/convert_ckpt_to_diffusers.py

@@ -1121,7 +1121,7 @@ def download_from_original_stable_diffusion_ckpt(
     prediction_type: str = None,
     model_type: str = None,
     extract_ema: bool = False,
-    precision: torch.dtype = torch.float16,
+    precision: torch.dtype = torch.float32,
     scheduler_type: str = "pndm",
     num_in_channels: Optional[int] = None,
     upcast_attention: Optional[bool] = None,
@@ -1250,7 +1250,7 @@ def download_from_original_stable_diffusion_ckpt(
     while "state_dict" in checkpoint:
         checkpoint = checkpoint["state_dict"]
         
-    print(f'DEBUG: model_type = {model_type}; original_config_file = {original_config_file}')
+    logger.debug(f'model_type = {model_type}; original_config_file = {original_config_file}')
 
     if original_config_file is None:
         key_name_v2_1 = "model.diffusion_model.input_blocks.2.1.transformer_blocks.0.attn2.to_k.weight"
@@ -1624,7 +1624,7 @@ def download_controlnet_from_original_ckpt(
     original_config_file: str,
     image_size: int = 512,
     extract_ema: bool = False,
-    precision: torch.dtype = torch.float16,
+    precision: torch.dtype = torch.float32,
     num_in_channels: Optional[int] = None,
     upcast_attention: Optional[bool] = None,
     device: str = None,
@@ -1702,7 +1702,7 @@ def convert_ldm_vae_to_diffusers(checkpoint, vae_config: DictConfig, image_size:
 def convert_ckpt_to_diffusers(
         checkpoint_path: Union[str, Path],
         dump_path: Union[str, Path],
-        no_safetensors: bool = False,
+        use_safetensors: bool=True,
         **kwargs,
 ):
     """
@@ -1714,7 +1714,7 @@ def convert_ckpt_to_diffusers(
 
     pipe.save_pretrained(
         dump_path,
-        safe_serialization=is_safetensors_available() and not no_safetensors,
+        safe_serialization=use_safetensors and is_safetensors_available(),
     )
 
 def convert_controlnet_to_diffusers(

invokeai/backend/model_management/model_probe.py

@@ -253,10 +253,12 @@ class PipelineCheckpointProbe(CheckpointProbeBase):
             return BaseModelType.StableDiffusion1
         if key_name in state_dict and state_dict[key_name].shape[-1] == 1024:
             return BaseModelType.StableDiffusion2
-        # TODO: This is just a guess based on N=1
         key_name = 'model.diffusion_model.input_blocks.4.1.transformer_blocks.0.attn2.to_k.weight'
         if key_name in state_dict and state_dict[key_name].shape[-1] == 2048:
             return BaseModelType.StableDiffusionXL
+        elif key_name in state_dict and state_dict[key_name].shape[-1] == 1280:
+            return BaseModelType.StableDiffusionXLRefiner
+        else:
             raise InvalidModelException("Cannot determine base type")
 
     def get_scheduler_prediction_type(self)->SchedulerPredictionType:

invokeai/backend/model_management/models/sdxl.py

@@ -1,5 +1,6 @@
 import os
 import json
+import invokeai.backend.util.logging as logger
 from enum import Enum
 from pydantic import Field
 from typing import Literal, Optional
@@ -108,14 +109,20 @@ class StableDiffusionXLModel(DiffusersModel):
         config: ModelConfigBase,
         base_model: BaseModelType,
     ) -> str:
+        # The convert script adapted from the diffusers package uses
+        # strings for the base model type. To avoid making too many
+        # source code changes, we simply translate here
+        model_base_to_model_type = {BaseModelType.StableDiffusionXL: 'SDXL',
+                                    BaseModelType.StableDiffusionXLRefiner: 'SDXL-Refiner',
+                                    }
         if isinstance(config, cls.CheckpointConfig):
             from invokeai.backend.model_management.models.stable_diffusion import _convert_ckpt_and_cache
             return _convert_ckpt_and_cache(
                 version=base_model,
                 model_config=config,
                 output_path=output_path,
-                model_type='SDXL',
+                model_type=model_base_to_model_type[base_model],
-                no_safetensors=True, # giving errors for some reason
+                use_safetensors=False, # corrupts sdxl models for some reason
             )
         else:
             return model_path

invokeai/backend/model_management/models/stable_diffusion.py

@@ -16,9 +16,11 @@ from .base import (
     InvalidModelException,
 )
 from .sdxl import StableDiffusionXLModel
+import invokeai.backend.util.logging as logger
 from invokeai.app.services.config import InvokeAIAppConfig
 from omegaconf import OmegaConf
 
+
 class StableDiffusion1ModelFormat(str, Enum):
     Checkpoint = "checkpoint"
     Diffusers = "diffusers"
@@ -265,6 +267,9 @@ def _convert_ckpt_and_cache(
 
     # to avoid circular import errors
     from ..convert_ckpt_to_diffusers import convert_ckpt_to_diffusers
+    from ...util.devices import choose_torch_device, torch_dtype
+    
+    logger.info(f'Converting {weights} to diffusers format')
     with SilenceWarnings():        
         convert_ckpt_to_diffusers(
             weights,
@@ -275,6 +280,7 @@ def _convert_ckpt_and_cache(
             extract_ema=True,
             scan_needed=True,
             from_safetensors = weights.suffix == ".safetensors",
+            precision = torch_dtype(choose_torch_device()),
             **kwargs,
         )
     return output_path

invokeai/backend/util/logging.py

@@ -1,7 +1,7 @@
 # Copyright (c) 2023 Lincoln D. Stein and The InvokeAI Development Team
 
 """
-invokeai.util.logging
+invokeai.backend.util.logging
 
 Logging class for InvokeAI that produces console messages