[WIP] Add sdxl lora support

invokeai/app/invocations/compel.py

@@ -173,7 +173,7 @@ def _lora_loader():
 
 
 class SDXLPromptInvocationBase:
-    def run_clip_raw(self, context, clip_field, prompt, get_pooled):
+    def run_clip_raw(self, context, clip_field, prompt, get_pooled, lora_prefix):
         tokenizer_info = context.services.model_manager.get_model(
             **clip_field.tokenizer.dict(),
             context=context,
@@ -210,8 +210,8 @@ def _lora_loader():
                 # print(traceback.format_exc())
                 print(f'Warn: trigger: "{trigger}" not found')
 
-        with ModelPatcher.apply_lora_text_encoder(
-            text_encoder_info.context.model, _lora_loader()
+        with ModelPatcher.apply_lora(
+            text_encoder_info.context.model, _lora_loader(), lora_prefix
         ), ModelPatcher.apply_ti(tokenizer_info.context.model, text_encoder_info.context.model, ti_list) as (
             tokenizer,
             ti_manager,
@@ -247,7 +247,7 @@ def _lora_loader():
 
         return c, c_pooled, None
 
-    def run_clip_compel(self, context, clip_field, prompt, get_pooled):
+    def run_clip_compel(self, context, clip_field, prompt, get_pooled, lora_prefix):
         tokenizer_info = context.services.model_manager.get_model(
             **clip_field.tokenizer.dict(),
             context=context,
@@ -284,8 +284,8 @@ def _lora_loader():
                 # print(traceback.format_exc())
                 print(f'Warn: trigger: "{trigger}" not found')
 
-        with ModelPatcher.apply_lora_text_encoder(
-            text_encoder_info.context.model, _lora_loader()
+        with ModelPatcher.apply_lora(
+            text_encoder_info.context.model, _lora_loader(), lora_prefix
         ), ModelPatcher.apply_ti(tokenizer_info.context.model, text_encoder_info.context.model, ti_list) as (
             tokenizer,
             ti_manager,
@@ -357,11 +357,11 @@ class Config(InvocationConfig):
 
     @torch.no_grad()
     def invoke(self, context: InvocationContext) -> CompelOutput:
-        c1, c1_pooled, ec1 = self.run_clip_compel(context, self.clip, self.prompt, False)
+        c1, c1_pooled, ec1 = self.run_clip_compel(context, self.clip, self.prompt, False, "lora_te1_")
         if self.style.strip() == "":
-            c2, c2_pooled, ec2 = self.run_clip_compel(context, self.clip2, self.prompt, True)
+            c2, c2_pooled, ec2 = self.run_clip_compel(context, self.clip2, self.prompt, True, "lora_te2_")
         else:
-            c2, c2_pooled, ec2 = self.run_clip_compel(context, self.clip2, self.style, True)
+            c2, c2_pooled, ec2 = self.run_clip_compel(context, self.clip2, self.style, True, "lora_te2_")
 
         original_size = (self.original_height, self.original_width)
         crop_coords = (self.crop_top, self.crop_left)
@@ -415,7 +415,8 @@ class Config(InvocationConfig):
 
     @torch.no_grad()
     def invoke(self, context: InvocationContext) -> CompelOutput:
-        c2, c2_pooled, ec2 = self.run_clip_compel(context, self.clip2, self.style, True)
+        # TODO: if there will appear lora for refiner - write proper prefix
+        c2, c2_pooled, ec2 = self.run_clip_compel(context, self.clip2, self.style, True, "<NONE>")
 
         original_size = (self.original_height, self.original_width)
         crop_coords = (self.crop_top, self.crop_left)
@@ -467,11 +468,11 @@ class Config(InvocationConfig):
 
     @torch.no_grad()
     def invoke(self, context: InvocationContext) -> CompelOutput:
-        c1, c1_pooled, ec1 = self.run_clip_raw(context, self.clip, self.prompt, False)
+        c1, c1_pooled, ec1 = self.run_clip_raw(context, self.clip, self.prompt, False, "lora_te1_")
         if self.style.strip() == "":
-            c2, c2_pooled, ec2 = self.run_clip_raw(context, self.clip2, self.prompt, True)
+            c2, c2_pooled, ec2 = self.run_clip_raw(context, self.clip2, self.prompt, True, "lora_te2_")
         else:
-            c2, c2_pooled, ec2 = self.run_clip_raw(context, self.clip2, self.style, True)
+            c2, c2_pooled, ec2 = self.run_clip_raw(context, self.clip2, self.style, True, "lora_te2_")
 
         original_size = (self.original_height, self.original_width)
         crop_coords = (self.crop_top, self.crop_left)
@@ -525,7 +526,8 @@ class Config(InvocationConfig):
 
     @torch.no_grad()
     def invoke(self, context: InvocationContext) -> CompelOutput:
-        c2, c2_pooled, ec2 = self.run_clip_raw(context, self.clip2, self.style, True)
+        # TODO: if there will appear lora for refiner - write proper prefix
+        c2, c2_pooled, ec2 = self.run_clip_raw(context, self.clip2, self.style, True, "<NONE>")
 
         original_size = (self.original_height, self.original_width)
         crop_coords = (self.crop_top, self.crop_left)

invokeai/app/invocations/latent.py

@@ -14,7 +14,7 @@
 from invokeai.app.util.step_callback import stable_diffusion_step_callback
 from invokeai.backend.model_management.models import ModelType, SilenceWarnings
 
-from ...backend.model_management.lora import ModelPatcher
+from ...backend.model_management import ModelPatcher
 from ...backend.stable_diffusion import PipelineIntermediateState
 from ...backend.stable_diffusion.diffusers_pipeline import (
     ConditioningData,

invokeai/app/invocations/model.py

@@ -262,6 +262,103 @@ def invoke(self, context: InvocationContext) -> LoraLoaderOutput:
         return output
 
 
+class SDXLLoraLoaderOutput(BaseInvocationOutput):
+    """Model loader output"""
+
+    # fmt: off
+    type: Literal["sdxl_lora_loader_output"] = "sdxl_lora_loader_output"
+
+    unet: Optional[UNetField] = Field(default=None, description="UNet submodel")
+    clip: Optional[ClipField] = Field(default=None, description="Tokenizer and text_encoder submodels")
+    clip2: Optional[ClipField] = Field(default=None, description="Tokenizer2 and text_encoder2 submodels")
+    # fmt: on
+
+
+class SDXLLoraLoaderInvocation(BaseInvocation):
+    """Apply selected lora to unet and text_encoder."""
+
+    type: Literal["sdxl_lora_loader"] = "sdxl_lora_loader"
+
+    lora: Union[LoRAModelField, None] = Field(default=None, description="Lora model name")
+    weight: float = Field(default=0.75, description="With what weight to apply lora")
+
+    unet: Optional[UNetField] = Field(description="UNet model for applying lora")
+    clip: Optional[ClipField] = Field(description="Clip model for applying lora")
+    clip2: Optional[ClipField] = Field(description="Clip2 model for applying lora")
+
+    class Config(InvocationConfig):
+        schema_extra = {
+            "ui": {
+                "title": "SDXL Lora Loader",
+                "tags": ["lora", "loader"],
+                "type_hints": {"lora": "lora_model"},
+            },
+        }
+
+    def invoke(self, context: InvocationContext) -> SDXLLoraLoaderOutput:
+        if self.lora is None:
+            raise Exception("No LoRA provided")
+
+        base_model = self.lora.base_model
+        lora_name = self.lora.model_name
+
+        if not context.services.model_manager.model_exists(
+            base_model=base_model,
+            model_name=lora_name,
+            model_type=ModelType.Lora,
+        ):
+            raise Exception(f"Unknown lora name: {lora_name}!")
+
+        if self.unet is not None and any(lora.model_name == lora_name for lora in self.unet.loras):
+            raise Exception(f'Lora "{lora_name}" already applied to unet')
+
+        if self.clip is not None and any(lora.model_name == lora_name for lora in self.clip.loras):
+            raise Exception(f'Lora "{lora_name}" already applied to clip')
+
+        if self.clip2 is not None and any(lora.model_name == lora_name for lora in self.clip2.loras):
+            raise Exception(f'Lora "{lora_name}" already applied to clip2')
+
+        output = SDXLLoraLoaderOutput()
+
+        if self.unet is not None:
+            output.unet = copy.deepcopy(self.unet)
+            output.unet.loras.append(
+                LoraInfo(
+                    base_model=base_model,
+                    model_name=lora_name,
+                    model_type=ModelType.Lora,
+                    submodel=None,
+                    weight=self.weight,
+                )
+            )
+
+        if self.clip is not None:
+            output.clip = copy.deepcopy(self.clip)
+            output.clip.loras.append(
+                LoraInfo(
+                    base_model=base_model,
+                    model_name=lora_name,
+                    model_type=ModelType.Lora,
+                    submodel=None,
+                    weight=self.weight,
+                )
+            )
+
+        if self.clip2 is not None:
+            output.clip2 = copy.deepcopy(self.clip2)
+            output.clip2.loras.append(
+                LoraInfo(
+                    base_model=base_model,
+                    model_name=lora_name,
+                    model_type=ModelType.Lora,
+                    submodel=None,
+                    weight=self.weight,
+                )
+            )
+
+        return output
+
+
 class VAEModelField(BaseModel):
     """Vae model field"""
 

invokeai/app/invocations/sdxl.py

@@ -5,7 +5,7 @@
 
 from pydantic import Field, validator
 
-from ...backend.model_management import ModelType, SubModelType
+from ...backend.model_management import ModelType, SubModelType, ModelPatcher
 from invokeai.app.util.step_callback import stable_diffusion_xl_step_callback
 from .baseinvocation import BaseInvocation, BaseInvocationOutput, InvocationConfig, InvocationContext
 
@@ -293,10 +293,20 @@ def invoke(self, context: InvocationContext) -> LatentsOutput:
 
         num_inference_steps = self.steps
 
+        def _lora_loader():
+            for lora in self.unet.loras:
+                lora_info = context.services.model_manager.get_model(
+                    **lora.dict(exclude={"weight"}),
+                    context=context,
+                )
+                yield (lora_info.context.model, lora.weight)
+                del lora_info
+            return
+
         unet_info = context.services.model_manager.get_model(**self.unet.unet.dict(), context=context)
         do_classifier_free_guidance = True
         cross_attention_kwargs = None
-        with unet_info as unet:
+        with ModelPatcher.apply_lora_unet(unet_info.context.model, _lora_loader()), unet_info as unet:
             scheduler.set_timesteps(num_inference_steps, device=unet.device)
             timesteps = scheduler.timesteps
 
@@ -543,9 +553,19 @@ def invoke(self, context: InvocationContext) -> LatentsOutput:
             context=context,
         )
 
+        def _lora_loader():
+            for lora in self.unet.loras:
+                lora_info = context.services.model_manager.get_model(
+                    **lora.dict(exclude={"weight"}),
+                    context=context,
+                )
+                yield (lora_info.context.model, lora.weight)
+                del lora_info
+            return
+
         do_classifier_free_guidance = True
         cross_attention_kwargs = None
-        with unet_info as unet:
+        with ModelPatcher.apply_lora_unet(unet_info.context.model, _lora_loader()), unet_info as unet:
             # apply denoising_start
             num_inference_steps = self.steps
             scheduler.set_timesteps(num_inference_steps, device=unet.device)

invokeai/backend/model_management/__init__.py

@@ -13,3 +13,4 @@
     DuplicateModelException,
 )
 from .model_merge import ModelMerger, MergeInterpolationMethod
+from .lora import ModelPatcher