Initial implementation of IP-Adapter "begin_step_percent" and "end_step_percent" for controlling on which steps IP-Adapter is applied in the denoising loop.

invokeai/app/invocations/ip_adapter.py

@@ -34,6 +34,8 @@ class IPAdapterField(BaseModel):
     ip_adapter_model: IPAdapterModelField = Field(description="The IP-Adapter model to use.")
     image_encoder_model: CLIPVisionModelField = Field(description="The name of the CLIP image encoder model.")
     weight: float = Field(default=1.0, ge=0, description="The weight of the IP-Adapter.")
+    begin_step_percent: float = Field(default=0.0, ge=0, le=1.0)
+    end_step_percent: float = Field(default=1.0, ge=0, le=1.0)
 
 
 @invocation_output("ip_adapter_output")
@@ -54,6 +56,12 @@ class IPAdapterInvocation(BaseInvocation):
         input=Input.Direct,
     )
     weight: float = InputField(default=1.0, description="The weight of the IP-Adapter.", ui_type=UIType.Float)
+    begin_step_percent: float = InputField(
+        default=0, ge=-1, le=2, description="When the IP-Adapter is first applied (% of total steps)"
+    )
+    end_step_percent: float = InputField(
+        default=1, ge=0, le=1, description="When the IP-Adapter is last applied (% of total steps)"
+    )
 
     def invoke(self, context: InvocationContext) -> IPAdapterOutput:
         # Lookup the CLIP Vision encoder that is intended to be used with the IP-Adapter model.
@@ -80,5 +88,7 @@ class IPAdapterInvocation(BaseInvocation):
                 ip_adapter_model=self.ip_adapter_model,
                 image_encoder_model=image_encoder_model,
                 weight=self.weight,
+                begin_step_percent=self.begin_step_percent,
+                end_step_percent=self.end_step_percent,
             ),
         )

invokeai/app/invocations/latent.py

@@ -446,6 +446,8 @@ class DenoiseLatentsInvocation(BaseInvocation):
         return IPAdapterData(
             ip_adapter_model=ip_adapter_model,
             weight=ip_adapter.weight,
+            begin_step_percent=ip_adapter.begin_step_percent,
+            end_step_percent=ip_adapter.end_step_percent,
         )
 
     # original idea by https://github.com/AmericanPresidentJimmyCarter

invokeai/backend/ip_adapter/ip_adapter.py

@@ -131,8 +131,18 @@ class IPAdapter:
         self._attn_processors = attn_procs
         self._state_dict = None
 
+    # @genomancer: pushed scaling back out into its own method (like original Tencent implementation)
+    #      which makes implementing begin_step_percent and end_step_percent easier
+    #  but based on self._attn_processors (ala @Ryan) instead of original Tencent unet.attn_processors,
+    #      which should make it easier to implement multiple IPAdapters
+    def set_scale(self, scale):
+        if self._attn_processors is not None:
+            for attn_processor in self._attn_processors.values():
+                if isinstance(attn_processor, IPAttnProcessor):
+                    attn_processor.scale = scale
+
     @contextmanager
-    def apply_ip_adapter_attention(self, unet: UNet2DConditionModel, scale: int):
+    def apply_ip_adapter_attention(self, unet: UNet2DConditionModel, scale: float):
         """A context manager that patches `unet` with this IP-Adapter's attention processors while it is active.
 
         Yields:
@@ -143,10 +153,11 @@ class IPAdapter:
             # used on any UNet model (with the same dimensions).
             self._prepare_attention_processors(unet)
 
-        # Set scale.
-        for attn_processor in self._attn_processors.values():
-            if isinstance(attn_processor, IPAttnProcessor):
-                attn_processor.scale = scale
+        # Set scale
+        self.set_scale(scale)
+        # for attn_processor in self._attn_processors.values():
+        #     if isinstance(attn_processor, IPAttnProcessor):
+        #         attn_processor.scale = scale
 
         orig_attn_processors = unet.attn_processors
 

invokeai/backend/stable_diffusion/diffusers_pipeline.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 from contextlib import nullcontext
 from dataclasses import dataclass
+import math
 from typing import Any, Callable, List, Optional, Union
 
 import einops
@@ -168,6 +169,8 @@ class IPAdapterData:
     # TODO: change to polymorphic so can do different weights per step (once implemented...)
     # weight: Union[float, List[float]] = Field(default=1.0)
     weight: float = Field(default=1.0)
+    begin_step_percent: float = Field(default=0.0)
+    end_step_percent: float = Field(default=1.0)
 
 
 @dataclass
@@ -445,6 +448,7 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
                     total_step_count=len(timesteps),
                     additional_guidance=additional_guidance,
                     control_data=control_data,
+                    ip_adapter_data=ip_adapter_data,
                 )
                 latents = step_output.prev_sample
 
@@ -490,6 +494,7 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         total_step_count: int,
         additional_guidance: List[Callable] = None,
         control_data: List[ControlNetData] = None,
+        ip_adapter_data: Optional[IPAdapterData] = None,
     ):
         # invokeai_diffuser has batched timesteps, but diffusers schedulers expect a single value
         timestep = t[0]
@@ -500,6 +505,18 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         #     i.e. before or after passing it to InvokeAIDiffuserComponent
         latent_model_input = self.scheduler.scale_model_input(latents, timestep)
 
+        # handle IP-Adapter
+        if ip_adapter_data is not None:
+            first_adapter_step = math.floor(ip_adapter_data.begin_step_percent * total_step_count)
+            last_adapter_step = math.ceil(ip_adapter_data.end_step_percent * total_step_count)
+            if step_index >= first_adapter_step and step_index <= last_adapter_step:
+                # only apply IP-Adapter if current step is within the IP-Adapter's begin/end step range
+                ip_adapter_data.ip_adapter_model.set_scale(ip_adapter_data.weight)
+            else:
+                # otherwise, set IP-Adapter scale to 0, so it has no effect
+                ip_adapter_data.ip_adapter_model.set_scale(0.0)
+
+        # handle ControlNet(s)
         # default is no controlnet, so set controlnet processing output to None
         controlnet_down_block_samples, controlnet_mid_block_sample = None, None
         if control_data is not None: