Update IP-Adapter model to enable running multiple IP-Adapters at once. (Not tested yet.)

invokeai/backend/ip_adapter/attention_processor.py

@@ -8,6 +8,8 @@ import torch.nn as nn
 import torch.nn.functional as F
 from diffusers.models.attention_processor import AttnProcessor2_0 as DiffusersAttnProcessor2_0
 
+from invokeai.backend.ip_adapter.ip_attention_weights import IPAttentionProcessorWeights
+
 
 # Create a version of AttnProcessor2_0 that is a sub-class of nn.Module. This is required for IP-Adapter state_dict
 # loading.
@@ -45,18 +47,13 @@ class IPAttnProcessor2_0(torch.nn.Module):
             the weight scale of image prompt.
     """
 
-    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0):
+    def __init__(self, weights: list[IPAttentionProcessorWeights]):
         super().__init__()
 
         if not hasattr(F, "scaled_dot_product_attention"):
             raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
 
-        self.hidden_size = hidden_size
-        self.cross_attention_dim = cross_attention_dim
-        self.scale = scale
-
-        self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
-        self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.weights = weights
 
     def __call__(
         self,
@@ -67,16 +64,6 @@ class IPAttnProcessor2_0(torch.nn.Module):
         temb=None,
         ip_adapter_image_prompt_embeds=None,
     ):
-        if encoder_hidden_states is not None:
-            # If encoder_hidden_states is not None, then we are doing cross-attention, not self-attention. In this case,
-            # we will apply IP-Adapter conditioning. We validate the inputs for IP-Adapter conditioning here.
-            assert ip_adapter_image_prompt_embeds is not None
-            # The batch dimensions should match.
-            assert ip_adapter_image_prompt_embeds.shape[0] == encoder_hidden_states.shape[0]
-            # The channel dimensions should match.
-            assert ip_adapter_image_prompt_embeds.shape[2] == encoder_hidden_states.shape[2]
-            ip_hidden_states = ip_adapter_image_prompt_embeds
-
         residual = hidden_states
 
         if attn.spatial_norm is not None:
@@ -128,23 +115,36 @@ class IPAttnProcessor2_0(torch.nn.Module):
         hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
         hidden_states = hidden_states.to(query.dtype)
 
-        if ip_hidden_states is not None:
-            ip_key = self.to_k_ip(ip_hidden_states)
-            ip_value = self.to_v_ip(ip_hidden_states)
+        if encoder_hidden_states is not None:
+            # If encoder_hidden_states is not None, then we are doing cross-attention, not self-attention. In this case,
+            # we will apply IP-Adapter conditioning. We validate the inputs for IP-Adapter conditioning here.
+            assert ip_adapter_image_prompt_embeds is not None
+            assert len(ip_adapter_image_prompt_embeds) == len(self.weights)
 
-            ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
-            ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+            for ipa_embed, ipa_weights in zip(ip_adapter_image_prompt_embeds, self.weights):
+                # The batch dimensions should match.
+                assert ipa_embed.shape[0] == encoder_hidden_states.shape[0]
+                # The channel dimensions should match.
+                assert ipa_embed.shape[2] == encoder_hidden_states.shape[2]
 
-            # the output of sdp = (batch, num_heads, seq_len, head_dim)
-            # TODO: add support for attn.scale when we move to Torch 2.1
-            ip_hidden_states = F.scaled_dot_product_attention(
-                query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
-            )
+                ip_hidden_states = ipa_embed
 
-            ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
-            ip_hidden_states = ip_hidden_states.to(query.dtype)
+                ip_key = ipa_weights.to_k_ip(ip_hidden_states)
+                ip_value = ipa_weights.to_v_ip(ip_hidden_states)
 
-            hidden_states = hidden_states + self.scale * ip_hidden_states
+                ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+                ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+                # The output of sdpa has shape: (batch, num_heads, seq_len, head_dim)
+                # TODO: add support for attn.scale when we move to Torch 2.1
+                ip_hidden_states = F.scaled_dot_product_attention(
+                    query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
+                )
+
+                ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+                ip_hidden_states = ip_hidden_states.to(query.dtype)
+
+                hidden_states = hidden_states + ipa_weights.scale * ip_hidden_states
 
         # linear proj
         hidden_states = attn.to_out[0](hidden_states)

invokeai/backend/ip_adapter/ip_adapter.py

@@ -1,17 +1,15 @@
 # copied from https://github.com/tencent-ailab/IP-Adapter (Apache License 2.0)
 #   and modified as needed
 
-from contextlib import contextmanager
 from typing import Optional, Union
 
 import torch
-from diffusers.models import UNet2DConditionModel
 from PIL import Image
 from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
 
+from invokeai.backend.ip_adapter.ip_attention_weights import IPAttentionWeights
 from invokeai.backend.model_management.models.base import calc_model_size_by_data
 
-from .attention_processor import AttnProcessor2_0, IPAttnProcessor2_0
 from .resampler import Resampler
 
 
@@ -61,7 +59,7 @@ class IPAdapter:
 
     def __init__(
         self,
-        state_dict: dict[torch.Tensor],
+        state_dict: dict[str, torch.Tensor],
         device: torch.device,
         dtype: torch.dtype = torch.float16,
         num_tokens: int = 4,
@@ -73,12 +71,11 @@ class IPAdapter:
 
         self._clip_image_processor = CLIPImageProcessor()
 
-        self._state_dict = state_dict
+        self._image_proj_model = self._init_image_proj_model(state_dict["image_proj"])
 
-        self._image_proj_model = self._init_image_proj_model(self._state_dict["image_proj"])
-
-        # The _attn_processors will be initialized later when we have access to the UNet.
-        self._attn_processors = None
+        self.attn_weights = IPAttentionWeights.from_state_dict(state_dict["ip_adapter"]).to(
+            self.device, dtype=self.dtype
+        )
 
     def to(self, device: torch.device, dtype: Optional[torch.dtype] = None):
         self.device = device
@@ -86,99 +83,14 @@ class IPAdapter:
             self.dtype = dtype
 
         self._image_proj_model.to(device=self.device, dtype=self.dtype)
-        if self._attn_processors is not None:
-            torch.nn.ModuleList(self._attn_processors.values()).to(device=self.device, dtype=self.dtype)
+        self.attn_weights.to(device=self.device, dtype=self.dtype)
 
     def calc_size(self):
-        if self._state_dict is not None:
-            image_proj_size = sum(
-                [tensor.nelement() * tensor.element_size() for tensor in self._state_dict["image_proj"].values()]
-            )
-            ip_adapter_size = sum(
-                [tensor.nelement() * tensor.element_size() for tensor in self._state_dict["ip_adapter"].values()]
-            )
-            return image_proj_size + ip_adapter_size
-        else:
-            return calc_model_size_by_data(self._image_proj_model) + calc_model_size_by_data(
-                torch.nn.ModuleList(self._attn_processors.values())
-            )
+        return calc_model_size_by_data(self._image_proj_model) + calc_model_size_by_data(self._attn_weights)
 
     def _init_image_proj_model(self, state_dict):
         return ImageProjModel.from_state_dict(state_dict, self._num_tokens).to(self.device, dtype=self.dtype)
 
-    def _prepare_attention_processors(self, unet: UNet2DConditionModel):
-        """Prepare a dict of attention processors that can later be injected into a unet, and load the IP-Adapter
-        attention weights into them.
-
-        Note that the `unet` param is only used to determine attention block dimensions and naming.
-        TODO(ryand): As a future improvement, this could all be inferred from the state_dict when the IPAdapter is
-        intialized.
-        """
-        attn_procs = {}
-        for name in unet.attn_processors.keys():
-            cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
-            if name.startswith("mid_block"):
-                hidden_size = unet.config.block_out_channels[-1]
-            elif name.startswith("up_blocks"):
-                block_id = int(name[len("up_blocks.")])
-                hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
-            elif name.startswith("down_blocks"):
-                block_id = int(name[len("down_blocks.")])
-                hidden_size = unet.config.block_out_channels[block_id]
-            if cross_attention_dim is None:
-                attn_procs[name] = AttnProcessor2_0()
-            else:
-                attn_procs[name] = IPAttnProcessor2_0(
-                    hidden_size=hidden_size,
-                    cross_attention_dim=cross_attention_dim,
-                    scale=1.0,
-                ).to(self.device, dtype=self.dtype)
-
-        ip_layers = torch.nn.ModuleList(attn_procs.values())
-        ip_layers.load_state_dict(self._state_dict["ip_adapter"])
-        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, IPAttnProcessor2_0):
-                    attn_processor.scale = scale
-
-    @contextmanager
-    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:
-            None
-        """
-        if self._attn_processors is None:
-            # We only have to call _prepare_attention_processors(...) once, and then the result is cached and can be
-            # used on any UNet model (with the same dimensions).
-            self._prepare_attention_processors(unet)
-
-        # Set scale
-        self.set_scale(scale)
-        # for attn_processor in self._attn_processors.values():
-        #     if isinstance(attn_processor, IPAttnProcessor2_0):
-        #         attn_processor.scale = scale
-
-        orig_attn_processors = unet.attn_processors
-
-        # Make a (moderately-) shallow copy of the self._attn_processors dict, because unet.set_attn_processor(...)
-        # actually pops elements from the passed dict.
-        ip_adapter_attn_processors = {k: v for k, v in self._attn_processors.items()}
-
-        try:
-            unet.set_attn_processor(ip_adapter_attn_processors)
-            yield None
-        finally:
-            unet.set_attn_processor(orig_attn_processors)
-
     @torch.inference_mode()
     def get_image_embeds(self, pil_image, image_encoder: CLIPVisionModelWithProjection):
         if isinstance(pil_image, Image.Image):

invokeai/backend/ip_adapter/ip_attention_weights.py

@@ -0,0 +1,46 @@
+import torch
+
+
+class IPAttentionProcessorWeights(torch.nn.Module):
+    """The IP-Adapter weights for a single attention processor.
+
+    This class is a torch.nn.Module sub-class to facilitate loading from a state_dict. It does not have a forward(...)
+    method.
+    """
+
+    def __init__(self, in_dim: int, out_dim: int, scale: float = 1.0):
+        super().__init__()
+        self.scale = scale
+        self.to_k_ip = torch.nn.Linear(in_dim, out_dim, bias=False)
+        self.to_v_ip = torch.nn.Linear(in_dim, out_dim, bias=False)
+
+
+class IPAttentionWeights(torch.nn.Module):
+    """A collection of all the `IPAttentionProcessorWeights` objects for an IP-Adapter model.
+
+    This class is a torch.nn.Module sub-class so that it inherits the `.to(...)` functionality. It does not have a
+    forward(...) method.
+    """
+
+    def __init__(self, weights: dict[int, IPAttentionProcessorWeights]):
+        super().__init__()
+        self.weights = weights
+
+    def set_scale(self, scale: float):
+        """Set the scale (a.k.a. 'weight') for all of the `IPAttentionProcessorWeights` in this collection."""
+        for w in self.weights.values():
+            w.scale = scale
+
+    @classmethod
+    def from_state_dict(cls, state_dict: dict[str, torch.Tensor]):
+        attn_proc_weights: dict[int, IPAttentionProcessorWeights] = {}
+
+        for tensor_name, tensor in state_dict.items():
+            if "to_k_ip.weight" in tensor_name:
+                index = int(tensor_name.split(".")[0])
+                attn_proc_weights[index] = IPAttentionProcessorWeights(tensor.shape[1], tensor.shape[0])
+
+        attn_proc_weights_module_dict = torch.nn.ModuleDict(attn_proc_weights)
+        attn_proc_weights_module_dict.load_state_dict(state_dict)
+
+        return cls(attn_proc_weights)

invokeai/backend/ip_adapter/unet_patcher.py

@@ -0,0 +1,51 @@
+from contextlib import contextmanager
+
+from diffusers.models import UNet2DConditionModel
+
+from invokeai.backend.ip_adapter.attention_processor import AttnProcessor2_0, IPAttnProcessor2_0
+from invokeai.backend.ip_adapter.ip_adapter import IPAdapter
+
+
+def _prepare_attention_processors(unet: UNet2DConditionModel, ip_adapters: list[IPAdapter]):
+    """Prepare a dict of attention processors that can be injected into a unet, and load the IP-Adapter attention
+    weights into them.
+
+    Note that the `unet` param is only used to determine attention block dimensions and naming.
+    """
+    # TODO(ryand): This logic can be simplified.
+
+    # Construct a dict of attention processors based on the UNet's architecture.
+    attn_procs = {}
+    for idx, name in enumerate(unet.attn_processors.keys()):
+        cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+        if name.startswith("mid_block"):
+            hidden_size = unet.config.block_out_channels[-1]
+        elif name.startswith("up_blocks"):
+            block_id = int(name[len("up_blocks.")])
+            hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+        elif name.startswith("down_blocks"):
+            block_id = int(name[len("down_blocks.")])
+            hidden_size = unet.config.block_out_channels[block_id]
+
+        if cross_attention_dim is None:
+            attn_procs[name] = AttnProcessor2_0()
+        else:
+            # Collect the weights from each IP Adapter for the idx'th attention processor.
+            attn_procs[name] = IPAttnProcessor2_0([ip_adapter.attn_weights.weights[idx] for ip_adapter in ip_adapters])
+
+
+@contextmanager
+def apply_ip_adapter_attention(cls, unet: UNet2DConditionModel, ip_adapters: list[IPAdapter]):
+    """A context manager that patches `unet` with IP-Adapter attention processors."""
+    attn_procs = _prepare_attention_processors(unet, ip_adapters)
+
+    orig_attn_processors = unet.attn_processors
+
+    try:
+        # Note to future devs: set_attn_processor(...) does something slightly unexpected - it pops elements from the
+        # passed dict. So, if you wanted to keep the dict for future use, you'd have to make a moderately-shallow copy
+        # of it. E.g. `attn_procs_copy = {k: v for k, v in attn_procs.items()}`.
+        unet.set_attn_processor(attn_procs)
+        yield None
+    finally:
+        unet.set_attn_processor(orig_attn_processors)

invokeai/backend/stable_diffusion/diffusers_pipeline.py

@@ -24,6 +24,7 @@ from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
 
 from invokeai.app.services.config import InvokeAIAppConfig
 from invokeai.backend.ip_adapter.ip_adapter import IPAdapter
+from invokeai.backend.ip_adapter.unet_patcher import apply_ip_adapter_attention
 from invokeai.backend.stable_diffusion.diffusion.conditioning_data import ConditioningData
 
 from ..util import auto_detect_slice_size, normalize_device
@@ -434,10 +435,8 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         elif ip_adapter_data is not None:
             # TODO(ryand): Should we raise an exception if both custom attention and IP-Adapter attention are active?
             # As it is now, the IP-Adapter will silently be skipped.
-            weight = ip_adapter_data.weight[0] if isinstance(ip_adapter_data.weight, List) else ip_adapter_data.weight
-            attn_ctx = ip_adapter_data.ip_adapter_model.apply_ip_adapter_attention(
-                unet=self.invokeai_diffuser.model,
-                scale=weight,
+            attn_ctx = apply_ip_adapter_attention(
+                unet=self.invokeai_diffuser.model, ip_adapters=[ipa.ip_adapter_model for ipa in ip_adapter_data]
             )
             self.use_ip_adapter = True
         else:
@@ -513,7 +512,7 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         total_step_count: int,
         additional_guidance: List[Callable] = None,
         control_data: List[ControlNetData] = None,
-        ip_adapter_data: Optional[IPAdapterData] = None,
+        ip_adapter_data: Optional[list[IPAdapterData]] = None,
         t2i_adapter_data: Optional[list[T2IAdapterData]] = None,
     ):
         # invokeai_diffuser has batched timesteps, but diffusers schedulers expect a single value
@@ -527,20 +526,20 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
 
         # handle IP-Adapter
         if self.use_ip_adapter and ip_adapter_data is not None:  # somewhat redundant but logic is clearer
-            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)
-            weight = (
-                ip_adapter_data.weight[step_index]
-                if isinstance(ip_adapter_data.weight, List)
-                else ip_adapter_data.weight
-            )
-            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)
-                ip_adapter_data.ip_adapter_model.set_scale(weight)
-            else:
-                # otherwise, set IP-Adapter scale to 0, so it has no effect
-                ip_adapter_data.ip_adapter_model.set_scale(0.0)
+            for single_ip_adapter_data in ip_adapter_data:
+                first_adapter_step = math.floor(single_ip_adapter_data.begin_step_percent * total_step_count)
+                last_adapter_step = math.ceil(single_ip_adapter_data.end_step_percent * total_step_count)
+                weight = (
+                    single_ip_adapter_data.weight[step_index]
+                    if isinstance(single_ip_adapter_data.weight, List)
+                    else single_ip_adapter_data.weight
+                )
+                if step_index >= first_adapter_step and step_index <= last_adapter_step:
+                    # Only apply this IP-Adapter if the current step is within the IP-Adapter's begin/end step range.
+                    single_ip_adapter_data.ip_adapter_model.attn_weights.set_scale(weight)
+                else:
+                    # Otherwise, set the IP-Adapter's scale to 0, so it has no effect.
+                    single_ip_adapter_data.ip_adapter_model.attn_weights.set_scale(0.0)
 
         # Handle ControlNet(s) and T2I-Adapter(s)
         down_block_additional_residuals = None

invokeai/backend/stable_diffusion/diffusion/shared_invokeai_diffusion.py

@@ -346,12 +346,10 @@ class InvokeAIDiffuserComponent:
         cross_attention_kwargs = None
         if conditioning_data.ip_adapter_conditioning is not None:
             cross_attention_kwargs = {
-                "ip_adapter_image_prompt_embeds": torch.cat(
-                    [
-                        conditioning_data.ip_adapter_conditioning.uncond_image_prompt_embeds,
-                        conditioning_data.ip_adapter_conditioning.cond_image_prompt_embeds,
-                    ]
-                )
+                "ip_adapter_image_prompt_embeds": [
+                    torch.cat([ipa_conditioning.uncond_image_prompt_embeds, ipa_conditioning.cond_image_prompt_embeds])
+                    for ipa_conditioning in conditioning_data.ip_adapter_conditioning
+                ]
             }
 
         added_cond_kwargs = None
@@ -418,7 +416,10 @@ class InvokeAIDiffuserComponent:
         cross_attention_kwargs = None
         if conditioning_data.ip_adapter_conditioning is not None:
             cross_attention_kwargs = {
-                "ip_adapter_image_prompt_embeds": conditioning_data.ip_adapter_conditioning.uncond_image_prompt_embeds
+                "ip_adapter_image_prompt_embeds": [
+                    ipa_conditioning.uncond_image_prompt_embeds
+                    for ipa_conditioning in conditioning_data.ip_adapter_conditioning
+                ]
             }
 
         added_cond_kwargs = None
@@ -444,7 +445,10 @@ class InvokeAIDiffuserComponent:
         cross_attention_kwargs = None
         if conditioning_data.ip_adapter_conditioning is not None:
             cross_attention_kwargs = {
-                "ip_adapter_image_prompt_embeds": conditioning_data.ip_adapter_conditioning.cond_image_prompt_embeds
+                "ip_adapter_image_prompt_embeds": [
+                    ipa_conditioning.cond_image_prompt_embeds
+                    for ipa_conditioning in conditioning_data.ip_adapter_conditioning
+                ]
             }
 
         added_cond_kwargs = None