Support multiple IP-Adapters (workflow editor only)

invokeai/app/invocations/latent.py

@@ -10,7 +10,6 @@
 import torchvision.transforms as T
 from diffusers import AutoencoderKL, AutoencoderTiny
 from diffusers.image_processor import VaeImageProcessor
-from diffusers.models import UNet2DConditionModel
 from diffusers.models.adapter import FullAdapterXL, T2IAdapter
 from diffusers.models.attention_processor import (
     AttnProcessor2_0,
@@ -199,7 +198,7 @@ def get_scheduler(
     title="Denoise Latents",
     tags=["latents", "denoise", "txt2img", "t2i", "t2l", "img2img", "i2i", "l2l"],
     category="latents",
-    version="1.2.0",
+    version="1.3.0",
 )
 class DenoiseLatentsInvocation(BaseInvocation):
     """Denoises noisy latents to decodable images"""
@@ -226,7 +225,7 @@ class DenoiseLatentsInvocation(BaseInvocation):
         input=Input.Connection,
         ui_order=5,
     )
-    ip_adapter: Optional[IPAdapterField] = InputField(
+    ip_adapter: Optional[Union[IPAdapterField, list[IPAdapterField]]] = InputField(
         description=FieldDescriptions.ip_adapter, title="IP-Adapter", default=None, input=Input.Connection, ui_order=6
     )
     t2i_adapter: Union[T2IAdapterField, list[T2IAdapterField]] = InputField(
@@ -410,52 +409,65 @@ def prep_control_data(
     def prep_ip_adapter_data(
         self,
         context: InvocationContext,
-        ip_adapter: Optional[IPAdapterField],
+        ip_adapter: Optional[Union[IPAdapterField, list[IPAdapterField]]],
         conditioning_data: ConditioningData,
-        unet: UNet2DConditionModel,
         exit_stack: ExitStack,
-    ) -> Optional[IPAdapterData]:
+    ) -> Optional[list[IPAdapterData]]:
         """If IP-Adapter is enabled, then this function loads the requisite models, and adds the image prompt embeddings
         to the `conditioning_data` (in-place).
         """
         if ip_adapter is None:
             return None
 
-        image_encoder_model_info = context.services.model_manager.get_model(
-            model_name=ip_adapter.image_encoder_model.model_name,
-            model_type=ModelType.CLIPVision,
-            base_model=ip_adapter.image_encoder_model.base_model,
-            context=context,
-        )
+        # ip_adapter could be a list or a single IPAdapterField. Normalize to a list here.
+        if not isinstance(ip_adapter, list):
+            ip_adapter = [ip_adapter]
+
+        if len(ip_adapter) == 0:
+            return None
+
+        ip_adapter_data_list = []
+        conditioning_data.ip_adapter_conditioning = []
+        for single_ip_adapter in ip_adapter:
+            ip_adapter_model: Union[IPAdapter, IPAdapterPlus] = exit_stack.enter_context(
+                context.services.model_manager.get_model(
+                    model_name=single_ip_adapter.ip_adapter_model.model_name,
+                    model_type=ModelType.IPAdapter,
+                    base_model=single_ip_adapter.ip_adapter_model.base_model,
+                    context=context,
+                )
+            )
 
-        ip_adapter_model: Union[IPAdapter, IPAdapterPlus] = exit_stack.enter_context(
-            context.services.model_manager.get_model(
-                model_name=ip_adapter.ip_adapter_model.model_name,
-                model_type=ModelType.IPAdapter,
-                base_model=ip_adapter.ip_adapter_model.base_model,
+            image_encoder_model_info = context.services.model_manager.get_model(
+                model_name=single_ip_adapter.image_encoder_model.model_name,
+                model_type=ModelType.CLIPVision,
+                base_model=single_ip_adapter.image_encoder_model.base_model,
                 context=context,
             )
-        )
 
-        input_image = context.services.images.get_pil_image(ip_adapter.image.image_name)
+            input_image = context.services.images.get_pil_image(single_ip_adapter.image.image_name)
 
-        # TODO(ryand): With some effort, the step of running the CLIP Vision encoder could be done before any other
-        # models are needed in memory. This would help to reduce peak memory utilization in low-memory environments.
-        with image_encoder_model_info as image_encoder_model:
-            # Get image embeddings from CLIP and ImageProjModel.
-            image_prompt_embeds, uncond_image_prompt_embeds = ip_adapter_model.get_image_embeds(
-                input_image, image_encoder_model
-            )
-            conditioning_data.ip_adapter_conditioning = IPAdapterConditioningInfo(
-                image_prompt_embeds, uncond_image_prompt_embeds
+            # TODO(ryand): With some effort, the step of running the CLIP Vision encoder could be done before any other
+            # models are needed in memory. This would help to reduce peak memory utilization in low-memory environments.
+            with image_encoder_model_info as image_encoder_model:
+                # Get image embeddings from CLIP and ImageProjModel.
+                image_prompt_embeds, uncond_image_prompt_embeds = ip_adapter_model.get_image_embeds(
+                    input_image, image_encoder_model
+                )
+                conditioning_data.ip_adapter_conditioning.append(
+                    IPAdapterConditioningInfo(image_prompt_embeds, uncond_image_prompt_embeds)
+                )
+
+            ip_adapter_data_list.append(
+                IPAdapterData(
+                    ip_adapter_model=ip_adapter_model,
+                    weight=single_ip_adapter.weight,
+                    begin_step_percent=single_ip_adapter.begin_step_percent,
+                    end_step_percent=single_ip_adapter.end_step_percent,
+                )
             )
 
-        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,
-        )
+        return ip_adapter_data_list
 
     def run_t2i_adapters(
         self,
@@ -677,7 +689,6 @@ def _lora_loader():
                     context=context,
                     ip_adapter=self.ip_adapter,
                     conditioning_data=conditioning_data,
-                    unet=unet,
                     exit_stack=exit_stack,
                 )
 

invokeai/backend/ip_adapter/attention_processor.py

@@ -8,6 +8,8 @@
 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 +47,16 @@ 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], scales: list[float]):
         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
+        assert len(weights) == len(scales)
 
-        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
+        self._scales = scales
 
     def __call__(
         self,
@@ -67,16 +67,6 @@ def __call__(
         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 +118,36 @@ def __call__(
         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)
+
+            for ipa_embed, ipa_weights, scale in zip(ip_adapter_image_prompt_embeds, self._weights, self._scales):
+                # 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]
+
+                ip_hidden_states = ipa_embed
+
+                ip_key = ipa_weights.to_k_ip(ip_hidden_states)
+                ip_value = ipa_weights.to_v_ip(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)
+                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 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
-            )
+                # 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)
+                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 + self.scale * ip_hidden_states
+                hidden_states = hidden_states + 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,
@@ -73,112 +71,26 @@ def __init__(
 
         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
         if dtype is not None:
             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):