Merge remote-tracking branch 'origin/main' into feat/taesd

# Conflicts:
#	invokeai/backend/model_management/model_probe.py

invokeai/backend/install/model_install_backend.py

@@ -326,6 +326,16 @@ class ModelInstall(object):
                     elif f"learned_embeds.{suffix}" in files:
                         location = self._download_hf_model(repo_id, [f"learned_embeds.{suffix}"], staging)
                         break
+                    elif "image_encoder.txt" in files and f"ip_adapter.{suffix}" in files:  # IP-Adapter
+                        files = ["image_encoder.txt", f"ip_adapter.{suffix}"]
+                        location = self._download_hf_model(repo_id, files, staging)
+                        break
+                    elif f"model.{suffix}" in files and "config.json" in files:
+                        # This elif-condition is pretty fragile, but it is intended to handle CLIP Vision models hosted
+                        # by InvokeAI for use with IP-Adapters.
+                        files = ["config.json", f"model.{suffix}"]
+                        location = self._download_hf_model(repo_id, files, staging)
+                        break
             if not location:
                 logger.warning(f"Could not determine type of repo {repo_id}. Skipping install.")
                 return {}
@@ -534,14 +544,17 @@ def hf_download_with_resume(
         logger.info(f"{model_name}: Downloading...")
 
     try:
-        with open(model_dest, open_mode) as file, tqdm(
-            desc=model_name,
-            initial=exist_size,
-            total=total + exist_size,
-            unit="iB",
-            unit_scale=True,
-            unit_divisor=1000,
-        ) as bar:
+        with (
+            open(model_dest, open_mode) as file,
+            tqdm(
+                desc=model_name,
+                initial=exist_size,
+                total=total + exist_size,
+                unit="iB",
+                unit_scale=True,
+                unit_divisor=1000,
+            ) as bar,
+        ):
             for data in resp.iter_content(chunk_size=1024):
                 size = file.write(data)
                 bar.update(size)

invokeai/backend/ip_adapter/README.md

@@ -0,0 +1,45 @@
+# IP-Adapter Model Formats
+
+The official IP-Adapter models are released here: [h94/IP-Adapter](https://huggingface.co/h94/IP-Adapter)
+
+This official model repo does not integrate well with InvokeAI's current approach to model management, so we have defined a new file structure for IP-Adapter models. The InvokeAI format is described below.
+
+## CLIP Vision Models
+
+CLIP Vision models are organized in `diffusers`` format. The expected directory structure is:
+
+```bash
+ip_adapter_sd_image_encoder/
+├── config.json
+└── model.safetensors
+```
+
+## IP-Adapter Models
+
+IP-Adapter models are stored in a directory containing two files
+- `image_encoder.txt`: A text file containing the model identifier for the CLIP Vision encoder that is intended to be used with this IP-Adapter model.
+- `ip_adapter.bin`: The IP-Adapter weights.
+
+Sample directory structure:
+```bash
+ip_adapter_sd15/
+├── image_encoder.txt
+└── ip_adapter.bin
+```
+
+### Why save the weights in a .safetensors file?
+
+The weights in `ip_adapter.bin` are stored in a nested dict, which is not supported by `safetensors`. This could be solved by splitting `ip_adapter.bin` into multiple files, but for now we have decided to maintain consistency with the checkpoint structure used in the official [h94/IP-Adapter](https://huggingface.co/h94/IP-Adapter) repo.
+
+## InvokeAI Hosted IP-Adapters
+
+Image Encoders:
+- [InvokeAI/ip_adapter_sd_image_encoder](https://huggingface.co/InvokeAI/ip_adapter_sd_image_encoder)
+- [InvokeAI/ip_adapter_sdxl_image_encoder](https://huggingface.co/InvokeAI/ip_adapter_sdxl_image_encoder)
+
+IP-Adapters:
+- [InvokeAI/ip_adapter_sd15](https://huggingface.co/InvokeAI/ip_adapter_sd15)
+- [InvokeAI/ip_adapter_plus_sd15](https://huggingface.co/InvokeAI/ip_adapter_plus_sd15)
+- [InvokeAI/ip_adapter_plus_face_sd15](https://huggingface.co/InvokeAI/ip_adapter_plus_face_sd15)
+- [InvokeAI/ip_adapter_sdxl](https://huggingface.co/InvokeAI/ip_adapter_sdxl)
+- Not yet supported: [InvokeAI/ip_adapter_sdxl_vit_h](https://huggingface.co/InvokeAI/ip_adapter_sdxl_vit_h)

invokeai/backend/ip_adapter/attention_processor.py

@@ -0,0 +1,162 @@
+# copied from https://github.com/tencent-ailab/IP-Adapter (Apache License 2.0)
+#   and modified as needed
+
+# tencent-ailab comment:
+# modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from diffusers.models.attention_processor import AttnProcessor2_0 as DiffusersAttnProcessor2_0
+
+
+# Create a version of AttnProcessor2_0 that is a sub-class of nn.Module. This is required for IP-Adapter state_dict
+# loading.
+class AttnProcessor2_0(DiffusersAttnProcessor2_0, nn.Module):
+    def __init__(self):
+        DiffusersAttnProcessor2_0.__init__(self)
+        nn.Module.__init__(self)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        ip_adapter_image_prompt_embeds=None,
+    ):
+        """Re-definition of DiffusersAttnProcessor2_0.__call__(...) that accepts and ignores the
+        ip_adapter_image_prompt_embeds parameter.
+        """
+        return DiffusersAttnProcessor2_0.__call__(
+            self, attn, hidden_states, encoder_hidden_states, attention_mask, temb
+        )
+
+
+class IPAttnProcessor2_0(torch.nn.Module):
+    r"""
+    Attention processor for IP-Adapater for PyTorch 2.0.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0):
+        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)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        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:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = 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
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        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)
+
+            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
+            )
+
+            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
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states

invokeai/backend/ip_adapter/ip_adapter.py

@@ -0,0 +1,217 @@
+# 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 .attention_processor import AttnProcessor2_0, IPAttnProcessor2_0
+from .resampler import Resampler
+
+
+class ImageProjModel(torch.nn.Module):
+    """Image Projection Model"""
+
+    def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4):
+        super().__init__()
+
+        self.cross_attention_dim = cross_attention_dim
+        self.clip_extra_context_tokens = clip_extra_context_tokens
+        self.proj = torch.nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim)
+        self.norm = torch.nn.LayerNorm(cross_attention_dim)
+
+    @classmethod
+    def from_state_dict(cls, state_dict: dict[torch.Tensor], clip_extra_context_tokens=4):
+        """Initialize an ImageProjModel from a state_dict.
+
+        The cross_attention_dim and clip_embeddings_dim are inferred from the shape of the tensors in the state_dict.
+
+        Args:
+            state_dict (dict[torch.Tensor]): The state_dict of model weights.
+            clip_extra_context_tokens (int, optional): Defaults to 4.
+
+        Returns:
+            ImageProjModel
+        """
+        cross_attention_dim = state_dict["norm.weight"].shape[0]
+        clip_embeddings_dim = state_dict["proj.weight"].shape[-1]
+
+        model = cls(cross_attention_dim, clip_embeddings_dim, clip_extra_context_tokens)
+
+        model.load_state_dict(state_dict)
+        return model
+
+    def forward(self, image_embeds):
+        embeds = image_embeds
+        clip_extra_context_tokens = self.proj(embeds).reshape(
+            -1, self.clip_extra_context_tokens, self.cross_attention_dim
+        )
+        clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
+        return clip_extra_context_tokens
+
+
+class IPAdapter:
+    """IP-Adapter: https://arxiv.org/pdf/2308.06721.pdf"""
+
+    def __init__(
+        self,
+        state_dict: dict[torch.Tensor],
+        device: torch.device,
+        dtype: torch.dtype = torch.float16,
+        num_tokens: int = 4,
+    ):
+        self.device = device
+        self.dtype = dtype
+
+        self._num_tokens = num_tokens
+
+        self._clip_image_processor = CLIPImageProcessor()
+
+        self._state_dict = state_dict
+
+        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
+
+    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)
+
+    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):
+            pil_image = [pil_image]
+        clip_image = self._clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+        clip_image_embeds = image_encoder(clip_image.to(self.device, dtype=self.dtype)).image_embeds
+        image_prompt_embeds = self._image_proj_model(clip_image_embeds)
+        uncond_image_prompt_embeds = self._image_proj_model(torch.zeros_like(clip_image_embeds))
+        return image_prompt_embeds, uncond_image_prompt_embeds
+
+
+class IPAdapterPlus(IPAdapter):
+    """IP-Adapter with fine-grained features"""
+
+    def _init_image_proj_model(self, state_dict):
+        return Resampler.from_state_dict(
+            state_dict=state_dict,
+            depth=4,
+            dim_head=64,
+            heads=12,
+            num_queries=self._num_tokens,
+            ff_mult=4,
+        ).to(self.device, dtype=self.dtype)
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image, image_encoder: CLIPVisionModelWithProjection):
+        if isinstance(pil_image, Image.Image):
+            pil_image = [pil_image]
+        clip_image = self._clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+        clip_image = clip_image.to(self.device, dtype=self.dtype)
+        clip_image_embeds = image_encoder(clip_image, output_hidden_states=True).hidden_states[-2]
+        image_prompt_embeds = self._image_proj_model(clip_image_embeds)
+        uncond_clip_image_embeds = image_encoder(torch.zeros_like(clip_image), output_hidden_states=True).hidden_states[
+            -2
+        ]
+        uncond_image_prompt_embeds = self._image_proj_model(uncond_clip_image_embeds)
+        return image_prompt_embeds, uncond_image_prompt_embeds
+
+
+def build_ip_adapter(
+    ip_adapter_ckpt_path: str, device: torch.device, dtype: torch.dtype = torch.float16
+) -> Union[IPAdapter, IPAdapterPlus]:
+    state_dict = torch.load(ip_adapter_ckpt_path, map_location="cpu")
+
+    # Determine if the state_dict is from an IPAdapter or IPAdapterPlus based on the image_proj weights that it
+    # contains.
+    is_plus = "proj.weight" not in state_dict["image_proj"]
+
+    if is_plus:
+        return IPAdapterPlus(state_dict, device=device, dtype=dtype)
+    else:
+        return IPAdapter(state_dict, device=device, dtype=dtype)

invokeai/backend/ip_adapter/resampler.py

@@ -0,0 +1,158 @@
+# copied from https://github.com/tencent-ailab/IP-Adapter (Apache License 2.0)
+
+# tencent ailab comment: modified from
+# https://github.com/mlfoundations/open_flamingo/blob/main/open_flamingo/src/helpers.py
+import math
+
+import torch
+import torch.nn as nn
+
+
+# FFN
+def FeedForward(dim, mult=4):
+    inner_dim = int(dim * mult)
+    return nn.Sequential(
+        nn.LayerNorm(dim),
+        nn.Linear(dim, inner_dim, bias=False),
+        nn.GELU(),
+        nn.Linear(inner_dim, dim, bias=False),
+    )
+
+
+def reshape_tensor(x, heads):
+    bs, length, width = x.shape
+    # (bs, length, width) --> (bs, length, n_heads, dim_per_head)
+    x = x.view(bs, length, heads, -1)
+    # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
+    x = x.transpose(1, 2)
+    # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
+    x = x.reshape(bs, heads, length, -1)
+    return x
+
+
+class PerceiverAttention(nn.Module):
+    def __init__(self, *, dim, dim_head=64, heads=8):
+        super().__init__()
+        self.scale = dim_head**-0.5
+        self.dim_head = dim_head
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+
+        b, l, _ = latents.shape
+
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
+
+        q = reshape_tensor(q, self.heads)
+        k = reshape_tensor(k, self.heads)
+        v = reshape_tensor(v, self.heads)
+
+        # attention
+        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
+        weight = (q * scale) @ (k * scale).transpose(-2, -1)  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        out = weight @ v
+
+        out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
+
+        return self.to_out(out)
+
+
+class Resampler(nn.Module):
+    def __init__(
+        self,
+        dim=1024,
+        depth=8,
+        dim_head=64,
+        heads=16,
+        num_queries=8,
+        embedding_dim=768,
+        output_dim=1024,
+        ff_mult=4,
+    ):
+        super().__init__()
+
+        self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
+
+        self.proj_in = nn.Linear(embedding_dim, dim)
+
+        self.proj_out = nn.Linear(dim, output_dim)
+        self.norm_out = nn.LayerNorm(output_dim)
+
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        FeedForward(dim=dim, mult=ff_mult),
+                    ]
+                )
+            )
+
+    @classmethod
+    def from_state_dict(cls, state_dict: dict[torch.Tensor], depth=8, dim_head=64, heads=16, num_queries=8, ff_mult=4):
+        """A convenience function that initializes a Resampler from a state_dict.
+
+        Some of the shape parameters are inferred from the state_dict (e.g. dim, embedding_dim, etc.). At the time of
+        writing, we did not have a need for inferring ALL of the shape parameters from the state_dict, but this would be
+        possible if needed in the future.
+
+        Args:
+            state_dict (dict[torch.Tensor]): The state_dict to load.
+            depth (int, optional):
+            dim_head (int, optional):
+            heads (int, optional):
+            ff_mult (int, optional):
+
+        Returns:
+            Resampler
+        """
+        dim = state_dict["latents"].shape[2]
+        num_queries = state_dict["latents"].shape[1]
+        embedding_dim = state_dict["proj_in.weight"].shape[-1]
+        output_dim = state_dict["norm_out.weight"].shape[0]
+
+        model = cls(
+            dim=dim,
+            depth=depth,
+            dim_head=dim_head,
+            heads=heads,
+            num_queries=num_queries,
+            embedding_dim=embedding_dim,
+            output_dim=output_dim,
+            ff_mult=ff_mult,
+        )
+        model.load_state_dict(state_dict)
+        return model
+
+    def forward(self, x):
+        latents = self.latents.repeat(x.size(0), 1, 1)
+
+        x = self.proj_in(x)
+
+        for attn, ff in self.layers:
+            latents = attn(x, latents) + latents
+            latents = ff(latents) + latents
+
+        latents = self.proj_out(latents)
+        return self.norm_out(latents)

invokeai/backend/model_management/model_manager.py

@@ -25,6 +25,7 @@ Models are described using four attributes:
          ModelType.Lora -- a LoRA or LyCORIS fine-tune
          ModelType.TextualInversion -- a textual inversion embedding
          ModelType.ControlNet -- a ControlNet model
+         ModelType.IPAdapter -- an IPAdapter model
 
   3) BaseModelType -- an enum indicating the stable diffusion base model, one of:
          BaseModelType.StableDiffusion1
@@ -1000,8 +1001,8 @@ class ModelManager(object):
                                 new_models_found = True
                             except DuplicateModelException as e:
                                 self.logger.warning(e)
-                            except InvalidModelException:
-                                self.logger.warning(f"Not a valid model: {model_path}")
+                            except InvalidModelException as e:
+                                self.logger.warning(f"Not a valid model: {model_path}. {e}")
                             except NotImplementedError as e:
                                 self.logger.warning(e)
 

invokeai/backend/model_management/model_probe.py

@@ -8,6 +8,8 @@ import torch
 from diffusers import ConfigMixin, ModelMixin
 from picklescan.scanner import scan_file_path
 
+from invokeai.backend.model_management.models.ip_adapter import IPAdapterModelFormat
+
 from .models import (
     BaseModelType,
     InvalidModelException,
@@ -53,6 +55,7 @@ class ModelProbe(object):
         "AutoencoderKL": ModelType.Vae,
         "AutoencoderTiny": ModelType.Vae,
         "ControlNetModel": ModelType.ControlNet,
+        "CLIPVisionModelWithProjection": ModelType.CLIPVision,
     }
 
     @classmethod
@@ -119,14 +122,18 @@ class ModelProbe(object):
                     and prediction_type == SchedulerPredictionType.VPrediction
                 ),
                 format=format,
-                image_size=1024
-                if (base_type in {BaseModelType.StableDiffusionXL, BaseModelType.StableDiffusionXLRefiner})
-                else 768
-                if (
-                    base_type == BaseModelType.StableDiffusion2
-                    and prediction_type == SchedulerPredictionType.VPrediction
-                )
-                else 512,
+                image_size=(
+                    1024
+                    if (base_type in {BaseModelType.StableDiffusionXL, BaseModelType.StableDiffusionXLRefiner})
+                    else (
+                        768
+                        if (
+                            base_type == BaseModelType.StableDiffusion2
+                            and prediction_type == SchedulerPredictionType.VPrediction
+                        )
+                        else 512
+                    )
+                ),
             )
         except Exception:
             raise
@@ -179,9 +186,10 @@ class ModelProbe(object):
                 return ModelType.ONNX
             if (folder_path / "learned_embeds.bin").exists():
                 return ModelType.TextualInversion
-
             if (folder_path / "pytorch_lora_weights.bin").exists():
                 return ModelType.Lora
+            if (folder_path / "image_encoder.txt").exists():
+                return ModelType.IPAdapter
 
             i = folder_path / "model_index.json"
             c = folder_path / "config.json"
@@ -190,7 +198,12 @@ class ModelProbe(object):
             if config_path:
                 with open(config_path, "r") as file:
                     conf = json.load(file)
-                class_name = conf["_class_name"]
+                if "_class_name" in conf:
+                    class_name = conf["_class_name"]
+                elif "architectures" in conf:
+                    class_name = conf["architectures"][0]
+                else:
+                    class_name = None
             else:
                 error_hint = f"No model_index.json or config.json found in {folder_path}."
 
@@ -374,6 +387,16 @@ class ControlNetCheckpointProbe(CheckpointProbeBase):
         raise InvalidModelException("Unable to determine base type for {self.checkpoint_path}")
 
 
+class IPAdapterCheckpointProbe(CheckpointProbeBase):
+    def get_base_type(self) -> BaseModelType:
+        raise NotImplementedError()
+
+
+class CLIPVisionCheckpointProbe(CheckpointProbeBase):
+    def get_base_type(self) -> BaseModelType:
+        raise NotImplementedError()
+
+
 ########################################################
 # classes for probing folders
 #######################################################
@@ -493,11 +516,13 @@ class ControlNetFolderProbe(FolderProbeBase):
         base_model = (
             BaseModelType.StableDiffusion1
             if dimension == 768
-            else BaseModelType.StableDiffusion2
-            if dimension == 1024
-            else BaseModelType.StableDiffusionXL
-            if dimension == 2048
-            else None
+            else (
+                BaseModelType.StableDiffusion2
+                if dimension == 1024
+                else BaseModelType.StableDiffusionXL
+                if dimension == 2048
+                else None
+            )
         )
         if not base_model:
             raise InvalidModelException(f"Unable to determine model base for {self.folder_path}")
@@ -517,15 +542,47 @@ class LoRAFolderProbe(FolderProbeBase):
         return LoRACheckpointProbe(model_file, None).get_base_type()
 
 
+class IPAdapterFolderProbe(FolderProbeBase):
+    def get_format(self) -> str:
+        return IPAdapterModelFormat.InvokeAI.value
+
+    def get_base_type(self) -> BaseModelType:
+        model_file = self.folder_path / "ip_adapter.bin"
+        if not model_file.exists():
+            raise InvalidModelException("Unknown IP-Adapter model format.")
+
+        state_dict = torch.load(model_file, map_location="cpu")
+        cross_attention_dim = state_dict["ip_adapter"]["1.to_k_ip.weight"].shape[-1]
+        if cross_attention_dim == 768:
+            return BaseModelType.StableDiffusion1
+        elif cross_attention_dim == 1024:
+            return BaseModelType.StableDiffusion2
+        elif cross_attention_dim == 2048:
+            return BaseModelType.StableDiffusionXL
+        else:
+            raise InvalidModelException(f"IP-Adapter had unexpected cross-attention dimension: {cross_attention_dim}.")
+
+
+class CLIPVisionFolderProbe(FolderProbeBase):
+    def get_base_type(self) -> BaseModelType:
+        return BaseModelType.Any
+
+
 ############## register probe classes ######
 ModelProbe.register_probe("diffusers", ModelType.Main, PipelineFolderProbe)
 ModelProbe.register_probe("diffusers", ModelType.Vae, VaeFolderProbe)
 ModelProbe.register_probe("diffusers", ModelType.Lora, LoRAFolderProbe)
 ModelProbe.register_probe("diffusers", ModelType.TextualInversion, TextualInversionFolderProbe)
 ModelProbe.register_probe("diffusers", ModelType.ControlNet, ControlNetFolderProbe)
+ModelProbe.register_probe("diffusers", ModelType.IPAdapter, IPAdapterFolderProbe)
+ModelProbe.register_probe("diffusers", ModelType.CLIPVision, CLIPVisionFolderProbe)
+
 ModelProbe.register_probe("checkpoint", ModelType.Main, PipelineCheckpointProbe)
 ModelProbe.register_probe("checkpoint", ModelType.Vae, VaeCheckpointProbe)
 ModelProbe.register_probe("checkpoint", ModelType.Lora, LoRACheckpointProbe)
 ModelProbe.register_probe("checkpoint", ModelType.TextualInversion, TextualInversionCheckpointProbe)
 ModelProbe.register_probe("checkpoint", ModelType.ControlNet, ControlNetCheckpointProbe)
+ModelProbe.register_probe("checkpoint", ModelType.IPAdapter, IPAdapterCheckpointProbe)
+ModelProbe.register_probe("checkpoint", ModelType.CLIPVision, CLIPVisionCheckpointProbe)
+
 ModelProbe.register_probe("onnx", ModelType.ONNX, ONNXFolderProbe)

invokeai/backend/model_management/model_search.py

@@ -79,7 +79,7 @@ class ModelSearch(ABC):
                         self._models_found += 1
                         self._scanned_dirs.add(path)
                     except Exception as e:
-                        self.logger.warning(str(e))
+                        self.logger.warning(f"Failed to process '{path}': {e}")
 
             for f in files:
                 path = Path(root) / f
@@ -90,7 +90,7 @@ class ModelSearch(ABC):
                         self.on_model_found(path)
                         self._models_found += 1
                     except Exception as e:
-                        self.logger.warning(str(e))
+                        self.logger.warning(f"Failed to process '{path}': {e}")
 
 
 class FindModels(ModelSearch):

invokeai/backend/model_management/models/__init__.py

@@ -18,7 +18,9 @@ from .base import (  # noqa: F401
     SilenceWarnings,
     SubModelType,
 )
+from .clip_vision import CLIPVisionModel
 from .controlnet import ControlNetModel  # TODO:
+from .ip_adapter import IPAdapterModel
 from .lora import LoRAModel
 from .sdxl import StableDiffusionXLModel
 from .stable_diffusion import StableDiffusion1Model, StableDiffusion2Model
@@ -34,6 +36,8 @@ MODEL_CLASSES = {
         ModelType.Lora: LoRAModel,
         ModelType.ControlNet: ControlNetModel,
         ModelType.TextualInversion: TextualInversionModel,
+        ModelType.IPAdapter: IPAdapterModel,
+        ModelType.CLIPVision: CLIPVisionModel,
     },
     BaseModelType.StableDiffusion2: {
         ModelType.ONNX: ONNXStableDiffusion2Model,
@@ -42,6 +46,8 @@ MODEL_CLASSES = {
         ModelType.Lora: LoRAModel,
         ModelType.ControlNet: ControlNetModel,
         ModelType.TextualInversion: TextualInversionModel,
+        ModelType.IPAdapter: IPAdapterModel,
+        ModelType.CLIPVision: CLIPVisionModel,
     },
     BaseModelType.StableDiffusionXL: {
         ModelType.Main: StableDiffusionXLModel,
@@ -51,6 +57,8 @@ MODEL_CLASSES = {
         ModelType.ControlNet: ControlNetModel,
         ModelType.TextualInversion: TextualInversionModel,
         ModelType.ONNX: ONNXStableDiffusion2Model,
+        ModelType.IPAdapter: IPAdapterModel,
+        ModelType.CLIPVision: CLIPVisionModel,
     },
     BaseModelType.StableDiffusionXLRefiner: {
         ModelType.Main: StableDiffusionXLModel,
@@ -60,6 +68,19 @@ MODEL_CLASSES = {
         ModelType.ControlNet: ControlNetModel,
         ModelType.TextualInversion: TextualInversionModel,
         ModelType.ONNX: ONNXStableDiffusion2Model,
+        ModelType.IPAdapter: IPAdapterModel,
+        ModelType.CLIPVision: CLIPVisionModel,
+    },
+    BaseModelType.Any: {
+        ModelType.CLIPVision: CLIPVisionModel,
+        # The following model types are not expected to be used with BaseModelType.Any.
+        ModelType.ONNX: ONNXStableDiffusion2Model,
+        ModelType.Main: StableDiffusion2Model,
+        ModelType.Vae: VaeModel,
+        ModelType.Lora: LoRAModel,
+        ModelType.ControlNet: ControlNetModel,
+        ModelType.TextualInversion: TextualInversionModel,
+        ModelType.IPAdapter: IPAdapterModel,
     },
     # BaseModelType.Kandinsky2_1: {
     #    ModelType.Main: Kandinsky2_1Model,

invokeai/backend/model_management/models/base.py

@@ -36,6 +36,7 @@ class ModelNotFoundException(Exception):
 
 
 class BaseModelType(str, Enum):
+    Any = "any"  # For models that are not associated with any particular base model.
     StableDiffusion1 = "sd-1"
     StableDiffusion2 = "sd-2"
     StableDiffusionXL = "sdxl"
@@ -50,6 +51,8 @@ class ModelType(str, Enum):
     Lora = "lora"
     ControlNet = "controlnet"  # used by model_probe
     TextualInversion = "embedding"
+    IPAdapter = "ip_adapter"
+    CLIPVision = "clip_vision"
 
 
 class SubModelType(str, Enum):

invokeai/backend/model_management/models/clip_vision.py

@@ -0,0 +1,82 @@
+import os
+from enum import Enum
+from typing import Literal, Optional
+
+import torch
+from transformers import CLIPVisionModelWithProjection
+
+from invokeai.backend.model_management.models.base import (
+    BaseModelType,
+    InvalidModelException,
+    ModelBase,
+    ModelConfigBase,
+    ModelType,
+    SubModelType,
+    calc_model_size_by_data,
+    calc_model_size_by_fs,
+    classproperty,
+)
+
+
+class CLIPVisionModelFormat(str, Enum):
+    Diffusers = "diffusers"
+
+
+class CLIPVisionModel(ModelBase):
+    class DiffusersConfig(ModelConfigBase):
+        model_format: Literal[CLIPVisionModelFormat.Diffusers]
+
+    def __init__(self, model_path: str, base_model: BaseModelType, model_type: ModelType):
+        assert model_type == ModelType.CLIPVision
+        super().__init__(model_path, base_model, model_type)
+
+        self.model_size = calc_model_size_by_fs(self.model_path)
+
+    @classmethod
+    def detect_format(cls, path: str) -> str:
+        if not os.path.exists(path):
+            raise ModuleNotFoundError(f"No CLIP Vision model at path '{path}'.")
+
+        if os.path.isdir(path) and os.path.exists(os.path.join(path, "config.json")):
+            return CLIPVisionModelFormat.Diffusers
+
+        raise InvalidModelException(f"Unexpected CLIP Vision model format: {path}")
+
+    @classproperty
+    def save_to_config(cls) -> bool:
+        return True
+
+    def get_size(self, child_type: Optional[SubModelType] = None) -> int:
+        if child_type is not None:
+            raise ValueError("There are no child models in a CLIP Vision model.")
+
+        return self.model_size
+
+    def get_model(
+        self,
+        torch_dtype: Optional[torch.dtype],
+        child_type: Optional[SubModelType] = None,
+    ) -> CLIPVisionModelWithProjection:
+        if child_type is not None:
+            raise ValueError("There are no child models in a CLIP Vision model.")
+
+        model = CLIPVisionModelWithProjection.from_pretrained(self.model_path, torch_dtype=torch_dtype)
+
+        # Calculate a more accurate model size.
+        self.model_size = calc_model_size_by_data(model)
+
+        return model
+
+    @classmethod
+    def convert_if_required(
+        cls,
+        model_path: str,
+        output_path: str,
+        config: ModelConfigBase,
+        base_model: BaseModelType,
+    ) -> str:
+        format = cls.detect_format(model_path)
+        if format == CLIPVisionModelFormat.Diffusers:
+            return model_path
+        else:
+            raise ValueError(f"Unsupported format: '{format}'.")

invokeai/backend/model_management/models/ip_adapter.py

@@ -0,0 +1,92 @@
+import os
+import typing
+from enum import Enum
+from typing import Literal, Optional
+
+import torch
+
+from invokeai.backend.ip_adapter.ip_adapter import IPAdapter, IPAdapterPlus, build_ip_adapter
+from invokeai.backend.model_management.models.base import (
+    BaseModelType,
+    InvalidModelException,
+    ModelBase,
+    ModelConfigBase,
+    ModelType,
+    SubModelType,
+    classproperty,
+)
+
+
+class IPAdapterModelFormat(str, Enum):
+    # The custom IP-Adapter model format defined by InvokeAI.
+    InvokeAI = "invokeai"
+
+
+class IPAdapterModel(ModelBase):
+    class InvokeAIConfig(ModelConfigBase):
+        model_format: Literal[IPAdapterModelFormat.InvokeAI]
+
+    def __init__(self, model_path: str, base_model: BaseModelType, model_type: ModelType):
+        assert model_type == ModelType.IPAdapter
+        super().__init__(model_path, base_model, model_type)
+
+        self.model_size = os.path.getsize(self.model_path)
+
+    @classmethod
+    def detect_format(cls, path: str) -> str:
+        if not os.path.exists(path):
+            raise ModuleNotFoundError(f"No IP-Adapter model at path '{path}'.")
+
+        if os.path.isdir(path):
+            model_file = os.path.join(path, "ip_adapter.bin")
+            image_encoder_config_file = os.path.join(path, "image_encoder.txt")
+            if os.path.exists(model_file) and os.path.exists(image_encoder_config_file):
+                return IPAdapterModelFormat.InvokeAI
+
+        raise InvalidModelException(f"Unexpected IP-Adapter model format: {path}")
+
+    @classproperty
+    def save_to_config(cls) -> bool:
+        return True
+
+    def get_size(self, child_type: Optional[SubModelType] = None) -> int:
+        if child_type is not None:
+            raise ValueError("There are no child models in an IP-Adapter model.")
+
+        return self.model_size
+
+    def get_model(
+        self,
+        torch_dtype: Optional[torch.dtype],
+        child_type: Optional[SubModelType] = None,
+    ) -> typing.Union[IPAdapter, IPAdapterPlus]:
+        if child_type is not None:
+            raise ValueError("There are no child models in an IP-Adapter model.")
+
+        return build_ip_adapter(
+            ip_adapter_ckpt_path=os.path.join(self.model_path, "ip_adapter.bin"), device="cpu", dtype=torch_dtype
+        )
+
+    @classmethod
+    def convert_if_required(
+        cls,
+        model_path: str,
+        output_path: str,
+        config: ModelConfigBase,
+        base_model: BaseModelType,
+    ) -> str:
+        format = cls.detect_format(model_path)
+        if format == IPAdapterModelFormat.InvokeAI:
+            return model_path
+        else:
+            raise ValueError(f"Unsupported format: '{format}'.")
+
+
+def get_ip_adapter_image_encoder_model_id(model_path: str):
+    """Read the ID of the image encoder associated with the IP-Adapter at `model_path`."""
+    image_encoder_config_file = os.path.join(model_path, "image_encoder.txt")
+
+    with open(image_encoder_config_file, "r") as f:
+        image_encoder_model = f.readline().strip()
+
+    return image_encoder_model

invokeai/backend/stable_diffusion/__init__.py

@@ -1,15 +1,6 @@
 """
 Initialization file for the invokeai.backend.stable_diffusion package
 """
-from .diffusers_pipeline import (  # noqa: F401
-    ConditioningData,
-    PipelineIntermediateState,
-    StableDiffusionGeneratorPipeline,
-)
+from .diffusers_pipeline import PipelineIntermediateState, StableDiffusionGeneratorPipeline  # noqa: F401
 from .diffusion import InvokeAIDiffuserComponent  # noqa: F401
 from .diffusion.cross_attention_map_saving import AttentionMapSaver  # noqa: F401
-from .diffusion.shared_invokeai_diffusion import (  # noqa: F401
-    BasicConditioningInfo,
-    PostprocessingSettings,
-    SDXLConditioningInfo,
-)

invokeai/backend/stable_diffusion/diffusers_pipeline.py

@@ -1,8 +1,8 @@
 from __future__ import annotations
 
-import dataclasses
-import inspect
-from dataclasses import dataclass, field
+import math
+from contextlib import nullcontext
+from dataclasses import dataclass
 from typing import Any, Callable, List, Optional, Union
 
 import einops
@@ -23,9 +23,11 @@ from pydantic import Field
 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.stable_diffusion.diffusion.conditioning_data import ConditioningData
 
 from ..util import auto_detect_slice_size, normalize_device
-from .diffusion import AttentionMapSaver, BasicConditioningInfo, InvokeAIDiffuserComponent, PostprocessingSettings
+from .diffusion import AttentionMapSaver, InvokeAIDiffuserComponent
 
 
 @dataclass
@@ -95,7 +97,7 @@ class AddsMaskGuidance:
         # Mask anything that has the same shape as prev_sample, return others as-is.
         return output_class(
             {
-                k: (self.apply_mask(v, self._t_for_field(k, t)) if are_like_tensors(prev_sample, v) else v)
+                k: self.apply_mask(v, self._t_for_field(k, t)) if are_like_tensors(prev_sample, v) else v
                 for k, v in step_output.items()
             }
         )
@@ -162,39 +164,13 @@ class ControlNetData:
 
 
 @dataclass
-class ConditioningData:
-    unconditioned_embeddings: BasicConditioningInfo
-    text_embeddings: BasicConditioningInfo
-    guidance_scale: Union[float, List[float]]
-    """
-    Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
-    `guidance_scale` is defined as `w` of equation 2. of [Imagen Paper](https://arxiv.org/pdf/2205.11487.pdf).
-    Guidance scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate
-    images that are closely linked to the text `prompt`, usually at the expense of lower image quality.
-    """
-    extra: Optional[InvokeAIDiffuserComponent.ExtraConditioningInfo] = None
-    scheduler_args: dict[str, Any] = field(default_factory=dict)
-    """
-    Additional arguments to pass to invokeai_diffuser.do_latent_postprocessing().
-    """
-    postprocessing_settings: Optional[PostprocessingSettings] = None
-
-    @property
-    def dtype(self):
-        return self.text_embeddings.dtype
-
-    def add_scheduler_args_if_applicable(self, scheduler, **kwargs):
-        scheduler_args = dict(self.scheduler_args)
-        step_method = inspect.signature(scheduler.step)
-        for name, value in kwargs.items():
-            try:
-                step_method.bind_partial(**{name: value})
-            except TypeError:
-                # FIXME: don't silently discard arguments
-                pass  # debug("%s does not accept argument named %r", scheduler, name)
-            else:
-                scheduler_args[name] = value
-        return dataclasses.replace(self, scheduler_args=scheduler_args)
+class IPAdapterData:
+    ip_adapter_model: IPAdapter = Field(default=None)
+    # 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
@@ -277,6 +253,7 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         )
         self.invokeai_diffuser = InvokeAIDiffuserComponent(self.unet, self._unet_forward)
         self.control_model = control_model
+        self.use_ip_adapter = False
 
     def _adjust_memory_efficient_attention(self, latents: torch.Tensor):
         """
@@ -349,6 +326,7 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         additional_guidance: List[Callable] = None,
         callback: Callable[[PipelineIntermediateState], None] = None,
         control_data: List[ControlNetData] = None,
+        ip_adapter_data: Optional[IPAdapterData] = None,
         mask: Optional[torch.Tensor] = None,
         masked_latents: Optional[torch.Tensor] = None,
         seed: Optional[int] = None,
@@ -400,6 +378,7 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
                 conditioning_data,
                 additional_guidance=additional_guidance,
                 control_data=control_data,
+                ip_adapter_data=ip_adapter_data,
                 callback=callback,
             )
         finally:
@@ -419,6 +398,7 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         *,
         additional_guidance: List[Callable] = None,
         control_data: List[ControlNetData] = None,
+        ip_adapter_data: Optional[IPAdapterData] = None,
         callback: Callable[[PipelineIntermediateState], None] = None,
     ):
         self._adjust_memory_efficient_attention(latents)
@@ -431,12 +411,26 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         if timesteps.shape[0] == 0:
             return latents, attention_map_saver
 
-        extra_conditioning_info = conditioning_data.extra
-        with self.invokeai_diffuser.custom_attention_context(
-            self.invokeai_diffuser.model,
-            extra_conditioning_info=extra_conditioning_info,
-            step_count=len(self.scheduler.timesteps),
-        ):
+        if conditioning_data.extra is not None and conditioning_data.extra.wants_cross_attention_control:
+            attn_ctx = self.invokeai_diffuser.custom_attention_context(
+                self.invokeai_diffuser.model,
+                extra_conditioning_info=conditioning_data.extra,
+                step_count=len(self.scheduler.timesteps),
+            )
+            self.use_ip_adapter = False
+        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,
+            )
+            self.use_ip_adapter = True
+        else:
+            attn_ctx = nullcontext()
+
+        with attn_ctx:
             if callback is not None:
                 callback(
                     PipelineIntermediateState(
@@ -459,6 +453,7 @@ 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
 
@@ -504,6 +499,7 @@ 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]
@@ -514,6 +510,24 @@ 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 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)
+
+        # 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:

invokeai/backend/stable_diffusion/diffusion/__init__.py

@@ -3,9 +3,4 @@ Initialization file for invokeai.models.diffusion
 """
 from .cross_attention_control import InvokeAICrossAttentionMixin  # noqa: F401
 from .cross_attention_map_saving import AttentionMapSaver  # noqa: F401
-from .shared_invokeai_diffusion import (  # noqa: F401
-    BasicConditioningInfo,
-    InvokeAIDiffuserComponent,
-    PostprocessingSettings,
-    SDXLConditioningInfo,
-)
+from .shared_invokeai_diffusion import InvokeAIDiffuserComponent  # noqa: F401

invokeai/backend/stable_diffusion/diffusion/conditioning_data.py

@@ -0,0 +1,101 @@
+import dataclasses
+import inspect
+from dataclasses import dataclass, field
+from typing import Any, List, Optional, Union
+
+import torch
+
+from .cross_attention_control import Arguments
+
+
+@dataclass
+class ExtraConditioningInfo:
+    tokens_count_including_eos_bos: int
+    cross_attention_control_args: Optional[Arguments] = None
+
+    @property
+    def wants_cross_attention_control(self):
+        return self.cross_attention_control_args is not None
+
+
+@dataclass
+class BasicConditioningInfo:
+    embeds: torch.Tensor
+    # TODO(ryand): Right now we awkwardly copy the extra conditioning info from here up to `ConditioningData`. This
+    # should only be stored in one place.
+    extra_conditioning: Optional[ExtraConditioningInfo]
+    # weight: float
+    # mode: ConditioningAlgo
+
+    def to(self, device, dtype=None):
+        self.embeds = self.embeds.to(device=device, dtype=dtype)
+        return self
+
+
+@dataclass
+class SDXLConditioningInfo(BasicConditioningInfo):
+    pooled_embeds: torch.Tensor
+    add_time_ids: torch.Tensor
+
+    def to(self, device, dtype=None):
+        self.pooled_embeds = self.pooled_embeds.to(device=device, dtype=dtype)
+        self.add_time_ids = self.add_time_ids.to(device=device, dtype=dtype)
+        return super().to(device=device, dtype=dtype)
+
+
+@dataclass(frozen=True)
+class PostprocessingSettings:
+    threshold: float
+    warmup: float
+    h_symmetry_time_pct: Optional[float]
+    v_symmetry_time_pct: Optional[float]
+
+
+@dataclass
+class IPAdapterConditioningInfo:
+    cond_image_prompt_embeds: torch.Tensor
+    """IP-Adapter image encoder conditioning embeddings.
+    Shape: (batch_size, num_tokens, encoding_dim).
+    """
+    uncond_image_prompt_embeds: torch.Tensor
+    """IP-Adapter image encoding embeddings to use for unconditional generation.
+    Shape: (batch_size, num_tokens, encoding_dim).
+    """
+
+
+@dataclass
+class ConditioningData:
+    unconditioned_embeddings: BasicConditioningInfo
+    text_embeddings: BasicConditioningInfo
+    guidance_scale: Union[float, List[float]]
+    """
+    Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+    `guidance_scale` is defined as `w` of equation 2. of [Imagen Paper](https://arxiv.org/pdf/2205.11487.pdf).
+    Guidance scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate
+    images that are closely linked to the text `prompt`, usually at the expense of lower image quality.
+    """
+    extra: Optional[ExtraConditioningInfo] = None
+    scheduler_args: dict[str, Any] = field(default_factory=dict)
+    """
+    Additional arguments to pass to invokeai_diffuser.do_latent_postprocessing().
+    """
+    postprocessing_settings: Optional[PostprocessingSettings] = None
+
+    ip_adapter_conditioning: Optional[IPAdapterConditioningInfo] = None
+
+    @property
+    def dtype(self):
+        return self.text_embeddings.dtype
+
+    def add_scheduler_args_if_applicable(self, scheduler, **kwargs):
+        scheduler_args = dict(self.scheduler_args)
+        step_method = inspect.signature(scheduler.step)
+        for name, value in kwargs.items():
+            try:
+                step_method.bind_partial(**{name: value})
+            except TypeError:
+                # FIXME: don't silently discard arguments
+                pass  # debug("%s does not accept argument named %r", scheduler, name)
+            else:
+                scheduler_args[name] = value
+        return dataclasses.replace(self, scheduler_args=scheduler_args)

invokeai/backend/stable_diffusion/diffusion/cross_attention_control.py

@@ -376,11 +376,11 @@ def get_cross_attention_modules(model, which: CrossAttentionType) -> list[tuple[
         # non-fatal error but .swap() won't work.
         logger.error(
             f"Error! CrossAttentionControl found an unexpected number of {cross_attention_class} modules in the model "
-            + f"(expected {expected_count}, found {cross_attention_modules_in_model_count}). Either monkey-patching failed "
-            + "or some assumption has changed about the structure of the model itself. Please fix the monkey-patching, "
-            + f"and/or update the {expected_count} above to an appropriate number, and/or find and inform someone who knows "
-            + "what it means. This error is non-fatal, but it is likely that .swap() and attention map display will not "
-            + "work properly until it is fixed."
+            f"(expected {expected_count}, found {cross_attention_modules_in_model_count}). Either monkey-patching "
+            "failed or some assumption has changed about the structure of the model itself. Please fix the "
+            f"monkey-patching, and/or update the {expected_count} above to an appropriate number, and/or find and "
+            "inform someone who knows what it means. This error is non-fatal, but it is likely that .swap() and "
+            "attention map display will not work properly until it is fixed."
         )
     return attention_module_tuples
 
@@ -577,6 +577,7 @@ class SlicedSwapCrossAttnProcesser(SlicedAttnProcessor):
         attention_mask=None,
         # kwargs
         swap_cross_attn_context: SwapCrossAttnContext = None,
+        **kwargs,
     ):
         attention_type = CrossAttentionType.SELF if encoder_hidden_states is None else CrossAttentionType.TOKENS
 

invokeai/backend/stable_diffusion/diffusion/shared_invokeai_diffusion.py

@@ -2,7 +2,6 @@ from __future__ import annotations
 
 import math
 from contextlib import contextmanager
-from dataclasses import dataclass
 from typing import Any, Callable, Optional, Union
 
 import torch
@@ -10,9 +9,14 @@ from diffusers import UNet2DConditionModel
 from typing_extensions import TypeAlias
 
 from invokeai.app.services.config import InvokeAIAppConfig
+from invokeai.backend.stable_diffusion.diffusion.conditioning_data import (
+    ConditioningData,
+    ExtraConditioningInfo,
+    PostprocessingSettings,
+    SDXLConditioningInfo,
+)
 
 from .cross_attention_control import (
-    Arguments,
     Context,
     CrossAttentionType,
     SwapCrossAttnContext,
@@ -31,37 +35,6 @@ ModelForwardCallback: TypeAlias = Union[
 ]
 
 
-@dataclass
-class BasicConditioningInfo:
-    embeds: torch.Tensor
-    extra_conditioning: Optional[InvokeAIDiffuserComponent.ExtraConditioningInfo]
-    # weight: float
-    # mode: ConditioningAlgo
-
-    def to(self, device, dtype=None):
-        self.embeds = self.embeds.to(device=device, dtype=dtype)
-        return self
-
-
-@dataclass
-class SDXLConditioningInfo(BasicConditioningInfo):
-    pooled_embeds: torch.Tensor
-    add_time_ids: torch.Tensor
-
-    def to(self, device, dtype=None):
-        self.pooled_embeds = self.pooled_embeds.to(device=device, dtype=dtype)
-        self.add_time_ids = self.add_time_ids.to(device=device, dtype=dtype)
-        return super().to(device=device, dtype=dtype)
-
-
-@dataclass(frozen=True)
-class PostprocessingSettings:
-    threshold: float
-    warmup: float
-    h_symmetry_time_pct: Optional[float]
-    v_symmetry_time_pct: Optional[float]
-
-
 class InvokeAIDiffuserComponent:
     """
     The aim of this component is to provide a single place for code that can be applied identically to
@@ -75,15 +48,6 @@ class InvokeAIDiffuserComponent:
     debug_thresholding = False
     sequential_guidance = False
 
-    @dataclass
-    class ExtraConditioningInfo:
-        tokens_count_including_eos_bos: int
-        cross_attention_control_args: Optional[Arguments] = None
-
-        @property
-        def wants_cross_attention_control(self):
-            return self.cross_attention_control_args is not None
-
     def __init__(
         self,
         model,
@@ -103,30 +67,26 @@ class InvokeAIDiffuserComponent:
     @contextmanager
     def custom_attention_context(
         self,
-        unet: UNet2DConditionModel,  # note: also may futz with the text encoder depending on requested LoRAs
+        unet: UNet2DConditionModel,
         extra_conditioning_info: Optional[ExtraConditioningInfo],
         step_count: int,
     ):
-        old_attn_processors = None
-        if extra_conditioning_info and (extra_conditioning_info.wants_cross_attention_control):
-            old_attn_processors = unet.attn_processors
-            # Load lora conditions into the model
-            if extra_conditioning_info.wants_cross_attention_control:
-                self.cross_attention_control_context = Context(
-                    arguments=extra_conditioning_info.cross_attention_control_args,
-                    step_count=step_count,
-                )
-                setup_cross_attention_control_attention_processors(
-                    unet,
-                    self.cross_attention_control_context,
-                )
+        old_attn_processors = unet.attn_processors
 
         try:
+            self.cross_attention_control_context = Context(
+                arguments=extra_conditioning_info.cross_attention_control_args,
+                step_count=step_count,
+            )
+            setup_cross_attention_control_attention_processors(
+                unet,
+                self.cross_attention_control_context,
+            )
+
             yield None
         finally:
             self.cross_attention_control_context = None
-            if old_attn_processors is not None:
-                unet.set_attn_processor(old_attn_processors)
+            unet.set_attn_processor(old_attn_processors)
             # TODO resuscitate attention map saving
             # self.remove_attention_map_saving()
 
@@ -376,11 +336,24 @@ class InvokeAIDiffuserComponent:
 
     # methods below are called from do_diffusion_step and should be considered private to this class.
 
-    def _apply_standard_conditioning(self, x, sigma, conditioning_data, **kwargs):
-        # fast batched path
+    def _apply_standard_conditioning(self, x, sigma, conditioning_data: ConditioningData, **kwargs):
+        """Runs the conditioned and unconditioned UNet forward passes in a single batch for faster inference speed at
+        the cost of higher memory usage.
+        """
         x_twice = torch.cat([x] * 2)
         sigma_twice = torch.cat([sigma] * 2)
 
+        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,
+                    ]
+                )
+            }
+
         added_cond_kwargs = None
         if type(conditioning_data.text_embeddings) is SDXLConditioningInfo:
             added_cond_kwargs = {
@@ -408,6 +381,7 @@ class InvokeAIDiffuserComponent:
             x_twice,
             sigma_twice,
             both_conditionings,
+            cross_attention_kwargs=cross_attention_kwargs,
             encoder_attention_mask=encoder_attention_mask,
             added_cond_kwargs=added_cond_kwargs,
             **kwargs,
@@ -419,9 +393,12 @@ class InvokeAIDiffuserComponent:
         self,
         x: torch.Tensor,
         sigma,
-        conditioning_data,
+        conditioning_data: ConditioningData,
         **kwargs,
     ):
+        """Runs the conditioned and unconditioned UNet forward passes sequentially for lower memory usage at the cost of
+        slower execution speed.
+        """
         # low-memory sequential path
         uncond_down_block, cond_down_block = None, None
         down_block_additional_residuals = kwargs.pop("down_block_additional_residuals", None)
@@ -437,6 +414,13 @@ class InvokeAIDiffuserComponent:
         if mid_block_additional_residual is not None:
             uncond_mid_block, cond_mid_block = mid_block_additional_residual.chunk(2)
 
+        # Run unconditional UNet denoising.
+        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
+            }
+
         added_cond_kwargs = None
         is_sdxl = type(conditioning_data.text_embeddings) is SDXLConditioningInfo
         if is_sdxl:
@@ -449,12 +433,21 @@ class InvokeAIDiffuserComponent:
             x,
             sigma,
             conditioning_data.unconditioned_embeddings.embeds,
+            cross_attention_kwargs=cross_attention_kwargs,
             down_block_additional_residuals=uncond_down_block,
             mid_block_additional_residual=uncond_mid_block,
             added_cond_kwargs=added_cond_kwargs,
             **kwargs,
         )
 
+        # Run conditional UNet denoising.
+        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
+            }
+
+        added_cond_kwargs = None
         if is_sdxl:
             added_cond_kwargs = {
                 "text_embeds": conditioning_data.text_embeddings.pooled_embeds,
@@ -465,6 +458,7 @@ class InvokeAIDiffuserComponent:
             x,
             sigma,
             conditioning_data.text_embeddings.embeds,
+            cross_attention_kwargs=cross_attention_kwargs,
             down_block_additional_residuals=cond_down_block,
             mid_block_additional_residual=cond_mid_block,
             added_cond_kwargs=added_cond_kwargs,