add IP-Adapter plus (aka fine-grained features)

scripts/conversion/convert_diffusers_ip_adapter.py

@@ -70,12 +70,14 @@ def main() -> None:
     assert isinstance(weights, dict)
     assert sorted(weights.keys()) == ["image_proj", "ip_adapter"]
 
+    fine_grained = "latents" in weights["image_proj"]  # aka IP-Adapter plus
+
     match len(weights["ip_adapter"]):
         case 32:
-            ip_adapter = SD1IPAdapter(target=SD1UNet(in_channels=4))
+            ip_adapter = SD1IPAdapter(target=SD1UNet(in_channels=4), fine_grained=fine_grained)
             cross_attn_mapping = CROSS_ATTN_MAPPING["sd15"]
         case 140:
-            ip_adapter = SDXLIPAdapter(target=SDXLUNet(in_channels=4))
+            ip_adapter = SDXLIPAdapter(target=SDXLUNet(in_channels=4), fine_grained=fine_grained)
             cross_attn_mapping = CROSS_ATTN_MAPPING["sdxl"]
         case _:
             raise ValueError("Unexpected number of keys in input checkpoint")
@@ -86,7 +88,38 @@ def main() -> None:
     state_dict: dict[str, torch.Tensor] = {}
 
     image_proj_weights = weights["image_proj"]
-    image_proj_state_dict: dict[str, torch.Tensor] = {
+    image_proj_state_dict: dict[str, torch.Tensor]
+
+    if fine_grained:
+        w = image_proj_weights
+        image_proj_state_dict = {
+            "LatentsEncoder.Parallel.Parameter.parameter": w["latents"].squeeze(0),  # drop batch dim = 1
+            "Linear_1.weight": w["proj_in.weight"],
+            "Linear_1.bias": w["proj_in.bias"],
+            "Linear_2.weight": w["proj_out.weight"],
+            "Linear_2.bias": w["proj_out.bias"],
+            "LayerNorm.weight": w["norm_out.weight"],
+            "LayerNorm.bias": w["norm_out.bias"],
+        }
+        for i in range(4):
+            t_pfx, s_pfx = f"Transformer.TransformerLayer_{i+1}.Residual_", f"layers.{i}."
+            image_proj_state_dict.update(
+                {
+                    f"{t_pfx}1.Chain.PerceiverAttention.Distribute.LayerNorm_1.weight": w[f"{s_pfx}0.norm1.weight"],
+                    f"{t_pfx}1.Chain.PerceiverAttention.Distribute.LayerNorm_1.bias": w[f"{s_pfx}0.norm1.bias"],
+                    f"{t_pfx}1.Chain.PerceiverAttention.Distribute.LayerNorm_2.weight": w[f"{s_pfx}0.norm2.weight"],
+                    f"{t_pfx}1.Chain.PerceiverAttention.Distribute.LayerNorm_2.bias": w[f"{s_pfx}0.norm2.bias"],
+                    f"{t_pfx}1.Chain.PerceiverAttention.Parallel.Chain_2.Linear.weight": w[f"{s_pfx}0.to_q.weight"],
+                    f"{t_pfx}1.Chain.PerceiverAttention.Parallel.Chain_1.Linear.weight": w[f"{s_pfx}0.to_kv.weight"],
+                    f"{t_pfx}1.Chain.PerceiverAttention.Linear.weight": w[f"{s_pfx}0.to_out.weight"],
+                    f"{t_pfx}2.Chain.LayerNorm.weight": w[f"{s_pfx}1.0.weight"],
+                    f"{t_pfx}2.Chain.LayerNorm.bias": w[f"{s_pfx}1.0.bias"],
+                    f"{t_pfx}2.Chain.FeedForward.Linear_1.weight": w[f"{s_pfx}1.1.weight"],
+                    f"{t_pfx}2.Chain.FeedForward.Linear_2.weight": w[f"{s_pfx}1.3.weight"],
+                }
+            )
+    else:
+        image_proj_state_dict = {
             "Linear.weight": image_proj_weights["proj.weight"],
             "Linear.bias": image_proj_weights["proj.bias"],
             "LayerNorm.weight": image_proj_weights["norm.weight"],

src/refiners/foundationals/latent_diffusion/image_prompt.py

@@ -1,14 +1,16 @@
 from enum import IntEnum
 from functools import partial
 from typing import Generic, TypeVar, Any, Callable, TYPE_CHECKING
+import math
 
-from torch import Tensor, cat, zeros_like, device as Device, dtype as DType
+from jaxtyping import Float
+from torch import Tensor, cat, softmax, zeros_like, device as Device, dtype as DType
 from PIL import Image
 
 from refiners.fluxion.adapters.adapter import Adapter
 from refiners.fluxion.adapters.lora import Lora
 from refiners.foundationals.clip.image_encoder import CLIPImageEncoderH
-from refiners.foundationals.latent_diffusion.cross_attention import CrossAttentionBlock2d
+from refiners.fluxion.context import Contexts
 from refiners.fluxion.layers.attentions import ScaledDotProductAttention
 from refiners.fluxion.utils import image_to_tensor, normalize
 import refiners.fluxion.layers as fl
@@ -26,25 +28,219 @@ class ImageProjection(fl.Chain):
         self,
         clip_image_embedding_dim: int = 1024,
         clip_text_embedding_dim: int = 768,
-        sequence_length: int = 4,
+        num_tokens: int = 4,
         device: Device | str | None = None,
         dtype: DType | None = None,
     ) -> None:
         self.clip_image_embedding_dim = clip_image_embedding_dim
         self.clip_text_embedding_dim = clip_text_embedding_dim
-        self.sequence_length = sequence_length
+        self.num_tokens = num_tokens
         super().__init__(
             fl.Linear(
                 in_features=clip_image_embedding_dim,
-                out_features=clip_text_embedding_dim * sequence_length,
+                out_features=clip_text_embedding_dim * num_tokens,
                 device=device,
                 dtype=dtype,
             ),
-            fl.Reshape(sequence_length, clip_text_embedding_dim),
+            fl.Reshape(num_tokens, clip_text_embedding_dim),
             fl.LayerNorm(normalized_shape=clip_text_embedding_dim, device=device, dtype=dtype),
         )
 
 
+class FeedForward(fl.Chain):
+    def __init__(
+        self,
+        embedding_dim: int,
+        feedforward_dim: int,
+        device: Device | str | None = None,
+        dtype: DType | None = None,
+    ) -> None:
+        self.embedding_dim = embedding_dim
+        self.feedforward_dim = feedforward_dim
+        super().__init__(
+            fl.Linear(
+                in_features=self.embedding_dim,
+                out_features=self.feedforward_dim,
+                bias=False,
+                device=device,
+                dtype=dtype,
+            ),
+            fl.GeLU(),
+            fl.Linear(
+                in_features=self.feedforward_dim,
+                out_features=self.embedding_dim,
+                bias=False,
+                device=device,
+                dtype=dtype,
+            ),
+        )
+
+
+# Adapted from https://github.com/tencent-ailab/IP-Adapter/blob/6212981/ip_adapter/resampler.py
+# See also:
+# - https://github.com/mlfoundations/open_flamingo/blob/main/open_flamingo/src/helpers.py
+# - https://github.com/lucidrains/flamingo-pytorch
+class PerceiverScaledDotProductAttention(fl.Module):
+    def __init__(self, head_dim: int, num_heads: int) -> None:
+        super().__init__()
+        self.num_heads = num_heads
+        # See https://github.com/tencent-ailab/IP-Adapter/blob/6212981/ip_adapter/resampler.py#L69
+        # -> "More stable with f16 than dividing afterwards"
+        self.scale = 1 / math.sqrt(math.sqrt(head_dim))
+
+    def forward(
+        self,
+        key_value: Float[Tensor, "batch sequence_length 2*head_dim*num_heads"],
+        query: Float[Tensor, "batch num_tokens head_dim*num_heads"],
+    ) -> Float[Tensor, "batch num_tokens head_dim*num_heads"]:
+        bs, length, _ = query.shape
+        key, value = key_value.chunk(2, dim=-1)
+
+        q = self.reshape_tensor(query)
+        k = self.reshape_tensor(key)
+        v = self.reshape_tensor(value)
+
+        attention = (q * self.scale) @ (k * self.scale).transpose(-2, -1)
+        attention = softmax(input=attention.float(), dim=-1).type(attention.dtype)
+        attention = attention @ v
+
+        return attention.permute(0, 2, 1, 3).reshape(bs, length, -1)
+
+    def reshape_tensor(
+        self, x: Float[Tensor, "batch length head_dim*num_heads"]
+    ) -> Float[Tensor, "batch num_heads length head_dim"]:
+        bs, length, _ = x.shape
+        x = x.view(bs, length, self.num_heads, -1)
+        x = x.transpose(1, 2)
+        x = x.reshape(bs, self.num_heads, length, -1)
+        return x
+
+
+class PerceiverAttention(fl.Chain):
+    def __init__(
+        self,
+        embedding_dim: int,
+        head_dim: int = 64,
+        num_heads: int = 8,
+        device: Device | str | None = None,
+        dtype: DType | None = None,
+    ) -> None:
+        self.embedding_dim = embedding_dim
+        self.head_dim = head_dim
+        self.inner_dim = head_dim * num_heads
+        super().__init__(
+            fl.Distribute(
+                fl.LayerNorm(normalized_shape=self.embedding_dim, device=device, dtype=dtype),
+                fl.LayerNorm(normalized_shape=self.embedding_dim, device=device, dtype=dtype),
+            ),
+            fl.Parallel(
+                fl.Chain(
+                    fl.Lambda(func=self.to_kv),
+                    fl.Linear(
+                        in_features=self.embedding_dim,
+                        out_features=2 * self.inner_dim,
+                        bias=False,
+                        device=device,
+                        dtype=dtype,
+                    ),  # Wkv
+                ),
+                fl.Chain(
+                    fl.GetArg(index=1),
+                    fl.Linear(
+                        in_features=self.embedding_dim,
+                        out_features=self.inner_dim,
+                        bias=False,
+                        device=device,
+                        dtype=dtype,
+                    ),  # Wq
+                ),
+            ),
+            PerceiverScaledDotProductAttention(head_dim=head_dim, num_heads=num_heads),
+            fl.Linear(
+                in_features=self.inner_dim, out_features=self.embedding_dim, bias=False, device=device, dtype=dtype
+            ),
+        )
+
+    def to_kv(self, x: Tensor, latents: Tensor) -> Tensor:
+        return cat((x, latents), dim=-2)
+
+
+class LatentsEncoder(fl.Chain):
+    def __init__(
+        self,
+        num_tokens: int,
+        embeddding_dim: int,
+        device: Device | str | None = None,
+        dtype: DType | None = None,
+    ) -> None:
+        super().__init__(
+            fl.Parallel(fl.Identity(), fl.Parameter(num_tokens, embeddding_dim, device=device, dtype=dtype)),
+            fl.Lambda(lambda x, p: p.expand(x.shape[0], -1, -1)),
+        )
+
+
+class Transformer(fl.Chain):
+    pass
+
+
+class TransformerLayer(fl.Chain):
+    pass
+
+
+class PerceiverResampler(fl.Chain):
+    def __init__(
+        self,
+        latents_dim: int = 1024,
+        num_attention_layers: int = 8,
+        num_attention_heads: int = 16,
+        head_dim: int = 64,
+        num_tokens: int = 8,
+        input_dim: int = 768,
+        output_dim: int = 1024,
+        device: Device | str | None = None,
+        dtype: DType | None = None,
+    ) -> None:
+        self.latents_dim = latents_dim
+        self.num_attention_layers = num_attention_layers
+        self.head_dim = head_dim
+        self.num_attention_heads = num_attention_heads
+        self.num_tokens = num_tokens
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+        self.feedforward_dim = 4 * self.latents_dim
+        super().__init__(
+            fl.Linear(in_features=input_dim, out_features=latents_dim, device=device, dtype=dtype),
+            fl.SetContext(context="perceiver_resampler", key="x"),
+            LatentsEncoder(num_tokens=num_tokens, embeddding_dim=latents_dim, device=device, dtype=dtype),
+            Transformer(
+                TransformerLayer(
+                    fl.Residual(
+                        fl.Parallel(fl.UseContext(context="perceiver_resampler", key="x"), fl.Identity()),
+                        PerceiverAttention(
+                            embedding_dim=latents_dim,
+                            head_dim=head_dim,
+                            num_heads=num_attention_heads,
+                            device=device,
+                            dtype=dtype,
+                        ),
+                    ),
+                    fl.Residual(
+                        fl.LayerNorm(normalized_shape=latents_dim, device=device, dtype=dtype),
+                        FeedForward(
+                            embedding_dim=latents_dim, feedforward_dim=self.feedforward_dim, device=device, dtype=dtype
+                        ),
+                    ),
+                )
+                for _ in range(num_attention_layers)
+            ),
+            fl.Linear(in_features=latents_dim, out_features=output_dim, device=device, dtype=dtype),
+            fl.LayerNorm(normalized_shape=output_dim, device=device, dtype=dtype),
+        )
+
+    def init_context(self) -> Contexts:
+        return {"perceiver_resampler": {"x": None}}
+
+
 class _CrossAttnIndex(IntEnum):
     TXT_CROSS_ATTN = 0  # text cross-attention
     IMG_CROSS_ATTN = 1  # image cross-attention
@@ -164,32 +360,29 @@ class CrossAttentionAdapter(fl.Chain, Adapter[fl.Attention]):
 class IPAdapter(Generic[T], fl.Chain, Adapter[T]):
     # Prevent PyTorch module registration
     _clip_image_encoder: list[CLIPImageEncoderH]
-    _image_proj: list[ImageProjection]
+    _grid_image_encoder: list[CLIPImageEncoderH]
+    _image_proj: list[fl.Module]
 
     def __init__(
         self,
         target: T,
-        clip_image_encoder: CLIPImageEncoderH | None = None,
+        clip_image_encoder: CLIPImageEncoderH,
+        image_proj: fl.Module,
         scale: float = 1.0,
+        fine_grained: bool = False,
         weights: dict[str, Tensor] | None = None,
     ) -> None:
         with self.setup_adapter(target):
             super().__init__(target)
 
-        cross_attn_2d = target.ensure_find(CrossAttentionBlock2d)
+        self.fine_grained = fine_grained
-
+        self._clip_image_encoder = [clip_image_encoder]
-        self._clip_image_encoder = [clip_image_encoder or CLIPImageEncoderH(device=target.device, dtype=target.dtype)]
+        if fine_grained:
-        self._image_proj = [
+            self._grid_image_encoder = [self.convert_to_grid_features(clip_image_encoder)]
-            ImageProjection(
+        self._image_proj = [image_proj]
-                clip_image_embedding_dim=self.clip_image_encoder.output_dim,
-                clip_text_embedding_dim=cross_attn_2d.context_embedding_dim,
-                device=target.device,
-                dtype=target.dtype,
-            )
-        ]
 
         self.sub_adapters = [
-            CrossAttentionAdapter(target=cross_attn, scale=scale)
+            CrossAttentionAdapter(target=cross_attn, scale=scale, image_sequence_length=self.image_proj.num_tokens)
             for cross_attn in filter(lambda attn: type(attn) != fl.SelfAttention, target.layers(fl.Attention))
         ]
 
@@ -214,7 +407,12 @@ class IPAdapter(Generic[T], fl.Chain, Adapter[T]):
         return self._clip_image_encoder[0]
 
     @property
-    def image_proj(self) -> ImageProjection:
+    def grid_image_encoder(self) -> CLIPImageEncoderH:
+        assert hasattr(self, "_grid_image_encoder")
+        return self._grid_image_encoder[0]
+
+    @property
+    def image_proj(self) -> fl.Module:
         return self._image_proj[0]
 
     def inject(self: "TIPAdapter", parent: fl.Chain | None = None) -> "TIPAdapter":
@@ -228,10 +426,16 @@ class IPAdapter(Generic[T], fl.Chain, Adapter[T]):
         super().eject()
 
     # These should be concatenated to the CLIP text embedding before setting the UNet context
-    def compute_clip_image_embedding(self, image_prompt: Tensor | None) -> Tensor:
+    def compute_clip_image_embedding(self, image_prompt: Tensor) -> Tensor:
-        clip_embedding = self.clip_image_encoder(image_prompt)
+        image_encoder = self.clip_image_encoder if not self.fine_grained else self.grid_image_encoder
+        clip_embedding = image_encoder(image_prompt)
         conditional_embedding = self.image_proj(clip_embedding)
+        if not self.fine_grained:
             negative_embedding = self.image_proj(zeros_like(clip_embedding))
+        else:
+            # See https://github.com/tencent-ailab/IP-Adapter/blob/d580c50/tutorial_train_plus.py#L351-L352
+            clip_embedding = image_encoder(zeros_like(image_prompt))
+            negative_embedding = self.image_proj(clip_embedding)
         return cat((negative_embedding, conditional_embedding))
 
     def preprocess_image(
@@ -247,3 +451,16 @@ class IPAdapter(Generic[T], fl.Chain, Adapter[T]):
             mean=[0.48145466, 0.4578275, 0.40821073] if mean is None else mean,
             std=[0.26862954, 0.26130258, 0.27577711] if std is None else std,
         )
+
+    @staticmethod
+    def convert_to_grid_features(clip_image_encoder: CLIPImageEncoderH) -> CLIPImageEncoderH:
+        encoder_clone = clip_image_encoder.structural_copy()
+        assert isinstance(encoder_clone[-1], fl.Linear)  # final proj
+        assert isinstance(encoder_clone[-2], fl.LayerNorm)  # final normalization
+        assert isinstance(encoder_clone[-3], fl.Lambda)  # pooling (classif token)
+        for _ in range(3):
+            encoder_clone.pop()
+        transfomer_layers = encoder_clone[-1]
+        assert isinstance(transfomer_layers, fl.Chain) and len(transfomer_layers) == 32
+        transfomer_layers.pop()
+        return encoder_clone

src/refiners/foundationals/latent_diffusion/stable_diffusion_1/image_prompt.py

@@ -1,6 +1,53 @@
-from refiners.foundationals.latent_diffusion.image_prompt import IPAdapter
+from torch import Tensor
+
+from refiners.foundationals.clip.image_encoder import CLIPImageEncoderH
+from refiners.foundationals.latent_diffusion.cross_attention import CrossAttentionBlock2d
+from refiners.foundationals.latent_diffusion.image_prompt import IPAdapter, ImageProjection, PerceiverResampler
 from refiners.foundationals.latent_diffusion.stable_diffusion_1 import SD1UNet
 
 
 class SD1IPAdapter(IPAdapter[SD1UNet]):
-    pass
+    def __init__(
+        self,
+        target: SD1UNet,
+        clip_image_encoder: CLIPImageEncoderH | None = None,
+        image_proj: ImageProjection | PerceiverResampler | None = None,
+        scale: float = 1.0,
+        fine_grained: bool = False,
+        weights: dict[str, Tensor] | None = None,
+    ) -> None:
+        clip_image_encoder = clip_image_encoder or CLIPImageEncoderH(device=target.device, dtype=target.dtype)
+
+        if image_proj is None:
+            cross_attn_2d = target.ensure_find(CrossAttentionBlock2d)
+            image_proj = (
+                ImageProjection(
+                    clip_image_embedding_dim=clip_image_encoder.output_dim,
+                    clip_text_embedding_dim=cross_attn_2d.context_embedding_dim,
+                    device=target.device,
+                    dtype=target.dtype,
+                )
+                if not fine_grained
+                else PerceiverResampler(
+                    latents_dim=cross_attn_2d.context_embedding_dim,
+                    num_attention_layers=4,
+                    num_attention_heads=12,
+                    head_dim=64,
+                    num_tokens=16,
+                    input_dim=clip_image_encoder.embedding_dim,  # = dim before final projection
+                    output_dim=cross_attn_2d.context_embedding_dim,
+                    device=target.device,
+                    dtype=target.dtype,
+                )
+            )
+        elif fine_grained:
+            assert isinstance(image_proj, PerceiverResampler)
+
+        super().__init__(
+            target=target,
+            clip_image_encoder=clip_image_encoder,
+            image_proj=image_proj,
+            scale=scale,
+            fine_grained=fine_grained,
+            weights=weights,
+        )

src/refiners/foundationals/latent_diffusion/stable_diffusion_xl/image_prompt.py

@@ -1,6 +1,53 @@
-from refiners.foundationals.latent_diffusion.image_prompt import IPAdapter
+from torch import Tensor
+
+from refiners.foundationals.clip.image_encoder import CLIPImageEncoderH
+from refiners.foundationals.latent_diffusion.cross_attention import CrossAttentionBlock2d
+from refiners.foundationals.latent_diffusion.image_prompt import IPAdapter, ImageProjection, PerceiverResampler
 from refiners.foundationals.latent_diffusion.stable_diffusion_xl import SDXLUNet
 
 
 class SDXLIPAdapter(IPAdapter[SDXLUNet]):
-    pass
+    def __init__(
+        self,
+        target: SDXLUNet,
+        clip_image_encoder: CLIPImageEncoderH | None = None,
+        image_proj: ImageProjection | PerceiverResampler | None = None,
+        scale: float = 1.0,
+        fine_grained: bool = False,
+        weights: dict[str, Tensor] | None = None,
+    ) -> None:
+        clip_image_encoder = clip_image_encoder or CLIPImageEncoderH(device=target.device, dtype=target.dtype)
+
+        if image_proj is None:
+            cross_attn_2d = target.ensure_find(CrossAttentionBlock2d)
+            image_proj = (
+                ImageProjection(
+                    clip_image_embedding_dim=clip_image_encoder.output_dim,
+                    clip_text_embedding_dim=cross_attn_2d.context_embedding_dim,
+                    device=target.device,
+                    dtype=target.dtype,
+                )
+                if not fine_grained
+                else PerceiverResampler(
+                    latents_dim=1280,  # not `cross_attn_2d.context_embedding_dim` in this case
+                    num_attention_layers=4,
+                    num_attention_heads=20,
+                    head_dim=64,
+                    num_tokens=16,
+                    input_dim=clip_image_encoder.embedding_dim,  # = dim before final projection
+                    output_dim=cross_attn_2d.context_embedding_dim,
+                    device=target.device,
+                    dtype=target.dtype,
+                )
+            )
+        elif fine_grained:
+            assert isinstance(image_proj, PerceiverResampler)
+
+        super().__init__(
+            target=target,
+            clip_image_encoder=clip_image_encoder,
+            image_proj=image_proj,
+            scale=scale,
+            fine_grained=fine_grained,
+            weights=weights,
+        )

tests/e2e/test_diffusion.py

@@ -54,6 +54,11 @@ def woman_image(ref_path: Path) -> Image.Image:
     return Image.open(ref_path / "woman.png").convert("RGB")
 
 
+@pytest.fixture(scope="module")
+def statue_image(ref_path: Path) -> Image.Image:
+    return Image.open(ref_path / "statue.png").convert("RGB")
+
+
 @pytest.fixture
 def expected_image_std_random_init(ref_path: Path) -> Image.Image:
     return Image.open(ref_path / "expected_std_random_init.png").convert("RGB")
@@ -79,11 +84,21 @@ def expected_image_ip_adapter_woman(ref_path: Path) -> Image.Image:
     return Image.open(ref_path / "expected_image_ip_adapter_woman.png").convert("RGB")
 
 
+@pytest.fixture
+def expected_image_ip_adapter_plus_statue(ref_path: Path) -> Image.Image:
+    return Image.open(ref_path / "expected_image_ip_adapter_plus_statue.png").convert("RGB")
+
+
 @pytest.fixture
 def expected_image_sdxl_ip_adapter_woman(ref_path: Path) -> Image.Image:
     return Image.open(ref_path / "expected_image_sdxl_ip_adapter_woman.png").convert("RGB")
 
 
+@pytest.fixture
+def expected_image_sdxl_ip_adapter_plus_woman(ref_path: Path) -> Image.Image:
+    return Image.open(ref_path / "expected_image_sdxl_ip_adapter_plus_woman.png").convert("RGB")
+
+
 @pytest.fixture
 def expected_image_ip_adapter_controlnet(ref_path: Path) -> Image.Image:
     return Image.open(ref_path / "expected_ip_adapter_controlnet.png").convert("RGB")
@@ -255,6 +270,15 @@ def ip_adapter_weights(test_weights_path: Path) -> Path:
     return ip_adapter_weights
 
 
+@pytest.fixture(scope="module")
+def ip_adapter_plus_weights(test_weights_path: Path) -> Path:
+    ip_adapter_weights = test_weights_path / "ip-adapter-plus_sd15.safetensors"
+    if not ip_adapter_weights.is_file():
+        warn(f"could not find weights at {ip_adapter_weights}, skipping")
+        pytest.skip(allow_module_level=True)
+    return ip_adapter_weights
+
+
 @pytest.fixture(scope="module")
 def sdxl_ip_adapter_weights(test_weights_path: Path) -> Path:
     ip_adapter_weights = test_weights_path / "ip-adapter_sdxl_vit-h.safetensors"
@@ -264,6 +288,15 @@ def sdxl_ip_adapter_weights(test_weights_path: Path) -> Path:
     return ip_adapter_weights
 
 
+@pytest.fixture(scope="module")
+def sdxl_ip_adapter_plus_weights(test_weights_path: Path) -> Path:
+    ip_adapter_weights = test_weights_path / "ip-adapter-plus_sdxl_vit-h.safetensors"
+    if not ip_adapter_weights.is_file():
+        warn(f"could not find weights at {ip_adapter_weights}, skipping")
+        pytest.skip(allow_module_level=True)
+    return ip_adapter_weights
+
+
 @pytest.fixture(scope="module")
 def image_encoder_weights(test_weights_path: Path) -> Path:
     image_encoder_weights = test_weights_path / "CLIPImageEncoderH.safetensors"
@@ -1189,6 +1222,115 @@ def test_diffusion_ip_adapter_controlnet(
     ensure_similar_images(predicted_image, expected_image_ip_adapter_controlnet)
 
 
+@torch.no_grad()
+def test_diffusion_ip_adapter_plus(
+    sd15_ddim_lda_ft_mse: StableDiffusion_1,
+    ip_adapter_plus_weights: Path,
+    image_encoder_weights: Path,
+    statue_image: Image.Image,
+    expected_image_ip_adapter_plus_statue: Image.Image,
+    test_device: torch.device,
+):
+    sd15 = sd15_ddim_lda_ft_mse.to(dtype=torch.float16)
+    n_steps = 50
+
+    prompt = "best quality, high quality"
+    negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"
+
+    ip_adapter = SD1IPAdapter(
+        target=sd15.unet, weights=load_from_safetensors(ip_adapter_plus_weights), fine_grained=True
+    )
+    ip_adapter.clip_image_encoder.load_from_safetensors(image_encoder_weights)
+    ip_adapter.inject()
+
+    clip_text_embedding = sd15.compute_clip_text_embedding(text=prompt, negative_text=negative_prompt)
+    clip_image_embedding = ip_adapter.compute_clip_image_embedding(ip_adapter.preprocess_image(statue_image))
+
+    negative_text_embedding, conditional_text_embedding = clip_text_embedding.chunk(2)
+    negative_image_embedding, conditional_image_embedding = clip_image_embedding.chunk(2)
+
+    clip_text_embedding = torch.cat(
+        (
+            torch.cat([negative_text_embedding, negative_image_embedding], dim=1),
+            torch.cat([conditional_text_embedding, conditional_image_embedding], dim=1),
+        )
+    )
+
+    sd15.set_num_inference_steps(n_steps)
+
+    manual_seed(42)  # seed=42 is used in the official IP-Adapter demo
+    x = torch.randn(1, 4, 64, 64, device=test_device, dtype=torch.float16)
+
+    for step in sd15.steps:
+        x = sd15(
+            x,
+            step=step,
+            clip_text_embedding=clip_text_embedding,
+            condition_scale=7.5,
+        )
+    predicted_image = sd15.lda.decode_latents(x)
+    predicted_image.save("output.png")
+
+    ensure_similar_images(predicted_image, expected_image_ip_adapter_plus_statue, min_psnr=35, min_ssim=0.98)
+
+
+@torch.no_grad()
+def test_diffusion_sdxl_ip_adapter_plus(
+    sdxl_ddim: StableDiffusion_XL,
+    sdxl_ip_adapter_plus_weights: Path,
+    image_encoder_weights: Path,
+    woman_image: Image.Image,
+    expected_image_sdxl_ip_adapter_plus_woman: Image.Image,
+    test_device: torch.device,
+):
+    sdxl = sdxl_ddim.to(dtype=torch.float16)
+    n_steps = 30
+
+    prompt = "best quality, high quality"
+    negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"
+
+    ip_adapter = SDXLIPAdapter(
+        target=sdxl.unet, weights=load_from_safetensors(sdxl_ip_adapter_plus_weights), fine_grained=True
+    )
+    ip_adapter.clip_image_encoder.load_from_safetensors(image_encoder_weights)
+    ip_adapter.inject()
+
+    clip_text_embedding, pooled_text_embedding = sdxl.compute_clip_text_embedding(
+        text=prompt, negative_text=negative_prompt
+    )
+    clip_image_embedding = ip_adapter.compute_clip_image_embedding(ip_adapter.preprocess_image(woman_image))
+
+    negative_text_embedding, conditional_text_embedding = clip_text_embedding.chunk(2)
+    negative_image_embedding, conditional_image_embedding = clip_image_embedding.chunk(2)
+
+    clip_text_embedding = torch.cat(
+        (
+            torch.cat([negative_text_embedding, negative_image_embedding], dim=1),
+            torch.cat([conditional_text_embedding, conditional_image_embedding], dim=1),
+        )
+    )
+    time_ids = sdxl.default_time_ids
+    sdxl.set_num_inference_steps(n_steps)
+
+    manual_seed(2)
+    x = torch.randn(1, 4, 128, 128, device=test_device, dtype=torch.float16)
+
+    for step in sdxl.steps:
+        x = sdxl(
+            x,
+            step=step,
+            clip_text_embedding=clip_text_embedding,
+            pooled_text_embedding=pooled_text_embedding,
+            time_ids=time_ids,
+            condition_scale=5,
+        )
+    sdxl.lda.to(dtype=torch.float32)
+    predicted_image = sdxl.lda.decode_latents(x.to(dtype=torch.float32))
+    predicted_image.save("output.png")
+
+    ensure_similar_images(predicted_image, expected_image_sdxl_ip_adapter_plus_woman)
+
+
 @torch.no_grad()
 def test_sdxl_random_init(
     sdxl_ddim: StableDiffusion_XL, expected_sdxl_ddim_random_init: Image.Image, test_device: torch.device

tests/e2e/test_diffusion_ref/README.md

@@ -41,6 +41,7 @@ Special cases:
     - `expected_image_sdxl_ip_adapter_woman.png`
     - `expected_ip_adapter_controlnet.png`
     - `expected_t2i_adapter_xl_canny.png`
+    - `expected_image_sdxl_ip_adapter_plus_woman.png`
 
 ## Other images
 
@@ -68,6 +69,8 @@ Special cases:
 
 - `woman.png` [comes from tencent-ailab/IP-Adapter](https://github.com/tencent-ailab/IP-Adapter/blob/8b96670cc5c8ef00278b42c0c7b62fe8a74510b9/assets/images/woman.png).
 
+- `statue.png` [comes from tencent-ailab/IP-Adapter](https://github.com/tencent-ailab/IP-Adapter/blob/d580c50a291566bbf9fc7ac0f760506607297e6d/assets/images/statue.png).
+
 ## VAE without randomness
 
 ```diff