add support for SDXL IP-Adapter

This only supports the latest SDXL IP-Adapter release (2023.9.8) which
builds upon the ViT-H/14 CLIP image encoder.

scripts/conversion/convert_diffusers_ip_adapter.py

@@ -4,60 +4,9 @@ import argparse
 
 import torch
 
-from refiners.foundationals.latent_diffusion import SD1UNet, SD1IPAdapter
+from refiners.foundationals.latent_diffusion import SD1UNet, SD1IPAdapter, SDXLUNet, SDXLIPAdapter
 from refiners.fluxion.utils import save_to_safetensors
 
-
-def main() -> None:
-    parser = argparse.ArgumentParser(description="Converts a IP-Adapter diffusers model to refiners.")
-    parser.add_argument(
-        "--from",
-        type=str,
-        dest="source_path",
-        default="ip-adapter_sd15.bin",
-        help="Path to the source model. (default: 'ip-adapter_sd15.bin').",
-    )
-    parser.add_argument(
-        "--to",
-        type=str,
-        dest="output_path",
-        default="ip-adapter_sd15.safetensors",
-        help="Path to save the converted model. (default: 'ip-adapter_sd15.safetensors').",
-    )
-    parser.add_argument("--verbose", action="store_true", dest="verbose")
-    parser.add_argument("--half", action="store_true", dest="half")
-    args = parser.parse_args()
-    if args.output_path is None:
-        args.output_path = f"{Path(args.source_path).stem}.safetensors"
-
-    weights: dict[str, Any] = torch.load(f=args.source_path, map_location="cpu")  # type: ignore
-    assert isinstance(weights, dict)
-    assert sorted(weights.keys()) == ["image_proj", "ip_adapter"]
-
-    unet = SD1UNet(in_channels=4)
-
-    ip_adapter = SD1IPAdapter(target=unet)
-
-    # Manual conversion to avoid any runtime dependency on IP-Adapter[1] custom classes
-    # [1]: https://github.com/tencent-ailab/IP-Adapter
-
-    state_dict: dict[str, torch.Tensor] = {}
-
-    image_proj_weights = weights["image_proj"]
-    image_proj_state_dict: dict[str, torch.Tensor] = {
-        "Linear.weight": image_proj_weights["proj.weight"],
-        "Linear.bias": image_proj_weights["proj.bias"],
-        "LayerNorm.weight": image_proj_weights["norm.weight"],
-        "LayerNorm.bias": image_proj_weights["norm.bias"],
-    }
-    ip_adapter.image_proj.load_state_dict(state_dict=image_proj_state_dict)
-
-    for k, v in image_proj_state_dict.items():
-        state_dict[f"image_proj.{k}"] = v
-
-    ip_adapter_weights: dict[str, torch.Tensor] = weights["ip_adapter"]
-    assert len(ip_adapter.sub_adapters) == len(ip_adapter_weights.keys()) // 2
-
 # Running:
 #
 #     from diffusers import UNet2DConditionModel
@@ -84,7 +33,72 @@ def main() -> None:
 #     DownBlocks  -> [1, 3, 5, 7, 9, 11]
 #     MiddleBlock -> [31]
 #     UpBlocks    -> [13, 15, 17, 19, 21, 23, 25, 27, 29]
-    cross_attn_mapping: list[int] = [1, 3, 5, 7, 9, 11, 31, 13, 15, 17, 19, 21, 23, 25, 27, 29]
+#
+# Same for SDXL with more layers (70 cross-attentions vs. 16)
+CROSS_ATTN_MAPPING: dict[str, list[int]] = {
+    "sd15": list(range(1, 12, 2)) + [31] + list(range(13, 30, 2)),
+    "sdxl": list(range(1, 48, 2)) + list(range(121, 140, 2)) + list(range(49, 120, 2)),
+}
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Converts a IP-Adapter diffusers model to refiners.")
+    parser.add_argument(
+        "--from",
+        type=str,
+        required=True,
+        dest="source_path",
+        help="Path to the source model. (e.g.: 'ip-adapter_sd15.bin').",
+    )
+    parser.add_argument(
+        "--to",
+        type=str,
+        dest="output_path",
+        default=None,
+        help=(
+            "Path to save the converted model. If not specified, the output path will be the source path with the"
+            " extension changed to .safetensors."
+        ),
+    )
+    parser.add_argument("--verbose", action="store_true", dest="verbose")
+    parser.add_argument("--half", action="store_true", dest="half")
+    args = parser.parse_args()
+    if args.output_path is None:
+        args.output_path = f"{Path(args.source_path).stem}.safetensors"
+
+    weights: dict[str, Any] = torch.load(f=args.source_path, map_location="cpu")  # type: ignore
+    assert isinstance(weights, dict)
+    assert sorted(weights.keys()) == ["image_proj", "ip_adapter"]
+
+    match len(weights["ip_adapter"]):
+        case 32:
+            ip_adapter = SD1IPAdapter(target=SD1UNet(in_channels=4))
+            cross_attn_mapping = CROSS_ATTN_MAPPING["sd15"]
+        case 140:
+            ip_adapter = SDXLIPAdapter(target=SDXLUNet(in_channels=4))
+            cross_attn_mapping = CROSS_ATTN_MAPPING["sdxl"]
+        case _:
+            raise ValueError("Unexpected number of keys in input checkpoint")
+
+    # Manual conversion to avoid any runtime dependency on IP-Adapter[1] custom classes
+    # [1]: https://github.com/tencent-ailab/IP-Adapter
+
+    state_dict: dict[str, torch.Tensor] = {}
+
+    image_proj_weights = weights["image_proj"]
+    image_proj_state_dict: dict[str, torch.Tensor] = {
+        "Linear.weight": image_proj_weights["proj.weight"],
+        "Linear.bias": image_proj_weights["proj.bias"],
+        "LayerNorm.weight": image_proj_weights["norm.weight"],
+        "LayerNorm.bias": image_proj_weights["norm.bias"],
+    }
+    ip_adapter.image_proj.load_state_dict(state_dict=image_proj_state_dict)
+
+    for k, v in image_proj_state_dict.items():
+        state_dict[f"image_proj.{k}"] = v
+
+    ip_adapter_weights: dict[str, torch.Tensor] = weights["ip_adapter"]
+    assert len(ip_adapter.sub_adapters) == len(ip_adapter_weights.keys()) // 2
 
     for i, cross_attn in enumerate(ip_adapter.sub_adapters):
         cross_attn_index = cross_attn_mapping[i]
@@ -107,6 +121,8 @@ def main() -> None:
 
     if args.half:
         state_dict = {key: value.half() for key, value in state_dict.items()}
+    if args.output_path is None:
+        args.output_path = f"{Path(args.source_path).stem}.safetensors"
     save_to_safetensors(path=args.output_path, tensors=state_dict)
 
 

src/refiners/foundationals/clip/image_encoder.py

@@ -121,6 +121,7 @@ class CLIPImageEncoder(fl.Chain):
     structural_attrs = [
         "image_size",
         "embedding_dim",
+        "output_dim",
         "patch_size",
         "num_layers",
         "num_attention_heads",
@@ -142,6 +143,7 @@ class CLIPImageEncoder(fl.Chain):
     ) -> None:
         self.image_size = image_size
         self.embedding_dim = embedding_dim
+        self.output_dim = output_dim
         self.patch_size = patch_size
         self.num_layers = num_layers
         self.num_attention_heads = num_attention_heads

src/refiners/foundationals/latent_diffusion/__init__.py

@@ -15,6 +15,7 @@ from refiners.foundationals.latent_diffusion.stable_diffusion_1 import (
 from refiners.foundationals.latent_diffusion.stable_diffusion_xl import (
     SDXLUNet,
     DoubleTextEncoder,
+    SDXLIPAdapter,
 )
 
 
@@ -26,6 +27,7 @@ __all__ = [
     "SD1IPAdapter",
     "SDXLUNet",
     "DoubleTextEncoder",
+    "SDXLIPAdapter",
     "DPMSolver",
     "Scheduler",
     "CLIPTextEncoderL",

src/refiners/foundationals/latent_diffusion/image_prompt.py

@@ -1,21 +1,23 @@
 from enum import IntEnum
 from functools import partial
-from typing import Generic, TypeVar, Any, Callable
+from typing import Generic, TypeVar, Any, Callable, TYPE_CHECKING
 
 from torch import Tensor, as_tensor, cat, 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 CLIPImageEncoder
+from refiners.foundationals.clip.image_encoder import CLIPImageEncoderH
-from refiners.foundationals.latent_diffusion.stable_diffusion_1.unet import SD1UNet
+from refiners.foundationals.latent_diffusion.cross_attention import CrossAttentionBlock2d
-from refiners.foundationals.latent_diffusion.stable_diffusion_xl.unet import SDXLUNet
-from refiners.fluxion.layers.module import Module
 from refiners.fluxion.layers.attentions import ScaledDotProductAttention
 from refiners.fluxion.utils import image_to_tensor
 import refiners.fluxion.layers as fl
 
-T = TypeVar("T", bound=SD1UNet | SDXLUNet)
+if TYPE_CHECKING:
+    from refiners.foundationals.latent_diffusion.stable_diffusion_1.unet import SD1UNet
+    from refiners.foundationals.latent_diffusion.stable_diffusion_xl.unet import SDXLUNet
+
+T = TypeVar("T", bound="SD1UNet | SDXLUNet")
 TIPAdapter = TypeVar("TIPAdapter", bound="IPAdapter[Any]")  # Self (see PEP 673)
 
 
@@ -128,7 +130,7 @@ class CrossAttentionAdapter(fl.Chain, Adapter[fl.Attention]):
     def scale_outputs(self, x: Tensor) -> Tensor:
         return x * self.scale
 
-    def _predicate(self, k: type[Module]) -> Callable[[fl.Module, fl.Chain], bool]:
+    def _predicate(self, k: type[fl.Module]) -> Callable[[fl.Module, fl.Chain], bool]:
         def f(m: fl.Module, _: fl.Chain) -> bool:
             if isinstance(m, Lora):  # do not adapt LoRAs
                 raise StopIteration
@@ -167,15 +169,22 @@ class IPAdapter(Generic[T], fl.Chain, Adapter[T]):
     def __init__(
         self,
         target: T,
-        clip_image_encoder: CLIPImageEncoder,
+        clip_image_encoder: CLIPImageEncoderH | None = None,
         scale: float = 1.0,
         weights: dict[str, Tensor] | None = None,
     ) -> None:
         with self.setup_adapter(target):
             super().__init__(target)
 
-        self.clip_image_encoder = clip_image_encoder
+        cross_attn_2d = target.ensure_find(CrossAttentionBlock2d)
-        self.image_proj = ImageProjection(device=target.device, dtype=target.dtype)
+
+        self.clip_image_encoder = clip_image_encoder or CLIPImageEncoderH(device=target.device, dtype=target.dtype)
+        self.image_proj = ImageProjection(
+            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)

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

@@ -1,22 +1,6 @@
-from torch import Tensor
-
 from refiners.foundationals.latent_diffusion.image_prompt import IPAdapter
 from refiners.foundationals.latent_diffusion.stable_diffusion_1 import SD1UNet
 
-from refiners.foundationals.clip.image_encoder import CLIPImageEncoderH
-
 
 class SD1IPAdapter(IPAdapter[SD1UNet]):
-    def __init__(
+    pass
-        self,
-        target: SD1UNet,
-        clip_image_encoder: CLIPImageEncoderH | None = None,
-        scale: float = 1.0,
-        weights: dict[str, Tensor] | None = None,
-    ) -> None:
-        super().__init__(
-            target=target,
-            clip_image_encoder=clip_image_encoder or CLIPImageEncoderH(device=target.device, dtype=target.dtype),
-            scale=scale,
-            weights=weights,
-        )

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

@@ -1,9 +1,12 @@
 from refiners.foundationals.latent_diffusion.stable_diffusion_xl.unet import SDXLUNet
 from refiners.foundationals.latent_diffusion.stable_diffusion_xl.text_encoder import DoubleTextEncoder
 from refiners.foundationals.latent_diffusion.stable_diffusion_xl.model import StableDiffusion_XL
+from refiners.foundationals.latent_diffusion.stable_diffusion_xl.image_prompt import SDXLIPAdapter
+
 
 __all__ = [
     "SDXLUNet",
     "DoubleTextEncoder",
     "StableDiffusion_XL",
+    "SDXLIPAdapter",
 ]

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

@@ -0,0 +1,6 @@
+from refiners.foundationals.latent_diffusion.image_prompt import IPAdapter
+from refiners.foundationals.latent_diffusion.stable_diffusion_xl import SDXLUNet
+
+
+class SDXLIPAdapter(IPAdapter[SDXLUNet]):
+    pass

tests/e2e/test_diffusion.py

@@ -14,6 +14,7 @@ from refiners.foundationals.latent_diffusion import (
     SD1UNet,
     SD1ControlnetAdapter,
     SD1IPAdapter,
+    SDXLIPAdapter,
 )
 from refiners.foundationals.latent_diffusion.lora import SD1LoraAdapter
 from refiners.foundationals.latent_diffusion.schedulers import DDIM
@@ -74,6 +75,11 @@ 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_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_sdxl_ddim_random_init(ref_path: Path) -> Image.Image:
     return Image.open(fp=ref_path / "expected_cutecat_sdxl_ddim_random_init.png").convert(mode="RGB")
@@ -217,6 +223,15 @@ def ip_adapter_weights(test_weights_path: Path) -> Path:
     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"
+    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"
@@ -1050,6 +1065,64 @@ def test_diffusion_ip_adapter(
     ensure_similar_images(predicted_image, expected_image_ip_adapter_woman)
 
 
+@torch.no_grad()
+def test_diffusion_sdxl_ip_adapter(
+    sdxl_ddim: StableDiffusion_XL,
+    sdxl_ip_adapter_weights: Path,
+    image_encoder_weights: Path,
+    woman_image: Image.Image,
+    expected_image_sdxl_ip_adapter_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_weights))
+    ip_adapter.clip_image_encoder.load_from_safetensors(image_encoder_weights)
+    ip_adapter.inject()
+
+    with torch.no_grad():
+        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)
+
+    with torch.no_grad():
+        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,
+            )
+        # See https://huggingface.co/madebyollin/sdxl-vae-fp16-fix: "SDXL-VAE generates NaNs in fp16 because the
+        # internal activation values are too big"
+        sdxl.lda.to(dtype=torch.float32)
+        predicted_image = sdxl.lda.decode_latents(x.to(dtype=torch.float32))
+
+    ensure_similar_images(predicted_image, expected_image_sdxl_ip_adapter_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

@@ -35,7 +35,7 @@ output.images[0].save("std_random_init_expected.png")
 Special cases:
 
 - `expected_refonly.png` has been generated [with Stable Diffusion web UI](https://github.com/AUTOMATIC1111/stable-diffusion-webui).
-- `expected_inpainting_refonly.png`, `expected_image_ip_adapter_woman.png` have been generated with refiners itself (and inspected so that it looks reasonable).
+- `expected_inpainting_refonly.png`, `expected_image_ip_adapter_woman.png`, `expected_image_sdxl_ip_adapter_woman.png` have been generated with refiners itself (and inspected so that they look reasonable).
 
 ## Other images