Unpin timm.

apps/stable_diffusion/shark_studio_imports.py

@@ -53,6 +53,7 @@ datas += collect_data_files("jsonschema_specifications")
 datas += collect_data_files("cpuinfo")
 datas += collect_data_files("langchain")
 datas += collect_data_files("cv2")
+datas += collect_data_files("einops")
 datas += [
     ("src/utils/resources/prompts.json", "resources"),
     ("src/utils/resources/model_db.json", "resources"),

apps/stable_diffusion/src/utils/stable_args.py

@@ -422,7 +422,7 @@ p.add_argument(
 
 p.add_argument(
     "--use_stencil",
-    choices=["canny", "openpose", "scribble"],
+    choices=["canny", "openpose", "scribble", "zoedepth"],
     help="Enable the stencil feature.",
 )
 

apps/stable_diffusion/src/utils/stencils/__init__.py

@@ -1,2 +1,3 @@
 from apps.stable_diffusion.src.utils.stencils.canny import CannyDetector
 from apps.stable_diffusion.src.utils.stencils.openpose import OpenposeDetector
+from apps.stable_diffusion.src.utils.stencils.zoe import ZoeDetector

apps/stable_diffusion/src/utils/stencils/stencil_utils.py

@@ -4,6 +4,7 @@ import torch
 from apps.stable_diffusion.src.utils.stencils import (
     CannyDetector,
     OpenposeDetector,
+    ZoeDetector,
 )
 
 stencil = {}
@@ -117,6 +118,9 @@ def controlnet_hint_conversion(
         case "scribble":
             print("Working with scribble")
             controlnet_hint = hint_scribble(image)
+        case "zoedepth":
+            print("Working with ZoeDepth")
+            controlnet_hint = hint_zoedepth(image)
         case _:
             return None
     controlnet_hint = controlnet_hint_shaping(
@@ -127,7 +131,7 @@ def controlnet_hint_conversion(
 
 stencil_to_model_id_map = {
     "canny": "lllyasviel/control_v11p_sd15_canny",
-    "depth": "lllyasviel/control_v11p_sd15_depth",
+    "zoedepth": "lllyasviel/control_v11f1p_sd15_depth",
     "hed": "lllyasviel/sd-controlnet-hed",
     "mlsd": "lllyasviel/control_v11p_sd15_mlsd",
     "normal": "lllyasviel/control_v11p_sd15_normalbae",
@@ -184,3 +188,16 @@ def hint_scribble(image: Image.Image):
         detected_map = np.zeros_like(input_image, dtype=np.uint8)
         detected_map[np.min(input_image, axis=2) < 127] = 255
         return detected_map
+
+
+# Stencil 4. Depth (Only Zoe Preprocessing)
+def hint_zoedepth(image: Image.Image):
+    with torch.no_grad():
+        input_image = np.array(image)
+
+        if not "depth" in stencil:
+            stencil["depth"] = ZoeDetector()
+
+        detected_map = stencil["depth"](input_image)
+        detected_map = HWC3(detected_map)
+        return detected_map

apps/stable_diffusion/src/utils/stencils/zoe/__init__.py

@@ -0,0 +1,58 @@
+import numpy as np
+import torch
+from pathlib import Path
+import requests
+
+
+from einops import rearrange
+
+remote_model_path = (
+    "https://huggingface.co/lllyasviel/Annotators/resolve/main/ZoeD_M12_N.pt"
+)
+
+
+class ZoeDetector:
+    def __init__(self):
+        cwd = Path.cwd()
+        ckpt_path = Path(cwd, "stencil_annotator")
+        ckpt_path.mkdir(parents=True, exist_ok=True)
+        modelpath = ckpt_path / "ZoeD_M12_N.pt"
+
+        with requests.get(remote_model_path, stream=True) as r:
+            r.raise_for_status()
+            with open(modelpath, "wb") as f:
+                for chunk in r.iter_content(chunk_size=8192):
+                    f.write(chunk)
+
+        model = torch.hub.load(
+            "monorimet/ZoeDepth:torch_update",
+            "ZoeD_N",
+            pretrained=False,
+            force_reload=False,
+        )
+        model.load_state_dict(
+            torch.load(modelpath, map_location=model.device)["model"]
+        )
+        model.eval()
+        self.model = model
+
+    def __call__(self, input_image):
+        assert input_image.ndim == 3
+        image_depth = input_image
+        with torch.no_grad():
+            image_depth = torch.from_numpy(image_depth).float()
+            image_depth = image_depth / 255.0
+            image_depth = rearrange(image_depth, "h w c -> 1 c h w")
+            depth = self.model.infer(image_depth)
+
+            depth = depth[0, 0].cpu().numpy()
+
+            vmin = np.percentile(depth, 2)
+            vmax = np.percentile(depth, 85)
+
+            depth -= vmin
+            depth /= vmax - vmin
+            depth = 1.0 - depth
+            depth_image = (depth * 255.0).clip(0, 255).astype(np.uint8)
+
+            return depth_image

apps/stable_diffusion/web/ui/img2img_ui.py

@@ -453,8 +453,13 @@ with gr.Blocks(title="Image-to-Image") as img2img_web:
                             elem_id="stencil_model",
                             label="Stencil model",
                             value="None",
-                            choices=["None", "canny", "openpose", "scribble"],
-                            allow_custom_value=True,
+                            choices=[
+                                "None",
+                                "canny",
+                                "openpose",
+                                "scribble",
+                                "zoedepth",
+                            ],
                         )
 
                     def show_canvas(choice):

requirements.txt

@@ -41,6 +41,7 @@ tiktoken # for codegen
 joblib # for langchain
 timm # for MiniGPT4
 langchain
+einops # for zoedepth
 
 # Keep PyInstaller at the end. Sometimes Windows Defender flags it but most folks can continue even if it errors
 pefile