support lcm and multi-controlnets

app.py

@@ -1,23 +1,34 @@
-import math
-import random
-
+import os
 import cv2
-import gradio as gr
+import torch
+import random
 import numpy as np
+
 import PIL
-import spaces
-import torch
-from diffusers.models import ControlNetModel
+from PIL import Image
+
+import diffusers
 from diffusers.utils import load_image
+from diffusers.models import ControlNetModel
+from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
+
+from huggingface_hub import hf_hub_download
+
 from insightface.app import FaceAnalysis
-from PIL import Image
 
-from pipeline_stable_diffusion_xl_instantid import StableDiffusionXLInstantIDPipeline
 from style_template import styles
+from pipeline_stable_diffusion_xl_instantid_full import StableDiffusionXLInstantIDPipeline, draw_kps
+from model_util import load_models_xl, get_torch_device
+from controlnet_util import openpose, get_depth_map, get_canny_image
+
+import gradio as gr
+
+import spaces
 
 # global variable
 MAX_SEED = np.iinfo(np.int32).max
-device = "cuda" if torch.cuda.is_available() else "cpu"
+device = get_torch_device()
+dtype = torch.float16 if str(device).__contains__("cuda") else torch.float32
 STYLE_NAMES = list(styles.keys())
 DEFAULT_STYLE_NAME = "Watercolor"
 
@@ -33,69 +44,120 @@ hf_hub_download(
 hf_hub_download(repo_id="InstantX/InstantID", filename="ip-adapter.bin", local_dir="./checkpoints")
 
 # Load face encoder
-app = FaceAnalysis(name="antelopev2", root="./", providers=["CPUExecutionProvider"])
+app = FaceAnalysis(
+    name="antelopev2",
+    root="./",
+    providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
+)
 app.prepare(ctx_id=0, det_size=(640, 640))
 
 # Path to InstantID models
-face_adapter = "./checkpoints/ip-adapter.bin"
-controlnet_path = "./checkpoints/ControlNetModel"
+face_adapter = f"./checkpoints/ip-adapter.bin"
+controlnet_path = f"./checkpoints/ControlNetModel"
 
-# Load pipeline
-controlnet = ControlNetModel.from_pretrained(controlnet_path, torch_dtype=torch.float16)
+# Load pipeline face ControlNetModel
+controlnet_identitynet = ControlNetModel.from_pretrained(
+    controlnet_path, torch_dtype=dtype
+)
 
-base_model_path = "wangqixun/YamerMIX_v8"
+# controlnet-pose
+controlnet_pose_model = "thibaud/controlnet-openpose-sdxl-1.0"
+controlnet_canny_model = "diffusers/controlnet-canny-sdxl-1.0"
+controlnet_depth_model = "diffusers/controlnet-depth-sdxl-1.0-small"
+
+controlnet_pose = ControlNetModel.from_pretrained(
+    controlnet_pose_model, torch_dtype=dtype
+).to(device)
+controlnet_canny = ControlNetModel.from_pretrained(
+    controlnet_canny_model, torch_dtype=dtype
+).to(device)
+controlnet_depth = ControlNetModel.from_pretrained(
+    controlnet_depth_model, torch_dtype=dtype
+).to(device)
+
+controlnet_map = {
+    "pose": controlnet_pose,
+    "canny": controlnet_canny,
+    "depth": controlnet_depth,
+}
+controlnet_map_fn = {
+    "pose": openpose,
+    "canny": get_canny_image,
+    "depth": get_depth_map,
+}
+
+pretrained_model_name_or_path = "wangqixun/YamerMIX_v8"
 
 pipe = StableDiffusionXLInstantIDPipeline.from_pretrained(
-    base_model_path,
-    controlnet=controlnet,
-    torch_dtype=torch.float16,
+    pretrained_model_name_or_path,
+    controlnet=[controlnet_identitynet],
+    torch_dtype=dtype,
     safety_checker=None,
     feature_extractor=None,
+).to(device)
+
+pipe.scheduler = diffusers.EulerDiscreteScheduler.from_config(
+    pipe.scheduler.config
 )
-pipe.cuda()
-pipe.load_ip_adapter_instantid(face_adapter)
-pipe.image_proj_model.to("cuda")
-pipe.unet.to("cuda")
 
+pipe.load_ip_adapter_instantid(face_adapter)
+# load and disable LCM
+pipe.load_lora_weights("latent-consistency/lcm-lora-sdxl")
+pipe.disable_lora()
+
+def toggle_lcm_ui(value):
+    if value:
+        return (
+            gr.update(minimum=0, maximum=100, step=1, value=5),
+            gr.update(minimum=0.1, maximum=20.0, step=0.1, value=1.5),
+        )
+    else:
+        return (
+            gr.update(minimum=5, maximum=100, step=1, value=30),
+            gr.update(minimum=0.1, maximum=20.0, step=0.1, value=5),
+        )
 
 def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
     if randomize_seed:
         seed = random.randint(0, MAX_SEED)
     return seed
 
-
 def remove_tips():
     return gr.update(visible=False)
 
-
 def get_example():
     case = [
         [
             "./examples/yann-lecun_resize.jpg",
+            None,
             "a man",
             "Snow",
             "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
         ],
         [
             "./examples/musk_resize.jpeg",
-            "a man",
+            "./examples/poses/pose2.jpg",
+            "a man flying in the sky in Mars",
             "Mars",
             "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
         ],
         [
             "./examples/sam_resize.png",
-            "a man",
+            "./examples/poses/pose4.jpg",
+            "a man doing a silly pose wearing a suite",
             "Jungle",
             "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, gree",
         ],
         [
             "./examples/schmidhuber_resize.png",
-            "a man",
+            "./examples/poses/pose3.jpg",
+            "a man sit on a chair",
             "Neon",
             "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
         ],
         [
             "./examples/kaifu_resize.png",
+            "./examples/poses/pose.jpg",
             "a man",
             "Vibrant Color",
             "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
@@ -103,50 +165,33 @@ def get_example():
     ]
     return case
 
-
-def run_for_examples(face_file, prompt, style, negative_prompt):
-    return generate_image(face_file, None, prompt, negative_prompt, style, True, 30, 0.8, 0.8, 5, 42)
-
+def run_for_examples(face_file, pose_file, prompt, style, negative_prompt):
+    return generate_image(
+        face_file,
+        pose_file,
+        prompt,
+        negative_prompt,
+        style,
+        20,  # num_steps
+        0.8,  # identitynet_strength_ratio
+        0.8,  # adapter_strength_ratio
+        0.4,  # pose_strength
+        0.3,  # canny_strength
+        0.5,  # depth_strength
+        ["pose", "canny"],  # controlnet_selection
+        5.0,  # guidance_scale
+        42,  # seed
+        "EulerDiscreteScheduler",  # scheduler
+        False,  # enable_LCM
+        True,  # enable_Face_Region
+    )
 
 def convert_from_cv2_to_image(img: np.ndarray) -> Image:
     return Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
 
-
 def convert_from_image_to_cv2(img: Image) -> np.ndarray:
     return cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR)
 
-
-def draw_kps(image_pil, kps, color_list=[(255, 0, 0), (0, 255, 0), (0, 0, 255), (255, 255, 0), (255, 0, 255)]):
-    stickwidth = 4
-    limbSeq = np.array([[0, 2], [1, 2], [3, 2], [4, 2]])
-    kps = np.array(kps)
-
-    w, h = image_pil.size
-    out_img = np.zeros([h, w, 3])
-
-    for i in range(len(limbSeq)):
-        index = limbSeq[i]
-        color = color_list[index[0]]
-
-        x = kps[index][:, 0]
-        y = kps[index][:, 1]
-        length = ((x[0] - x[1]) ** 2 + (y[0] - y[1]) ** 2) ** 0.5
-        angle = math.degrees(math.atan2(y[0] - y[1], x[0] - x[1]))
-        polygon = cv2.ellipse2Poly(
-            (int(np.mean(x)), int(np.mean(y))), (int(length / 2), stickwidth), int(angle), 0, 360, 1
-        )
-        out_img = cv2.fillConvexPoly(out_img.copy(), polygon, color)
-    out_img = (out_img * 0.6).astype(np.uint8)
-
-    for idx_kp, kp in enumerate(kps):
-        color = color_list[idx_kp]
-        x, y = kp
-        out_img = cv2.circle(out_img.copy(), (int(x), int(y)), 10, color, -1)
-
-    out_img_pil = Image.fromarray(out_img.astype(np.uint8))
-    return out_img_pil
-
-
 def resize_img(
     input_image,
     max_side=1280,
@@ -172,21 +217,18 @@ def resize_img(
         res = np.ones([max_side, max_side, 3], dtype=np.uint8) * 255
         offset_x = (max_side - w_resize_new) // 2
         offset_y = (max_side - h_resize_new) // 2
-        res[offset_y : offset_y + h_resize_new, offset_x : offset_x + w_resize_new] = np.array(input_image)
+        res[
+            offset_y : offset_y + h_resize_new, offset_x : offset_x + w_resize_new
+        ] = np.array(input_image)
         input_image = Image.fromarray(res)
     return input_image
 
-
-def apply_style(style_name: str, positive: str, negative: str = "") -> tuple[str, str]:
+def apply_style(
+    style_name: str, positive: str, negative: str = ""
+) -> tuple[str, str]:
     p, n = styles.get(style_name, styles[DEFAULT_STYLE_NAME])
     return p.replace("{prompt}", positive), n + " " + negative
 
-
-def check_input_image(face_image):
-    if face_image is None:
-        raise gr.Error("Cannot find any input face image! Please upload the face image")
-
-
 @spaces.GPU
 def generate_image(
     face_image_path,
@@ -194,14 +236,41 @@ def generate_image(
     prompt,
     negative_prompt,
     style_name,
-    enhance_face_region,
     num_steps,
     identitynet_strength_ratio,
     adapter_strength_ratio,
+    pose_strength,
+    canny_strength,
+    depth_strength,
+    controlnet_selection,
     guidance_scale,
     seed,
+    scheduler,
+    enable_LCM,
+    enhance_face_region,
     progress=gr.Progress(track_tqdm=True),
 ):
+
+    if enable_LCM:
+        pipe.scheduler = diffusers.LCMScheduler.from_config(pipe.scheduler.config)
+        pipe.enable_lora()
+    else:
+        pipe.disable_lora()
+        scheduler_class_name = scheduler.split("-")[0]
+
+        add_kwargs = {}
+        if len(scheduler.split("-")) > 1:
+            add_kwargs["use_karras_sigmas"] = True
+        if len(scheduler.split("-")) > 2:
+            add_kwargs["algorithm_type"] = "sde-dpmsolver++"
+        scheduler = getattr(diffusers, scheduler_class_name)
+        pipe.scheduler = scheduler.from_config(pipe.scheduler.config, **add_kwargs)
+
+    if face_image_path is None:
+        raise gr.Error(
+            f"Cannot find any input face image! Please upload the face image"
+        )
+
     if prompt is None:
         prompt = "a person"
 
@@ -209,7 +278,7 @@ def generate_image(
     prompt, negative_prompt = apply_style(style_name, prompt, negative_prompt)
 
     face_image = load_image(face_image_path)
-    face_image = resize_img(face_image)
+    face_image = resize_img(face_image, max_side=1024)
     face_image_cv2 = convert_from_image_to_cv2(face_image)
     height, width, _ = face_image_cv2.shape
 
@@ -217,23 +286,31 @@ def generate_image(
     face_info = app.get(face_image_cv2)
 
     if len(face_info) == 0:
-        raise gr.Error("Cannot find any face in the image! Please upload another person image")
+        raise gr.Error(
+            f"Unable to detect a face in the image. Please upload a different photo with a clear face."
+        )
 
-    face_info = sorted(face_info, key=lambda x: (x["bbox"][2] - x["bbox"][0]) * x["bbox"][3] - x["bbox"][1])[
+    face_info = sorted(
+        face_info,
+        key=lambda x: (x["bbox"][2] - x["bbox"][0]) * x["bbox"][3] - x["bbox"][1],
+    )[
         -1
     ]  # only use the maximum face
     face_emb = face_info["embedding"]
     face_kps = draw_kps(convert_from_cv2_to_image(face_image_cv2), face_info["kps"])
-
+    img_controlnet = face_image
     if pose_image_path is not None:
         pose_image = load_image(pose_image_path)
-        pose_image = resize_img(pose_image)
+        pose_image = resize_img(pose_image, max_side=1024)
+        img_controlnet = pose_image
         pose_image_cv2 = convert_from_image_to_cv2(pose_image)
 
         face_info = app.get(pose_image_cv2)
 
         if len(face_info) == 0:
-            raise gr.Error("Cannot find any face in the reference image! Please upload another person image")
+            raise gr.Error(
+                f"Cannot find any face in the reference image! Please upload another person image"
+            )
 
         face_info = face_info[-1]
         face_kps = draw_kps(pose_image, face_info["kps"])
@@ -249,6 +326,28 @@ def generate_image(
     else:
         control_mask = None
 
+    if len(controlnet_selection) > 0:
+        controlnet_scales = {
+            "pose": pose_strength,
+            "canny": canny_strength,
+            "depth": depth_strength,
+        }
+        pipe.controlnet = MultiControlNetModel(
+            [controlnet_identitynet]
+            + [controlnet_map[s] for s in controlnet_selection]
+        )
+        control_scales = [float(identitynet_strength_ratio)] + [
+            controlnet_scales[s] for s in controlnet_selection
+        ]
+        control_images = [face_kps] + [
+            controlnet_map_fn[s](img_controlnet).resize((width, height))
+            for s in controlnet_selection
+        ]
+    else:
+        pipe.controlnet = controlnet_identitynet
+        control_scales = float(identitynet_strength_ratio)
+        control_images = face_kps
+
     generator = torch.Generator(device=device).manual_seed(seed)
 
     print("Start inference...")
@@ -259,9 +358,9 @@ def generate_image(
         prompt=prompt,
         negative_prompt=negative_prompt,
         image_embeds=face_emb,
-        image=face_kps,
+        image=control_images,
         control_mask=control_mask,
-        controlnet_conditioning_scale=float(identitynet_strength_ratio),
+        controlnet_conditioning_scale=control_scales,
         num_inference_steps=num_steps,
         guidance_scale=guidance_scale,
         height=height,
@@ -271,8 +370,7 @@ def generate_image(
 
     return images[0], gr.update(visible=True)
 
-
-### Description
+# Description
 title = r"""
 <h1 align="center">InstantID: Zero-shot Identity-Preserving Generation in Seconds</h1>
 """
@@ -281,12 +379,12 @@ description = r"""
 <b>Official 🤗 Gradio demo</b> for <a href='https://github.com/InstantID/InstantID' target='_blank'><b>InstantID: Zero-shot Identity-Preserving Generation in Seconds</b></a>.<br>
 
 How to use:<br>
-1. Upload a person image. For multiple person images, we will only detect the biggest face. Make sure face is not too small and not significantly blocked or blurred.
-2. (Optionally) upload another person image as reference pose. If not uploaded, we will use the first person image to extract landmarks. If you use a cropped face at step1, it is recommeneded to upload it to extract a new pose.
-3. Enter a text prompt as done in normal text-to-image models.
-4. Click the <b>Submit</b> button to start customizing.
-5. Share your customizd photo with your friends, enjoy😊!
-"""
+1. Upload an image with a face. For images with multiple faces, we will only detect the largest face. Ensure the face is not too small and is clearly visible without significant obstructions or blurring.
+2. (Optional) You can upload another image as a reference for the face pose. If you don't, we will use the first detected face image to extract facial landmarks. If you use a cropped face at step 1, it is recommended to upload it to define a new face pose.
+3. (Optional) You can select multiple ControlNet models to control the generation process. The default is to use the IdentityNet only. The ControlNet models include pose skeleton, canny, and depth. You can adjust the strength of each ControlNet model to control the generation process.
+4. Enter a text prompt, as done in normal text-to-image models.
+5. Click the <b>Submit</b> button to begin customization.
+6. Share your customized photo with your friends and enjoy! 😊"""
 
 article = r"""
 ---
@@ -295,10 +393,10 @@ article = r"""
 If our work is helpful for your research or applications, please cite us via:
 ```bibtex
 @article{wang2024instantid,
-  title={InstantID: Zero-shot Identity-Preserving Generation in Seconds},
-  author={Wang, Qixun and Bai, Xu and Wang, Haofan and Qin, Zekui and Chen, Anthony},
-  journal={arXiv preprint arXiv:2401.07519},
-  year={2024}
+title={InstantID: Zero-shot Identity-Preserving Generation in Seconds},
+author={Wang, Qixun and Bai, Xu and Wang, Haofan and Qin, Zekui and Chen, Anthony},
+journal={arXiv preprint arXiv:2401.07519},
+year={2024}
 }
 ```
 📧 **Contact**
@@ -308,10 +406,10 @@ If you have any questions, please feel free to open an issue or directly reach u
 
 tips = r"""
 ### Usage tips of InstantID
-1. If you're unsatisfied with the similarity, increase the weight of controlnet_conditioning_scale (IdentityNet) and ip_adapter_scale (Adapter).
-2. If the generated image is over-saturated, decrease the ip_adapter_scale. If not work, decrease controlnet_conditioning_scale.
-3. If text control is not as expected, decrease ip_adapter_scale.
-4. Find a good base model always makes a difference.
+1. If you're not satisfied with the similarity, try increasing the weight of "IdentityNet Strength" and "Adapter Strength."    
+2. If you feel that the saturation is too high, first decrease the Adapter strength. If it remains too high, then decrease the IdentityNet strength.
+3. If you find that text control is not as expected, decrease Adapter strength.
+4. If you find that realistic style is not good enough, go for our Github repo and use a more realistic base model.
 """
 
 css = """
@@ -324,27 +422,39 @@ with gr.Blocks(css=css) as demo:
 
     with gr.Row():
         with gr.Column():
-            # upload face image
-            face_file = gr.Image(label="Upload a photo of your face", type="filepath")
-
-            # optional: upload a reference pose image
-            pose_file = gr.Image(label="Upload a reference pose image (optional)", type="filepath")
+            with gr.Row(equal_height=True):
+                # upload face image
+                face_file = gr.Image(
+                    label="Upload a photo of your face", type="filepath"
+                )
+                # optional: upload a reference pose image
+                pose_file = gr.Image(
+                    label="Upload a reference pose image (Optional)",
+                    type="filepath",
+                )
 
             # prompt
             prompt = gr.Textbox(
                 label="Prompt",
-                info="Give simple prompt is enough to achieve good face fedility",
+                info="Give simple prompt is enough to achieve good face fidelity",
                 placeholder="A photo of a person",
                 value="",
             )
 
             submit = gr.Button("Submit", variant="primary")
-
-            style = gr.Dropdown(label="Style template", choices=STYLE_NAMES, value=DEFAULT_STYLE_NAME)
+            enable_LCM = gr.Checkbox(
+                label="Enable Fast Inference with LCM", value=enable_lcm_arg,
+                info="LCM speeds up the inference step, the trade-off is the quality of the generated image. It performs better with portrait face images rather than distant faces",
+            )
+            style = gr.Dropdown(
+                label="Style template",
+                choices=STYLE_NAMES,
+                value=DEFAULT_STYLE_NAME,
+            )
 
             # strength
             identitynet_strength_ratio = gr.Slider(
-                label="IdentityNet strength (for fedility)",
+                label="IdentityNet strength (for fidelity)",
                 minimum=0,
                 maximum=1.5,
                 step=0.05,
@@ -357,26 +467,51 @@ with gr.Blocks(css=css) as demo:
                 step=0.05,
                 value=0.80,
             )
-
+            with gr.Accordion("Controlnet"):
+                controlnet_selection = gr.CheckboxGroup(
+                    ["pose", "canny", "depth"], label="Controlnet", value=["pose"],
+                    info="Use pose for skeleton inference, canny for edge detection, and depth for depth map estimation. You can try all three to control the generation process"
+                )
+                pose_strength = gr.Slider(
+                    label="Pose strength",
+                    minimum=0,
+                    maximum=1.5,
+                    step=0.05,
+                    value=0.40,
+                )
+                canny_strength = gr.Slider(
+                    label="Canny strength",
+                    minimum=0,
+                    maximum=1.5,
+                    step=0.05,
+                    value=0.40,
+                )
+                depth_strength = gr.Slider(
+                    label="Depth strength",
+                    minimum=0,
+                    maximum=1.5,
+                    step=0.05,
+                    value=0.40,
+                )
             with gr.Accordion(open=False, label="Advanced Options"):
                 negative_prompt = gr.Textbox(
                     label="Negative Prompt",
                     placeholder="low quality",
-                    value="(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, nudity,naked, bikini, skimpy, scanty, bare skin, lingerie, swimsuit, exposed, see-through, photo, anthropomorphic cat, monochrome, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
+                    value="(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green",
                 )
                 num_steps = gr.Slider(
                     label="Number of sample steps",
-                    minimum=20,
+                    minimum=1,
                     maximum=100,
                     step=1,
-                    value=30,
+                    value=5 if enable_lcm_arg else 30,
                 )
                 guidance_scale = gr.Slider(
                     label="Guidance scale",
                     minimum=0.1,
-                    maximum=10.0,
+                    maximum=20.0,
                     step=0.1,
-                    value=5,
+                    value=0.0 if enable_lcm_arg else 5.0,
                 )
                 seed = gr.Slider(
                     label="Seed",
@@ -385,18 +520,31 @@ with gr.Blocks(css=css) as demo:
                     step=1,
                     value=42,
                 )
+                schedulers = [
+                    "DEISMultistepScheduler",
+                    "HeunDiscreteScheduler",
+                    "EulerDiscreteScheduler",
+                    "DPMSolverMultistepScheduler",
+                    "DPMSolverMultistepScheduler-Karras",
+                    "DPMSolverMultistepScheduler-Karras-SDE",
+                ]
+                scheduler = gr.Dropdown(
+                    label="Schedulers",
+                    choices=schedulers,
+                    value="EulerDiscreteScheduler",
+                )
                 randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
                 enhance_face_region = gr.Checkbox(label="Enhance non-face region", value=True)
 
-        with gr.Column():
-            output_image = gr.Image(label="Generated Image")
-            usage_tips = gr.Markdown(label="Usage tips of InstantID", value=tips, visible=False)
+        with gr.Column(scale=1):
+            gallery = gr.Image(label="Generated Images")
+            usage_tips = gr.Markdown(
+                label="InstantID Usage Tips", value=tips, visible=False
+            )
 
         submit.click(
             fn=remove_tips,
             outputs=usage_tips,
-            queue=False,
-            api_name=False,
         ).then(
             fn=randomize_seed_fn,
             inputs=[seed, randomize_seed],
@@ -404,11 +552,6 @@ with gr.Blocks(css=css) as demo:
             queue=False,
             api_name=False,
         ).then(
-            fn=check_input_image,
-            inputs=face_file,
-            queue=False,
-            api_name=False,
-        ).success(
             fn=generate_image,
             inputs=[
                 face_file,
@@ -416,21 +559,34 @@ with gr.Blocks(css=css) as demo:
                 prompt,
                 negative_prompt,
                 style,
-                enhance_face_region,
                 num_steps,
                 identitynet_strength_ratio,
                 adapter_strength_ratio,
+                pose_strength,
+                canny_strength,
+                depth_strength,
+                controlnet_selection,
                 guidance_scale,
                 seed,
+                scheduler,
+                enable_LCM,
+                enhance_face_region,
             ],
-            outputs=[output_image, usage_tips],
+            outputs=[gallery, usage_tips],
+        )
+
+        enable_LCM.input(
+            fn=toggle_lcm_ui,
+            inputs=[enable_LCM],
+            outputs=[num_steps, guidance_scale],
+            queue=False,
         )
 
     gr.Examples(
         examples=get_example(),
-        inputs=[face_file, prompt, style, negative_prompt],
-        outputs=[output_image, usage_tips],
+        inputs=[face_file, pose_file, prompt, style, negative_prompt],
         fn=run_for_examples,
+        outputs=[gallery, usage_tips],
         cache_examples=True,
     )
 

controlnet_util.py

@@ -0,0 +1,39 @@
+import torch
+import numpy as np
+from PIL import Image
+from controlnet_aux import OpenposeDetector
+from model_util import get_torch_device
+import cv2
+
+
+from transformers import DPTImageProcessor, DPTForDepthEstimation
+
+device = get_torch_device()
+depth_estimator = DPTForDepthEstimation.from_pretrained("Intel/dpt-hybrid-midas").to(device)
+feature_extractor = DPTImageProcessor.from_pretrained("Intel/dpt-hybrid-midas")
+openpose = OpenposeDetector.from_pretrained("lllyasviel/ControlNet")
+
+def get_depth_map(image):
+    image = feature_extractor(images=image, return_tensors="pt").pixel_values.to("cuda")
+    with torch.no_grad(), torch.autocast("cuda"):
+        depth_map = depth_estimator(image).predicted_depth
+
+    depth_map = torch.nn.functional.interpolate(
+        depth_map.unsqueeze(1),
+        size=(1024, 1024),
+        mode="bicubic",
+        align_corners=False,
+    )
+    depth_min = torch.amin(depth_map, dim=[1, 2, 3], keepdim=True)
+    depth_max = torch.amax(depth_map, dim=[1, 2, 3], keepdim=True)
+    depth_map = (depth_map - depth_min) / (depth_max - depth_min)
+    image = torch.cat([depth_map] * 3, dim=1)
+
+    image = image.permute(0, 2, 3, 1).cpu().numpy()[0]
+    image = Image.fromarray((image * 255.0).clip(0, 255).astype(np.uint8))
+    return image
+
+def get_canny_image(image, t1=100, t2=200):
+    image = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR)
+    edges = cv2.Canny(image, t1, t2)
+    return Image.fromarray(edges, "L")

gradio_cached_examples/25/log.csv

@@ -1,6 +0,0 @@
-Generated Image,Usage tips of InstantID,flag,username,timestamp
-"{""path"":""gradio_cached_examples/25/Generated Image/7e6ea85f77dd925d842c/image.png"",""url"":null,""size"":null,""orig_name"":""image.png"",""mime_type"":null}","{'visible': True, '__type__': 'update'}",,,2024-01-24 08:55:38.846769
-"{""path"":""gradio_cached_examples/25/Generated Image/2880a3d19ef9b42e3ed2/image.png"",""url"":null,""size"":null,""orig_name"":""image.png"",""mime_type"":null}","{'visible': True, '__type__': 'update'}",,,2024-01-24 08:56:11.432078
-"{""path"":""gradio_cached_examples/25/Generated Image/38dde1388c41109c5d39/image.png"",""url"":null,""size"":null,""orig_name"":""image.png"",""mime_type"":null}","{'visible': True, '__type__': 'update'}",,,2024-01-24 08:56:45.563918
-"{""path"":""gradio_cached_examples/25/Generated Image/e56b029833685ff77e6a/image.png"",""url"":null,""size"":null,""orig_name"":""image.png"",""mime_type"":null}","{'visible': True, '__type__': 'update'}",,,2024-01-24 08:57:20.321876
-"{""path"":""gradio_cached_examples/25/Generated Image/6cb5a3af223906666bfd/image.png"",""url"":null,""size"":null,""orig_name"":""image.png"",""mime_type"":null}","{'visible': True, '__type__': 'update'}",,,2024-01-24 08:57:53.871716

ip_adapter/attention_processor.py

@@ -10,14 +10,11 @@ try:
 except Exception as e:
     xformers_available = False
 
-
-
 class RegionControler(object):
     def __init__(self) -> None:
         self.prompt_image_conditioning = []
 region_control = RegionControler()
 
-
 class AttnProcessor(nn.Module):
     r"""
     Default processor for performing attention-related computations.
@@ -29,7 +26,7 @@ class AttnProcessor(nn.Module):
     ):
         super().__init__()
 
-    def __call__(
+    def forward(
         self,
         attn,
         hidden_states,
@@ -115,7 +112,7 @@ class IPAttnProcessor(nn.Module):
         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__(
+    def forward(
         self,
         attn,
         hidden_states,
@@ -180,7 +177,7 @@ class IPAttnProcessor(nn.Module):
             ip_attention_probs = attn.get_attention_scores(query, ip_key, None)
             ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)
         ip_hidden_states = attn.batch_to_head_dim(ip_hidden_states)
-        
+
         # region control
         if len(region_control.prompt_image_conditioning) == 1:
             region_mask = region_control.prompt_image_conditioning[0].get('region_mask', None)
@@ -190,7 +187,7 @@ class IPAttnProcessor(nn.Module):
                 mask = F.interpolate(region_mask[None, None], scale_factor=1/ratio, mode='nearest').reshape([1, -1, 1])
             else:
                 mask = torch.ones_like(ip_hidden_states)
-            ip_hidden_states = ip_hidden_states * mask            
+            ip_hidden_states = ip_hidden_states * mask     
 
         hidden_states = hidden_states + self.scale * ip_hidden_states
 
@@ -233,7 +230,7 @@ class AttnProcessor2_0(torch.nn.Module):
         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.")
 
-    def __call__(
+    def forward(
         self,
         attn,
         hidden_states,
@@ -305,4 +302,145 @@ class AttnProcessor2_0(torch.nn.Module):
 
         hidden_states = hidden_states / attn.rescale_output_factor
 
+        return hidden_states
+
+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.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
+        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.num_tokens = num_tokens
+
+        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 forward(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        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
+        else:
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states, ip_hidden_states = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:, :],
+            )
+            if 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)
+
+        # for ip-adapter
+        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
+        )
+        with torch.no_grad():
+            self.attn_map = query @ ip_key.transpose(-2, -1).softmax(dim=-1)
+            #print(self.attn_map.shape)
+
+        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)
+
+        # region control
+        if len(region_control.prompt_image_conditioning) == 1:
+            region_mask = region_control.prompt_image_conditioning[0].get('region_mask', None)
+            if region_mask is not None:
+                h, w = region_mask.shape[:2]
+                ratio = (h * w / query.shape[1]) ** 0.5
+                mask = F.interpolate(region_mask[None, None], scale_factor=1/ratio, mode='nearest').reshape([1, -1, 1])
+            else:
+                mask = torch.ones_like(ip_hidden_states)
+            ip_hidden_states = ip_hidden_states * mask
+
+        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

model_util.py

@@ -0,0 +1,472 @@
+from typing import Literal, Union, Optional, Tuple, List
+
+import torch
+from transformers import CLIPTextModel, CLIPTokenizer, CLIPTextModelWithProjection
+from diffusers import (
+    UNet2DConditionModel,
+    SchedulerMixin,
+    StableDiffusionPipeline,
+    StableDiffusionXLPipeline,
+    AutoencoderKL,
+)
+from diffusers.pipelines.stable_diffusion.convert_from_ckpt import (
+    convert_ldm_unet_checkpoint,
+)
+from safetensors.torch import load_file
+from diffusers.schedulers import (
+    DDIMScheduler,
+    DDPMScheduler,
+    LMSDiscreteScheduler,
+    EulerDiscreteScheduler,
+    EulerAncestralDiscreteScheduler,
+    UniPCMultistepScheduler,
+)
+
+from omegaconf import OmegaConf
+
+# DiffUsers版StableDiffusionのモデルパラメータ
+NUM_TRAIN_TIMESTEPS = 1000
+BETA_START = 0.00085
+BETA_END = 0.0120
+
+UNET_PARAMS_MODEL_CHANNELS = 320
+UNET_PARAMS_CHANNEL_MULT = [1, 2, 4, 4]
+UNET_PARAMS_ATTENTION_RESOLUTIONS = [4, 2, 1]
+UNET_PARAMS_IMAGE_SIZE = 64  # fixed from old invalid value `32`
+UNET_PARAMS_IN_CHANNELS = 4
+UNET_PARAMS_OUT_CHANNELS = 4
+UNET_PARAMS_NUM_RES_BLOCKS = 2
+UNET_PARAMS_CONTEXT_DIM = 768
+UNET_PARAMS_NUM_HEADS = 8
+# UNET_PARAMS_USE_LINEAR_PROJECTION = False
+
+VAE_PARAMS_Z_CHANNELS = 4
+VAE_PARAMS_RESOLUTION = 256
+VAE_PARAMS_IN_CHANNELS = 3
+VAE_PARAMS_OUT_CH = 3
+VAE_PARAMS_CH = 128
+VAE_PARAMS_CH_MULT = [1, 2, 4, 4]
+VAE_PARAMS_NUM_RES_BLOCKS = 2
+
+# V2
+V2_UNET_PARAMS_ATTENTION_HEAD_DIM = [5, 10, 20, 20]
+V2_UNET_PARAMS_CONTEXT_DIM = 1024
+# V2_UNET_PARAMS_USE_LINEAR_PROJECTION = True
+
+TOKENIZER_V1_MODEL_NAME = "CompVis/stable-diffusion-v1-4"
+TOKENIZER_V2_MODEL_NAME = "stabilityai/stable-diffusion-2-1"
+
+AVAILABLE_SCHEDULERS = Literal["ddim", "ddpm", "lms", "euler_a", "euler", "uniPC"]
+
+SDXL_TEXT_ENCODER_TYPE = Union[CLIPTextModel, CLIPTextModelWithProjection]
+
+DIFFUSERS_CACHE_DIR = None  # if you want to change the cache dir, change this
+
+
+def load_checkpoint_with_text_encoder_conversion(ckpt_path: str, device="cpu"):
+    # text encoderの格納形式が違うモデルに対応する ('text_model'がない)
+    TEXT_ENCODER_KEY_REPLACEMENTS = [
+        (
+            "cond_stage_model.transformer.embeddings.",
+            "cond_stage_model.transformer.text_model.embeddings.",
+        ),
+        (
+            "cond_stage_model.transformer.encoder.",
+            "cond_stage_model.transformer.text_model.encoder.",
+        ),
+        (
+            "cond_stage_model.transformer.final_layer_norm.",
+            "cond_stage_model.transformer.text_model.final_layer_norm.",
+        ),
+    ]
+
+    if ckpt_path.endswith(".safetensors"):
+        checkpoint = None
+        state_dict = load_file(ckpt_path)  # , device) # may causes error
+    else:
+        checkpoint = torch.load(ckpt_path, map_location=device)
+        if "state_dict" in checkpoint:
+            state_dict = checkpoint["state_dict"]
+        else:
+            state_dict = checkpoint
+            checkpoint = None
+
+    key_reps = []
+    for rep_from, rep_to in TEXT_ENCODER_KEY_REPLACEMENTS:
+        for key in state_dict.keys():
+            if key.startswith(rep_from):
+                new_key = rep_to + key[len(rep_from) :]
+                key_reps.append((key, new_key))
+
+    for key, new_key in key_reps:
+        state_dict[new_key] = state_dict[key]
+        del state_dict[key]
+
+    return checkpoint, state_dict
+
+
+def create_unet_diffusers_config(v2, use_linear_projection_in_v2=False):
+    """
+    Creates a config for the diffusers based on the config of the LDM model.
+    """
+    # unet_params = original_config.model.params.unet_config.params
+
+    block_out_channels = [
+        UNET_PARAMS_MODEL_CHANNELS * mult for mult in UNET_PARAMS_CHANNEL_MULT
+    ]
+
+    down_block_types = []
+    resolution = 1
+    for i in range(len(block_out_channels)):
+        block_type = (
+            "CrossAttnDownBlock2D"
+            if resolution in UNET_PARAMS_ATTENTION_RESOLUTIONS
+            else "DownBlock2D"
+        )
+        down_block_types.append(block_type)
+        if i != len(block_out_channels) - 1:
+            resolution *= 2
+
+    up_block_types = []
+    for i in range(len(block_out_channels)):
+        block_type = (
+            "CrossAttnUpBlock2D"
+            if resolution in UNET_PARAMS_ATTENTION_RESOLUTIONS
+            else "UpBlock2D"
+        )
+        up_block_types.append(block_type)
+        resolution //= 2
+
+    config = dict(
+        sample_size=UNET_PARAMS_IMAGE_SIZE,
+        in_channels=UNET_PARAMS_IN_CHANNELS,
+        out_channels=UNET_PARAMS_OUT_CHANNELS,
+        down_block_types=tuple(down_block_types),
+        up_block_types=tuple(up_block_types),
+        block_out_channels=tuple(block_out_channels),
+        layers_per_block=UNET_PARAMS_NUM_RES_BLOCKS,
+        cross_attention_dim=UNET_PARAMS_CONTEXT_DIM
+        if not v2
+        else V2_UNET_PARAMS_CONTEXT_DIM,
+        attention_head_dim=UNET_PARAMS_NUM_HEADS
+        if not v2
+        else V2_UNET_PARAMS_ATTENTION_HEAD_DIM,
+        # use_linear_projection=UNET_PARAMS_USE_LINEAR_PROJECTION if not v2 else V2_UNET_PARAMS_USE_LINEAR_PROJECTION,
+    )
+    if v2 and use_linear_projection_in_v2:
+        config["use_linear_projection"] = True
+
+    return config
+
+
+def load_diffusers_model(
+    pretrained_model_name_or_path: str,
+    v2: bool = False,
+    clip_skip: Optional[int] = None,
+    weight_dtype: torch.dtype = torch.float32,
+) -> Tuple[CLIPTokenizer, CLIPTextModel, UNet2DConditionModel,]:
+    if v2:
+        tokenizer = CLIPTokenizer.from_pretrained(
+            TOKENIZER_V2_MODEL_NAME,
+            subfolder="tokenizer",
+            torch_dtype=weight_dtype,
+            cache_dir=DIFFUSERS_CACHE_DIR,
+        )
+        text_encoder = CLIPTextModel.from_pretrained(
+            pretrained_model_name_or_path,
+            subfolder="text_encoder",
+            # default is clip skip 2
+            num_hidden_layers=24 - (clip_skip - 1) if clip_skip is not None else 23,
+            torch_dtype=weight_dtype,
+            cache_dir=DIFFUSERS_CACHE_DIR,
+        )
+    else:
+        tokenizer = CLIPTokenizer.from_pretrained(
+            TOKENIZER_V1_MODEL_NAME,
+            subfolder="tokenizer",
+            torch_dtype=weight_dtype,
+            cache_dir=DIFFUSERS_CACHE_DIR,
+        )
+        text_encoder = CLIPTextModel.from_pretrained(
+            pretrained_model_name_or_path,
+            subfolder="text_encoder",
+            num_hidden_layers=12 - (clip_skip - 1) if clip_skip is not None else 12,
+            torch_dtype=weight_dtype,
+            cache_dir=DIFFUSERS_CACHE_DIR,
+        )
+
+    unet = UNet2DConditionModel.from_pretrained(
+        pretrained_model_name_or_path,
+        subfolder="unet",
+        torch_dtype=weight_dtype,
+        cache_dir=DIFFUSERS_CACHE_DIR,
+    )
+
+    vae = AutoencoderKL.from_pretrained(pretrained_model_name_or_path, subfolder="vae")
+
+    return tokenizer, text_encoder, unet, vae
+
+
+def load_checkpoint_model(
+    checkpoint_path: str,
+    v2: bool = False,
+    clip_skip: Optional[int] = None,
+    weight_dtype: torch.dtype = torch.float32,
+) -> Tuple[CLIPTokenizer, CLIPTextModel, UNet2DConditionModel,]:
+    pipe = StableDiffusionPipeline.from_single_file(
+        checkpoint_path,
+        upcast_attention=True if v2 else False,
+        torch_dtype=weight_dtype,
+        cache_dir=DIFFUSERS_CACHE_DIR,
+    )
+
+    _, state_dict = load_checkpoint_with_text_encoder_conversion(checkpoint_path)
+    unet_config = create_unet_diffusers_config(v2, use_linear_projection_in_v2=v2)
+    unet_config["class_embed_type"] = None
+    unet_config["addition_embed_type"] = None
+    converted_unet_checkpoint = convert_ldm_unet_checkpoint(state_dict, unet_config)
+    unet = UNet2DConditionModel(**unet_config)
+    unet.load_state_dict(converted_unet_checkpoint)
+
+    tokenizer = pipe.tokenizer
+    text_encoder = pipe.text_encoder
+    vae = pipe.vae
+    if clip_skip is not None:
+        if v2:
+            text_encoder.config.num_hidden_layers = 24 - (clip_skip - 1)
+        else:
+            text_encoder.config.num_hidden_layers = 12 - (clip_skip - 1)
+
+    del pipe
+
+    return tokenizer, text_encoder, unet, vae
+
+
+def load_models(
+    pretrained_model_name_or_path: str,
+    scheduler_name: str,
+    v2: bool = False,
+    v_pred: bool = False,
+    weight_dtype: torch.dtype = torch.float32,
+) -> Tuple[CLIPTokenizer, CLIPTextModel, UNet2DConditionModel, SchedulerMixin,]:
+    if pretrained_model_name_or_path.endswith(
+        ".ckpt"
+    ) or pretrained_model_name_or_path.endswith(".safetensors"):
+        tokenizer, text_encoder, unet, vae = load_checkpoint_model(
+            pretrained_model_name_or_path, v2=v2, weight_dtype=weight_dtype
+        )
+    else:  # diffusers
+        tokenizer, text_encoder, unet, vae = load_diffusers_model(
+            pretrained_model_name_or_path, v2=v2, weight_dtype=weight_dtype
+        )
+
+    if scheduler_name:
+        scheduler = create_noise_scheduler(
+            scheduler_name,
+            prediction_type="v_prediction" if v_pred else "epsilon",
+        )
+    else:
+        scheduler = None
+
+    return tokenizer, text_encoder, unet, scheduler, vae
+
+
+def load_diffusers_model_xl(
+    pretrained_model_name_or_path: str,
+    weight_dtype: torch.dtype = torch.float32,
+) -> Tuple[List[CLIPTokenizer], List[SDXL_TEXT_ENCODER_TYPE], UNet2DConditionModel,]:
+    # returns tokenizer, tokenizer_2, text_encoder, text_encoder_2, unet
+
+    tokenizers = [
+        CLIPTokenizer.from_pretrained(
+            pretrained_model_name_or_path,
+            subfolder="tokenizer",
+            torch_dtype=weight_dtype,
+            cache_dir=DIFFUSERS_CACHE_DIR,
+        ),
+        CLIPTokenizer.from_pretrained(
+            pretrained_model_name_or_path,
+            subfolder="tokenizer_2",
+            torch_dtype=weight_dtype,
+            cache_dir=DIFFUSERS_CACHE_DIR,
+            pad_token_id=0,  # same as open clip
+        ),
+    ]
+
+    text_encoders = [
+        CLIPTextModel.from_pretrained(
+            pretrained_model_name_or_path,
+            subfolder="text_encoder",
+            torch_dtype=weight_dtype,
+            cache_dir=DIFFUSERS_CACHE_DIR,
+        ),
+        CLIPTextModelWithProjection.from_pretrained(
+            pretrained_model_name_or_path,
+            subfolder="text_encoder_2",
+            torch_dtype=weight_dtype,
+            cache_dir=DIFFUSERS_CACHE_DIR,
+        ),
+    ]
+
+    unet = UNet2DConditionModel.from_pretrained(
+        pretrained_model_name_or_path,
+        subfolder="unet",
+        torch_dtype=weight_dtype,
+        cache_dir=DIFFUSERS_CACHE_DIR,
+    )
+    vae = AutoencoderKL.from_pretrained(pretrained_model_name_or_path, subfolder="vae")
+    return tokenizers, text_encoders, unet, vae
+
+
+def load_checkpoint_model_xl(
+    checkpoint_path: str,
+    weight_dtype: torch.dtype = torch.float32,
+) -> Tuple[List[CLIPTokenizer], List[SDXL_TEXT_ENCODER_TYPE], UNet2DConditionModel,]:
+    pipe = StableDiffusionXLPipeline.from_single_file(
+        checkpoint_path,
+        torch_dtype=weight_dtype,
+        cache_dir=DIFFUSERS_CACHE_DIR,
+    )
+
+    unet = pipe.unet
+    vae = pipe.vae
+    tokenizers = [pipe.tokenizer, pipe.tokenizer_2]
+    text_encoders = [pipe.text_encoder, pipe.text_encoder_2]
+    if len(text_encoders) == 2:
+        text_encoders[1].pad_token_id = 0
+
+    del pipe
+
+    return tokenizers, text_encoders, unet, vae
+
+
+def load_models_xl(
+    pretrained_model_name_or_path: str,
+    scheduler_name: str,
+    weight_dtype: torch.dtype = torch.float32,
+    noise_scheduler_kwargs=None,
+) -> Tuple[
+    List[CLIPTokenizer],
+    List[SDXL_TEXT_ENCODER_TYPE],
+    UNet2DConditionModel,
+    SchedulerMixin,
+]:
+    if pretrained_model_name_or_path.endswith(
+        ".ckpt"
+    ) or pretrained_model_name_or_path.endswith(".safetensors"):
+        (tokenizers, text_encoders, unet, vae) = load_checkpoint_model_xl(
+            pretrained_model_name_or_path, weight_dtype
+        )
+    else:  # diffusers
+        (tokenizers, text_encoders, unet, vae) = load_diffusers_model_xl(
+            pretrained_model_name_or_path, weight_dtype
+        )
+    if scheduler_name:
+        scheduler = create_noise_scheduler(scheduler_name, noise_scheduler_kwargs)
+    else:
+        scheduler = None
+
+    return tokenizers, text_encoders, unet, scheduler, vae
+
+def create_noise_scheduler(
+    scheduler_name: AVAILABLE_SCHEDULERS = "ddpm",
+    noise_scheduler_kwargs=None,
+    prediction_type: Literal["epsilon", "v_prediction"] = "epsilon",
+) -> SchedulerMixin:
+    name = scheduler_name.lower().replace(" ", "_")
+    if name.lower() == "ddim":
+        # https://huggingface.co/docs/diffusers/v0.17.1/en/api/schedulers/ddim
+        scheduler = DDIMScheduler(**OmegaConf.to_container(noise_scheduler_kwargs))
+    elif name.lower() == "ddpm":
+        # https://huggingface.co/docs/diffusers/v0.17.1/en/api/schedulers/ddpm
+        scheduler = DDPMScheduler(**OmegaConf.to_container(noise_scheduler_kwargs))
+    elif name.lower() == "lms":
+        # https://huggingface.co/docs/diffusers/v0.17.1/en/api/schedulers/lms_discrete
+        scheduler = LMSDiscreteScheduler(
+            **OmegaConf.to_container(noise_scheduler_kwargs)
+        )
+    elif name.lower() == "euler_a":
+        # https://huggingface.co/docs/diffusers/v0.17.1/en/api/schedulers/euler_ancestral
+        scheduler = EulerAncestralDiscreteScheduler(
+            **OmegaConf.to_container(noise_scheduler_kwargs)
+        )
+    elif name.lower() == "euler":
+        # https://huggingface.co/docs/diffusers/v0.17.1/en/api/schedulers/euler_ancestral
+        scheduler = EulerDiscreteScheduler(
+            **OmegaConf.to_container(noise_scheduler_kwargs)
+        )
+    elif name.lower() == "unipc":
+        # https://huggingface.co/docs/diffusers/v0.17.1/en/api/schedulers/unipc
+        scheduler = UniPCMultistepScheduler(
+            **OmegaConf.to_container(noise_scheduler_kwargs)
+        )
+    else:
+        raise ValueError(f"Unknown scheduler name: {name}")
+
+    return scheduler
+
+
+def torch_gc():
+    import gc
+
+    gc.collect()
+    if torch.cuda.is_available():
+        with torch.cuda.device("cuda"):
+            torch.cuda.empty_cache()
+            torch.cuda.ipc_collect()
+
+
+from enum import Enum
+
+
+class CPUState(Enum):
+    GPU = 0
+    CPU = 1
+    MPS = 2
+
+
+cpu_state = CPUState.GPU
+xpu_available = False
+directml_enabled = False
+
+
+def is_intel_xpu():
+    global cpu_state
+    global xpu_available
+    if cpu_state == CPUState.GPU:
+        if xpu_available:
+            return True
+    return False
+
+
+try:
+    import intel_extension_for_pytorch as ipex
+
+    if torch.xpu.is_available():
+        xpu_available = True
+except:
+    pass
+
+try:
+    if torch.backends.mps.is_available():
+        cpu_state = CPUState.MPS
+        import torch.mps
+except:
+    pass
+
+
+def get_torch_device():
+    global directml_enabled
+    global cpu_state
+    if directml_enabled:
+        global directml_device
+        return directml_device
+    if cpu_state == CPUState.MPS:
+        return torch.device("mps")
+    if cpu_state == CPUState.CPU:
+        return torch.device("cpu")
+    else:
+        if is_intel_xpu():
+            return torch.device("xpu")
+        else:
+            return torch.device(torch.cuda.current_device())

pipeline_stable_diffusion_xl_instantid.py → pipeline_stable_diffusion_xl_instantid_full.py

@@ -22,7 +22,6 @@ import numpy as np
 import PIL.Image
 import torch
 import torch.nn.functional as F
-from transformers import CLIPTokenizer
 
 from diffusers.image_processor import PipelineImageInput
 
@@ -41,8 +40,12 @@ from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
 from diffusers.utils.import_utils import is_xformers_available
 
 from ip_adapter.resampler import Resampler
+from ip_adapter.utils import is_torch2_available
 
-from ip_adapter.attention_processor import AttnProcessor, IPAttnProcessor
+if is_torch2_available():
+    from ip_adapter.attention_processor import IPAttnProcessor2_0 as IPAttnProcessor, AttnProcessor2_0 as AttnProcessor
+else:
+    from ip_adapter.attention_processor import IPAttnProcessor, AttnProcessor
 from ip_adapter.attention_processor import region_control
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
@@ -102,7 +105,7 @@ EXAMPLE_DOC_STRING = """
         ```
 """
 
-
+from transformers import CLIPTokenizer
 from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipeline
 class LongPromptWeight(object):
     
@@ -482,6 +485,34 @@ class LongPromptWeight(object):
         prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
         return prompt_embeds
 
+def draw_kps(image_pil, kps, color_list=[(255,0,0), (0,255,0), (0,0,255), (255,255,0), (255,0,255)]):
+    
+    stickwidth = 4
+    limbSeq = np.array([[0, 2], [1, 2], [3, 2], [4, 2]])
+    kps = np.array(kps)
+
+    w, h = image_pil.size
+    out_img = np.zeros([h, w, 3])
+
+    for i in range(len(limbSeq)):
+        index = limbSeq[i]
+        color = color_list[index[0]]
+
+        x = kps[index][:, 0]
+        y = kps[index][:, 1]
+        length = ((x[0] - x[1]) ** 2 + (y[0] - y[1]) ** 2) ** 0.5
+        angle = math.degrees(math.atan2(y[0] - y[1], x[0] - x[1]))
+        polygon = cv2.ellipse2Poly((int(np.mean(x)), int(np.mean(y))), (int(length / 2), stickwidth), int(angle), 0, 360, 1)
+        out_img = cv2.fillConvexPoly(out_img.copy(), polygon, color)
+    out_img = (out_img * 0.6).astype(np.uint8)
+
+    for idx_kp, kp in enumerate(kps):
+        color = color_list[idx_kp]
+        x, y = kp
+        out_img = cv2.circle(out_img.copy(), (int(x), int(y)), 10, color, -1)
+
+    out_img_pil = PIL.Image.fromarray(out_img.astype(np.uint8))
+    return out_img_pil
     
 class StableDiffusionXLInstantIDPipeline(StableDiffusionXLControlNetPipeline):
     
@@ -567,7 +598,7 @@ class StableDiffusionXLInstantIDPipeline(StableDiffusionXLControlNetPipeline):
             if isinstance(attn_processor, IPAttnProcessor):
                 attn_processor.scale = scale
 
-    def _encode_prompt_image_emb(self, prompt_image_emb, device, dtype, do_classifier_free_guidance):
+    def _encode_prompt_image_emb(self, prompt_image_emb, device, num_images_per_prompt, dtype, do_classifier_free_guidance):
         
         if isinstance(prompt_image_emb, torch.Tensor):
             prompt_image_emb = prompt_image_emb.clone().detach()
@@ -583,6 +614,11 @@ class StableDiffusionXLInstantIDPipeline(StableDiffusionXLControlNetPipeline):
             prompt_image_emb = torch.cat([prompt_image_emb], dim=0)
         
         prompt_image_emb = self.image_proj_model(prompt_image_emb)
+
+        bs_embed, seq_len, _ = prompt_image_emb.shape
+        prompt_image_emb = prompt_image_emb.repeat(1, num_images_per_prompt, 1)
+        prompt_image_emb = prompt_image_emb.view(bs_embed * num_images_per_prompt, seq_len, -1)
+        
         return prompt_image_emb
 
     @torch.no_grad()
@@ -623,7 +659,13 @@ class StableDiffusionXLInstantIDPipeline(StableDiffusionXLControlNetPipeline):
         clip_skip: Optional[int] = None,
         callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
         callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+
+        # IP adapter
+        ip_adapter_scale=None,
+
+        # Enhance Face Region
         control_mask = None,
+
         **kwargs,
     ):
         r"""
@@ -758,6 +800,7 @@ class StableDiffusionXLInstantIDPipeline(StableDiffusionXLControlNetPipeline):
                 If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
                 otherwise a `tuple` is returned containing the output images.
         """
+
         lpw = LongPromptWeight()
 
         callback = kwargs.pop("callback", None)
@@ -789,6 +832,10 @@ class StableDiffusionXLInstantIDPipeline(StableDiffusionXLControlNetPipeline):
                 mult * [control_guidance_start],
                 mult * [control_guidance_end],
             )
+        
+        # 0. set ip_adapter_scale
+        if ip_adapter_scale is not None:
+            self.set_ip_adapter_scale(ip_adapter_scale)
 
         # 1. Check inputs. Raise error if not correct
         self.check_inputs(
@@ -851,6 +898,7 @@ class StableDiffusionXLInstantIDPipeline(StableDiffusionXLControlNetPipeline):
         # 3.2 Encode image prompt
         prompt_image_emb = self._encode_prompt_image_emb(image_embeds, 
                                                          device,
+                                                         num_images_per_prompt,
                                                          self.unet.dtype,
                                                          self.do_classifier_free_guidance)
         
@@ -1031,24 +1079,57 @@ class StableDiffusionXLInstantIDPipeline(StableDiffusionXLControlNetPipeline):
                         controlnet_cond_scale = controlnet_cond_scale[0]
                     cond_scale = controlnet_cond_scale * controlnet_keep[i]
 
-                down_block_res_samples, mid_block_res_sample = self.controlnet(
-                    control_model_input,
-                    t,
-                    encoder_hidden_states=prompt_image_emb,
-                    controlnet_cond=image,
-                    conditioning_scale=cond_scale,
-                    guess_mode=guess_mode,
-                    added_cond_kwargs=controlnet_added_cond_kwargs,
-                    return_dict=False,
-                )
+                if isinstance(self.controlnet, MultiControlNetModel):
+                    down_block_res_samples_list, mid_block_res_sample_list = [], []
+                    for control_index in range(len(self.controlnet.nets)):
+                        controlnet = self.controlnet.nets[control_index]
+                        if control_index == 0:
+                            # assume fhe first controlnet is IdentityNet
+                            controlnet_prompt_embeds = prompt_image_emb
+                        else:
+                            controlnet_prompt_embeds = prompt_embeds
+                        down_block_res_samples, mid_block_res_sample = controlnet(control_model_input,
+                                                                                  t,
+                                                                                  encoder_hidden_states=controlnet_prompt_embeds,
+                                                                                  controlnet_cond=image[control_index],
+                                                                                  conditioning_scale=cond_scale[control_index],
+                                                                                  guess_mode=guess_mode,
+                                                                                  added_cond_kwargs=controlnet_added_cond_kwargs,
+                                                                                  return_dict=False)
+
+                        # controlnet mask
+                        if control_index == 0 and control_mask_wight_image_list is not None:
+                            down_block_res_samples = [
+                                down_block_res_sample * mask_weight
+                                for down_block_res_sample, mask_weight in zip(down_block_res_samples, control_mask_wight_image_list)
+                            ]
+                            mid_block_res_sample *= control_mask_wight_image_list[-1]
+
+                        down_block_res_samples_list.append(down_block_res_samples)
+                        mid_block_res_sample_list.append(mid_block_res_sample)
+
+                    mid_block_res_sample = torch.stack(mid_block_res_sample_list).sum(dim=0)
+                    down_block_res_samples = [torch.stack(down_block_res_samples).sum(dim=0) for down_block_res_samples in
+                                              zip(*down_block_res_samples_list)]
+                else:
+                    down_block_res_samples, mid_block_res_sample = self.controlnet(
+                        control_model_input,
+                        t,
+                        encoder_hidden_states=prompt_image_emb,
+                        controlnet_cond=image,
+                        conditioning_scale=cond_scale,
+                        guess_mode=guess_mode,
+                        added_cond_kwargs=controlnet_added_cond_kwargs,
+                        return_dict=False,
+                    )
 
-                # controlnet mask
-                if control_mask_wight_image_list is not None:
-                    down_block_res_samples = [
-                        down_block_res_sample * mask_weight
-                        for down_block_res_sample, mask_weight in zip(down_block_res_samples, control_mask_wight_image_list)
-                    ]
-                    mid_block_res_sample *= control_mask_wight_image_list[-1]
+                    # controlnet mask
+                    if control_mask_wight_image_list is not None:
+                        down_block_res_samples = [
+                            down_block_res_sample * mask_weight
+                            for down_block_res_sample, mask_weight in zip(down_block_res_samples, control_mask_wight_image_list)
+                        ]
+                        mid_block_res_sample *= control_mask_wight_image_list[-1]
 
                 if guess_mode and self.do_classifier_free_guidance:
                     # Infered ControlNet only for the conditional batch.

requirements.txt

@@ -1,7 +1,7 @@
-diffusers==0.25.0
+diffusers==0.25.1
 torch==2.0.0
 torchvision==0.15.1
-transformers==4.36.2
+transformers==4.37.1
 accelerate
 safetensors
 einops
@@ -11,4 +11,8 @@ omegaconf
 peft
 huggingface-hub==0.20.2
 opencv-python
-insightface
+insightface
+gradio
+controlnet_aux
+gdown
+peft