Upload 16 files

app.py

@@ -0,0 +1,80 @@
+#!/usr/bin/env python
+
+from __future__ import annotations
+
+import gradio as gr
+import torch
+
+torch.jit.script = lambda f: f
+import spaces
+
+from app_canny import create_demo as create_demo_canny
+from app_depth import create_demo as create_demo_depth
+from model import Model
+from settings import ALLOW_CHANGING_BASE_MODEL, DEFAULT_MODEL_ID, SHOW_DUPLICATE_BUTTON
+from transformers.utils.hub import move_cache
+
+move_cache()
+
+DESCRIPTION = "ControlNet"
+
+if not torch.cuda.is_available():
+    DESCRIPTION += "\n<p>Running on CPU.</p>"
+
+model = Model(base_model_id=DEFAULT_MODEL_ID, task_name="Canny")
+
+
+with gr.Blocks(css="style.css") as demo:
+    gr.Markdown(DESCRIPTION)
+    gr.DuplicateButton(
+        value="Duplicate Space for private use",
+        elem_id="duplicate-button",
+        visible=SHOW_DUPLICATE_BUTTON,
+    )
+
+    with gr.Tabs():
+        with gr.TabItem("Depth"):
+            create_demo_depth(model.process_depth)
+        with gr.TabItem("Canny"):
+            create_demo_canny(model.process_canny)
+
+    with gr.Accordion(label="Base model", open=False):
+        with gr.Row():
+            with gr.Column(scale=5):
+                current_base_model = gr.Text(label="Current base model")
+            with gr.Column(scale=1):
+                check_base_model_button = gr.Button("Check current base model")
+        with gr.Row():
+            with gr.Column(scale=5):
+                new_base_model_id = gr.Text(
+                    label="New base model",
+                    max_lines=1,
+                    placeholder="runwayml/stable-diffusion-v1-5",
+                    info="The base model must be compatible with Stable Diffusion v1.5.",
+                    interactive=ALLOW_CHANGING_BASE_MODEL,
+                )
+            with gr.Column(scale=1):
+                change_base_model_button = gr.Button(
+                    "Change base model", interactive=ALLOW_CHANGING_BASE_MODEL
+                )
+        if not ALLOW_CHANGING_BASE_MODEL:
+            gr.Markdown(
+                """The base model is not allowed to be changed in this Space so as not to slow down the demo, but it can be changed if you duplicate the Space."""
+            )
+
+    check_base_model_button.click(
+        fn=lambda: model.base_model_id,
+        outputs=current_base_model,
+        queue=False,
+        api_name="check_base_model",
+    )
+    gr.on(
+        triggers=[new_base_model_id.submit, change_base_model_button.click],
+        fn=model.set_base_model,
+        inputs=new_base_model_id,
+        outputs=current_base_model,
+        api_name=False,
+    )
+
+if __name__ == "__main__":
+    demo.queue(max_size=20).launch()

app_canny.py

@@ -0,0 +1,127 @@
+#!/usr/bin/env python
+
+import gradio as gr
+
+from settings import (
+    DEFAULT_IMAGE_RESOLUTION,
+    DEFAULT_NUM_IMAGES,
+    MAX_IMAGE_RESOLUTION,
+    MAX_NUM_IMAGES,
+    MAX_SEED,
+)
+from utils import randomize_seed_fn
+
+
+def create_demo(process):
+    with gr.Blocks() as demo:
+        with gr.Row():
+            with gr.Column():
+                image = gr.Image()
+                prompt = gr.Textbox(label="Prompt")
+                run_button = gr.Button("Run")
+                with gr.Accordion("Advanced options", open=False):
+                    num_samples = gr.Slider(
+                        label="Number of images",
+                        minimum=1,
+                        maximum=MAX_NUM_IMAGES,
+                        value=DEFAULT_NUM_IMAGES,
+                        step=1,
+                    )
+                    image_resolution = gr.Slider(
+                        label="Image resolution",
+                        minimum=256,
+                        maximum=MAX_IMAGE_RESOLUTION,
+                        value=DEFAULT_IMAGE_RESOLUTION,
+                        step=256,
+                    )
+                    canny_low_threshold = gr.Slider(
+                        label="Canny low threshold",
+                        minimum=0,
+                        maximum=1.0,
+                        value=0.1,
+                        step=0.05,
+                    )
+                    canny_high_threshold = gr.Slider(
+                        label="Canny high threshold",
+                        minimum=0,
+                        maximum=1.0,
+                        value=0.2,
+                        step=0.05,
+                    )
+                    num_steps = gr.Slider(
+                        label="Number of steps",
+                        minimum=1,
+                        maximum=100,
+                        value=20,
+                        step=1,
+                    )
+                    guidance_scale = gr.Slider(
+                        label="Guidance scale",
+                        minimum=0.1,
+                        maximum=30.0,
+                        value=7.5,
+                        step=0.1,
+                    )
+                    seed = gr.Slider(
+                        label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0
+                    )
+                    randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+                    a_prompt = gr.Textbox(
+                        label="Additional prompt",
+                        value="high-quality, extremely detailed, 4K",
+                    )
+                    n_prompt = gr.Textbox(
+                        label="Negative prompt",
+                        value="longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality",
+                    )
+            with gr.Column():
+                result = gr.Gallery(
+                    label="Output", show_label=False, columns=2, object_fit="scale-down"
+                )
+
+        inputs = [
+            image,
+            prompt,
+            a_prompt,
+            n_prompt,
+            num_samples,
+            image_resolution,
+            num_steps,
+            guidance_scale,
+            seed,
+            canny_low_threshold,
+            canny_high_threshold,
+        ]
+        prompt.submit(
+            fn=randomize_seed_fn,
+            inputs=[seed, randomize_seed],
+            outputs=seed,
+            queue=False,
+            api_name=False,
+        ).then(
+            fn=process,
+            inputs=inputs,
+            outputs=result,
+            api_name=False,
+        )
+        run_button.click(
+            fn=randomize_seed_fn,
+            inputs=[seed, randomize_seed],
+            outputs=seed,
+            queue=False,
+            api_name=False,
+        ).then(
+            fn=process,
+            inputs=inputs,
+            outputs=result,
+            api_name="canny",
+        )
+    return demo
+
+
+if __name__ == "__main__":
+    from model import Model
+
+    model = Model(task_name="Canny")
+    demo = create_demo(model.process_canny)
+    demo.queue().launch()

app_depth.py

@@ -0,0 +1,126 @@
+#!/usr/bin/env python
+
+import gradio as gr
+
+from settings import (
+    DEFAULT_IMAGE_RESOLUTION,
+    DEFAULT_NUM_IMAGES,
+    MAX_IMAGE_RESOLUTION,
+    MAX_NUM_IMAGES,
+    MAX_SEED,
+)
+from utils import randomize_seed_fn
+
+
+def create_demo(process):
+    with gr.Blocks() as demo:
+        with gr.Row():
+            with gr.Column():
+                image = gr.Image()
+                prompt = gr.Textbox(label="Prompt")
+                run_button = gr.Button("Run")
+                with gr.Accordion("Advanced options", open=False):
+                    preprocessor_name = gr.Radio(
+                        label="Preprocessor",
+                        choices=["Midas", "DPT", "None"],
+                        type="value",
+                        value="DPT",
+                    )
+                    num_samples = gr.Slider(
+                        label="Number of images",
+                        minimum=1,
+                        maximum=MAX_NUM_IMAGES,
+                        value=DEFAULT_NUM_IMAGES,
+                        step=1,
+                    )
+                    image_resolution = gr.Slider(
+                        label="Image resolution",
+                        minimum=256,
+                        maximum=MAX_IMAGE_RESOLUTION,
+                        value=DEFAULT_IMAGE_RESOLUTION,
+                        step=256,
+                    )
+                    preprocess_resolution = gr.Slider(
+                        label="Preprocess resolution",
+                        minimum=128,
+                        maximum=512,
+                        value=384,
+                        step=1,
+                    )
+                    num_steps = gr.Slider(
+                        label="Number of steps",
+                        minimum=1,
+                        maximum=100,
+                        value=20,
+                        step=1,
+                    )
+                    guidance_scale = gr.Slider(
+                        label="Guidance scale",
+                        minimum=0.1,
+                        maximum=30.0,
+                        value=7.5,
+                        step=0.1,
+                    )
+                    seed = gr.Slider(
+                        label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0
+                    )
+                    randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+                    a_prompt = gr.Textbox(
+                        label="Additional prompt",
+                        value="high-quality, extremely detailed, 4K",
+                    )
+                    n_prompt = gr.Textbox(
+                        label="Negative prompt",
+                        value="longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality",
+                    )
+            with gr.Column():
+                result = gr.Gallery(
+                    label="Output", show_label=False, columns=2, object_fit="scale-down"
+                )
+
+        inputs = [
+            image,
+            prompt,
+            a_prompt,
+            n_prompt,
+            num_samples,
+            image_resolution,
+            preprocess_resolution,
+            num_steps,
+            guidance_scale,
+            seed,
+            preprocessor_name,
+        ]
+        prompt.submit(
+            fn=randomize_seed_fn,
+            inputs=[seed, randomize_seed],
+            outputs=seed,
+            queue=False,
+            api_name=False,
+        ).then(
+            fn=process,
+            inputs=inputs,
+            outputs=result,
+            api_name=False,
+        )
+        run_button.click(
+            fn=randomize_seed_fn,
+            inputs=[seed, randomize_seed],
+            outputs=seed,
+            queue=False,
+            api_name=False,
+        ).then(
+            fn=process,
+            inputs=inputs,
+            outputs=result,
+            api_name="depth",
+        )
+    return demo
+
+
+if __name__ == "__main__":
+    from model import Model
+
+    model = Model(task_name="depth")
+    demo = create_demo(model.process_depth)
+    demo.queue().launch()

checkpoints/canny/controlnet/config.json

@@ -0,0 +1,52 @@
+{
+  "_class_name": "ControlNetModel",
+  "_diffusers_version": "0.26.3",
+  "_name_or_path": "work_dirs/finetune/MultiGen20M_canny/ft_controlnet_sd15_canny_res512_bs256_lr1e-5_warmup100_iter5k_fp16ft0-1000/checkpoint-5000",
+  "act_fn": "silu",
+  "addition_embed_type": null,
+  "addition_embed_type_num_heads": 64,
+  "addition_time_embed_dim": null,
+  "attention_head_dim": 8,
+  "block_out_channels": [
+    320,
+    640,
+    1280,
+    1280
+  ],
+  "class_embed_type": null,
+  "conditioning_channels": 3,
+  "conditioning_embedding_out_channels": [
+    16,
+    32,
+    96,
+    256
+  ],
+  "controlnet_conditioning_channel_order": "rgb",
+  "cross_attention_dim": 768,
+  "down_block_types": [
+    "CrossAttnDownBlock2D",
+    "CrossAttnDownBlock2D",
+    "CrossAttnDownBlock2D",
+    "DownBlock2D"
+  ],
+  "downsample_padding": 1,
+  "encoder_hid_dim": null,
+  "encoder_hid_dim_type": null,
+  "flip_sin_to_cos": true,
+  "freq_shift": 0,
+  "global_pool_conditions": false,
+  "in_channels": 4,
+  "layers_per_block": 2,
+  "mid_block_scale_factor": 1,
+  "mid_block_type": "UNetMidBlock2DCrossAttn",
+  "norm_eps": 1e-05,
+  "norm_num_groups": 32,
+  "num_attention_heads": null,
+  "num_class_embeds": null,
+  "only_cross_attention": false,
+  "projection_class_embeddings_input_dim": null,
+  "resnet_time_scale_shift": "default",
+  "transformer_layers_per_block": 1,
+  "upcast_attention": false,
+  "use_linear_projection": false
+}

checkpoints/canny/controlnet/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a3fd425077e65024addc5cf73c97195fcfd499b7a5e16868e4251b47cebb0d89
+size 1445157120

checkpoints/depth/controlnet/config.json

@@ -0,0 +1,52 @@
+{
+  "_class_name": "ControlNetModel",
+  "_diffusers_version": "0.26.3",
+  "_name_or_path": "work_dirs/finetune/MultiGen20M_depth/ft_controlnet_sd15_depth_res512_bs256_lr1e-5_warmup100_iter5k_fp16ft0-200/checkpoint-5000",
+  "act_fn": "silu",
+  "addition_embed_type": null,
+  "addition_embed_type_num_heads": 64,
+  "addition_time_embed_dim": null,
+  "attention_head_dim": 8,
+  "block_out_channels": [
+    320,
+    640,
+    1280,
+    1280
+  ],
+  "class_embed_type": null,
+  "conditioning_channels": 3,
+  "conditioning_embedding_out_channels": [
+    16,
+    32,
+    96,
+    256
+  ],
+  "controlnet_conditioning_channel_order": "rgb",
+  "cross_attention_dim": 768,
+  "down_block_types": [
+    "CrossAttnDownBlock2D",
+    "CrossAttnDownBlock2D",
+    "CrossAttnDownBlock2D",
+    "DownBlock2D"
+  ],
+  "downsample_padding": 1,
+  "encoder_hid_dim": null,
+  "encoder_hid_dim_type": null,
+  "flip_sin_to_cos": true,
+  "freq_shift": 0,
+  "global_pool_conditions": false,
+  "in_channels": 4,
+  "layers_per_block": 2,
+  "mid_block_scale_factor": 1,
+  "mid_block_type": "UNetMidBlock2DCrossAttn",
+  "norm_eps": 1e-05,
+  "norm_num_groups": 32,
+  "num_attention_heads": null,
+  "num_class_embeds": null,
+  "only_cross_attention": false,
+  "projection_class_embeddings_input_dim": null,
+  "resnet_time_scale_shift": "default",
+  "transformer_layers_per_block": 1,
+  "upcast_attention": false,
+  "use_linear_projection": false
+}

checkpoints/depth/controlnet/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d7450404d13ef888c9701433a3c17b2a86c021a6d042f9f5d2519602abd7f2f3
+size 1445157120

cv_utils.py

@@ -0,0 +1,17 @@
+import cv2
+import numpy as np
+
+
+def resize_image(input_image, resolution, interpolation=None):
+    H, W, C = input_image.shape
+    H = float(H)
+    W = float(W)
+    k = float(resolution) / max(H, W)
+    H *= k
+    W *= k
+    H = int(np.round(H / 64.0)) * 64
+    W = int(np.round(W / 64.0)) * 64
+    if interpolation is None:
+        interpolation = cv2.INTER_LANCZOS4 if k > 1 else cv2.INTER_AREA
+    img = cv2.resize(input_image, (W, H), interpolation=interpolation)
+    return img

depth_estimator.py

@@ -0,0 +1,25 @@
+import numpy as np
+import PIL.Image
+from controlnet_aux.util import HWC3
+from transformers import pipeline
+
+from cv_utils import resize_image
+
+
+class DepthEstimator:
+    def __init__(self):
+        self.model = pipeline("depth-estimation")
+
+    def __call__(self, image: np.ndarray, **kwargs) -> PIL.Image.Image:
+        detect_resolution = kwargs.pop("detect_resolution", 512)
+        image_resolution = kwargs.pop("image_resolution", 512)
+        image = np.array(image)
+        image = HWC3(image)
+        image = resize_image(image, resolution=detect_resolution)
+        image = PIL.Image.fromarray(image)
+        image = self.model(image)
+        image = image["depth"]
+        image = np.array(image)
+        image = HWC3(image)
+        image = resize_image(image, resolution=image_resolution)
+        return PIL.Image.fromarray(image)

image_segmentor.py

@@ -0,0 +1,33 @@
+import cv2
+import numpy as np
+import PIL.Image
+import torch
+from controlnet_aux.util import HWC3, ade_palette
+from transformers import AutoImageProcessor, UperNetForSemanticSegmentation
+
+from cv_utils import resize_image
+
+
+class ImageSegmentor:
+    def __init__(self):
+        self.image_processor = AutoImageProcessor.from_pretrained("openmmlab/upernet-convnext-small")
+        self.image_segmentor = UperNetForSemanticSegmentation.from_pretrained("openmmlab/upernet-convnext-small")
+
+    @torch.inference_mode()
+    def __call__(self, image: np.ndarray, **kwargs) -> PIL.Image.Image:
+        detect_resolution = kwargs.pop("detect_resolution", 512)
+        image_resolution = kwargs.pop("image_resolution", 512)
+        image = HWC3(image)
+        image = resize_image(image, resolution=detect_resolution)
+        image = PIL.Image.fromarray(image)
+
+        pixel_values = self.image_processor(image, return_tensors="pt").pixel_values
+        outputs = self.image_segmentor(pixel_values)
+        seg = self.image_processor.post_process_semantic_segmentation(outputs, target_sizes=[image.size[::-1]])[0]
+        color_seg = np.zeros((seg.shape[0], seg.shape[1], 3), dtype=np.uint8)
+        for label, color in enumerate(ade_palette()):
+            color_seg[seg == label, :] = color
+        color_seg = color_seg.astype(np.uint8)
+
+        color_seg = resize_image(color_seg, resolution=image_resolution, interpolation=cv2.INTER_NEAREST)
+        return PIL.Image.fromarray(color_seg)

model.py

@@ -0,0 +1,705 @@
+from __future__ import annotations
+
+import gc
+
+import numpy as np
+import PIL.Image
+import torch
+from controlnet_aux.util import HWC3
+from diffusers import (
+    ControlNetModel,
+    DiffusionPipeline,
+    StableDiffusionControlNetPipeline,
+    UniPCMultistepScheduler,
+)
+
+from cv_utils import resize_image
+from preprocessor import Preprocessor
+from settings import MAX_IMAGE_RESOLUTION, MAX_NUM_IMAGES
+
+CONTROLNET_MODEL_IDS = {
+    "Canny": "checkpoints/canny/controlnet",
+
+    "depth": "checkpoints/depth/controlnet",
+}
+
+
+def download_all_controlnet_weights() -> None:
+    for model_id in CONTROLNET_MODEL_IDS.values():
+        ControlNetModel.from_pretrained(model_id)
+
+
+class Model:
+    def __init__(self, base_model_id: str = "runwayml/stable-diffusion-v1-5", task_name: str = "Canny"):
+        self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+        self.base_model_id = ""
+        self.task_name = ""
+        self.pipe = self.load_pipe(base_model_id, task_name)
+        self.preprocessor = Preprocessor()
+
+    def load_pipe(self, base_model_id: str, task_name) -> DiffusionPipeline:
+        if (
+            base_model_id == self.base_model_id
+            and task_name == self.task_name
+            and hasattr(self, "pipe")
+            and self.pipe is not None
+        ):
+            return self.pipe
+        model_id = CONTROLNET_MODEL_IDS[task_name]
+        controlnet = ControlNetModel.from_pretrained(model_id, torch_dtype=torch.float16)
+        pipe = StableDiffusionControlNetPipeline.from_pretrained(
+            base_model_id, safety_checker=None, controlnet=controlnet, torch_dtype=torch.float16
+        )
+        pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
+        if self.device.type == "cuda":
+            pipe.disable_xformers_memory_efficient_attention()
+        pipe.to(self.device)
+        torch.cuda.empty_cache()
+        gc.collect()
+        self.base_model_id = base_model_id
+        self.task_name = task_name
+        return pipe
+
+    def set_base_model(self, base_model_id: str) -> str:
+        if not base_model_id or base_model_id == self.base_model_id:
+            return self.base_model_id
+        del self.pipe
+        torch.cuda.empty_cache()
+        gc.collect()
+        try:
+            self.pipe = self.load_pipe(base_model_id, self.task_name)
+        except Exception:
+            self.pipe = self.load_pipe(self.base_model_id, self.task_name)
+        return self.base_model_id
+
+    def load_controlnet_weight(self, task_name: str) -> None:
+        if task_name == self.task_name:
+            return
+        if self.pipe is not None and hasattr(self.pipe, "controlnet"):
+            del self.pipe.controlnet
+        torch.cuda.empty_cache()
+        gc.collect()
+        model_id = CONTROLNET_MODEL_IDS[task_name]
+        controlnet = ControlNetModel.from_pretrained(model_id, torch_dtype=torch.float16)
+        controlnet.to(self.device)
+        torch.cuda.empty_cache()
+        gc.collect()
+        self.pipe.controlnet = controlnet
+        self.task_name = task_name
+
+    def get_prompt(self, prompt: str, additional_prompt: str) -> str:
+        if not prompt:
+            prompt = additional_prompt
+        else:
+            prompt = f"{prompt}, {additional_prompt}"
+        return prompt
+
+    @torch.autocast("cuda")
+    def run_pipe(
+        self,
+        prompt: str,
+        negative_prompt: str,
+        control_image: PIL.Image.Image,
+        num_images: int,
+        num_steps: int,
+        guidance_scale: float,
+        seed: int,
+    ) -> list[PIL.Image.Image]:
+        generator = torch.Generator().manual_seed(seed)
+        return self.pipe(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            guidance_scale=guidance_scale,
+            num_images_per_prompt=num_images,
+            num_inference_steps=num_steps,
+            generator=generator,
+            image=control_image,
+        ).images
+
+    @torch.inference_mode()
+    def process_canny(
+        self,
+        image: np.ndarray,
+        prompt: str,
+        additional_prompt: str,
+        negative_prompt: str,
+        num_images: int,
+        image_resolution: int,
+        num_steps: int,
+        guidance_scale: float,
+        seed: int,
+        low_threshold: int,
+        high_threshold: int,
+    ) -> list[PIL.Image.Image]:
+        if image is None:
+            raise ValueError
+        if image_resolution > MAX_IMAGE_RESOLUTION:
+            raise ValueError
+        if num_images > MAX_NUM_IMAGES:
+            raise ValueError
+
+        self.preprocessor.load("Canny")
+        control_image = self.preprocessor(
+            image=image, low_threshold=low_threshold, high_threshold=high_threshold, detect_resolution=image_resolution
+        )
+
+        self.load_controlnet_weight("Canny")
+        results = self.run_pipe(
+            prompt=self.get_prompt(prompt, additional_prompt),
+            negative_prompt=negative_prompt,
+            control_image=control_image,
+            num_images=num_images,
+            num_steps=num_steps,
+            guidance_scale=guidance_scale,
+            seed=seed,
+        )
+        conditions_of_generated_imgs = [
+            self.preprocessor(
+                image=x, low_threshold=low_threshold, high_threshold=high_threshold, detect_resolution=image_resolution
+            ) for x in results
+        ]
+        return [control_image] * num_images + results + conditions_of_generated_imgs
+
+    @torch.inference_mode()
+    def process_mlsd(
+        self,
+        image: np.ndarray,
+        prompt: str,
+        additional_prompt: str,
+        negative_prompt: str,
+        num_images: int,
+        image_resolution: int,
+        preprocess_resolution: int,
+        num_steps: int,
+        guidance_scale: float,
+        seed: int,
+        value_threshold: float,
+        distance_threshold: float,
+    ) -> list[PIL.Image.Image]:
+        if image is None:
+            raise ValueError
+        if image_resolution > MAX_IMAGE_RESOLUTION:
+            raise ValueError
+        if num_images > MAX_NUM_IMAGES:
+            raise ValueError
+
+        self.preprocessor.load("MLSD")
+        control_image = self.preprocessor(
+            image=image,
+            image_resolution=image_resolution,
+            detect_resolution=preprocess_resolution,
+            thr_v=value_threshold,
+            thr_d=distance_threshold,
+        )
+        self.load_controlnet_weight("MLSD")
+        results = self.run_pipe(
+            prompt=self.get_prompt(prompt, additional_prompt),
+            negative_prompt=negative_prompt,
+            control_image=control_image,
+            num_images=num_images,
+            num_steps=num_steps,
+            guidance_scale=guidance_scale,
+            seed=seed,
+        )
+        return [control_image] + results
+
+    @torch.inference_mode()
+    def process_scribble(
+        self,
+        image: np.ndarray,
+        prompt: str,
+        additional_prompt: str,
+        negative_prompt: str,
+        num_images: int,
+        image_resolution: int,
+        preprocess_resolution: int,
+        num_steps: int,
+        guidance_scale: float,
+        seed: int,
+        preprocessor_name: str,
+    ) -> list[PIL.Image.Image]:
+        if image is None:
+            raise ValueError
+        if image_resolution > MAX_IMAGE_RESOLUTION:
+            raise ValueError
+        if num_images > MAX_NUM_IMAGES:
+            raise ValueError
+
+        if preprocessor_name == "None":
+            image = HWC3(image)
+            image = resize_image(image, resolution=image_resolution)
+            control_image = PIL.Image.fromarray(image)
+        elif preprocessor_name == "HED":
+            self.preprocessor.load(preprocessor_name)
+            control_image = self.preprocessor(
+                image=image,
+                image_resolution=image_resolution,
+                detect_resolution=preprocess_resolution,
+                scribble=False,
+            )
+        elif preprocessor_name == "PidiNet":
+            self.preprocessor.load(preprocessor_name)
+            control_image = self.preprocessor(
+                image=image,
+                image_resolution=image_resolution,
+                detect_resolution=preprocess_resolution,
+                safe=False,
+            )
+        self.load_controlnet_weight("scribble")
+        results = self.run_pipe(
+            prompt=self.get_prompt(prompt, additional_prompt),
+            negative_prompt=negative_prompt,
+            control_image=control_image,
+            num_images=num_images,
+            num_steps=num_steps,
+            guidance_scale=guidance_scale,
+            seed=seed,
+        )
+        return [control_image] + results
+
+    @torch.inference_mode()
+    def process_scribble_interactive(
+        self,
+        image_and_mask: dict[str, np.ndarray],
+        prompt: str,
+        additional_prompt: str,
+        negative_prompt: str,
+        num_images: int,
+        image_resolution: int,
+        num_steps: int,
+        guidance_scale: float,
+        seed: int,
+    ) -> list[PIL.Image.Image]:
+        if image_and_mask is None:
+            raise ValueError
+        if image_resolution > MAX_IMAGE_RESOLUTION:
+            raise ValueError
+        if num_images > MAX_NUM_IMAGES:
+            raise ValueError
+
+        image = image_and_mask["mask"]
+        image = HWC3(image)
+        image = resize_image(image, resolution=image_resolution)
+        control_image = PIL.Image.fromarray(image)
+
+        self.load_controlnet_weight("scribble")
+        results = self.run_pipe(
+            prompt=self.get_prompt(prompt, additional_prompt),
+            negative_prompt=negative_prompt,
+            control_image=control_image,
+            num_images=num_images,
+            num_steps=num_steps,
+            guidance_scale=guidance_scale,
+            seed=seed,
+        )
+        return [control_image] + results
+
+    @torch.inference_mode()
+    def process_softedge(
+        self,
+        image: np.ndarray,
+        prompt: str,
+        additional_prompt: str,
+        negative_prompt: str,
+        num_images: int,
+        image_resolution: int,
+        preprocess_resolution: int,
+        num_steps: int,
+        guidance_scale: float,
+        seed: int,
+        preprocessor_name: str,
+    ) -> list[PIL.Image.Image]:
+        if image is None:
+            raise ValueError
+        if image_resolution > MAX_IMAGE_RESOLUTION:
+            raise ValueError
+        if num_images > MAX_NUM_IMAGES:
+            raise ValueError
+
+        if preprocessor_name == "None":
+            image = HWC3(image)
+            image = resize_image(image, resolution=image_resolution)
+            control_image = PIL.Image.fromarray(image)
+        elif preprocessor_name in ["HED", "HED safe"]:
+            safe = "safe" in preprocessor_name
+            self.preprocessor.load("HED")
+            control_image = self.preprocessor(
+                image=image,
+                image_resolution=image_resolution,
+                detect_resolution=preprocess_resolution,
+                scribble=safe,
+            )
+        elif preprocessor_name in ["PidiNet", "PidiNet safe"]:
+            safe = "safe" in preprocessor_name
+            self.preprocessor.load("PidiNet")
+            control_image = self.preprocessor(
+                image=image,
+                image_resolution=image_resolution,
+                detect_resolution=preprocess_resolution,
+                safe=safe,
+            )
+        else:
+            raise ValueError
+        self.load_controlnet_weight("softedge")
+        results = self.run_pipe(
+            prompt=self.get_prompt(prompt, additional_prompt),
+            negative_prompt=negative_prompt,
+            control_image=control_image,
+            num_images=num_images,
+            num_steps=num_steps,
+            guidance_scale=guidance_scale,
+            seed=seed,
+        )
+        conditions_of_generated_imgs = [
+            self.preprocessor(
+                image=x,
+                image_resolution=image_resolution,
+                detect_resolution=preprocess_resolution,
+                scribble=safe,
+            ) for x in results
+        ]
+        return [control_image] * num_images + results + conditions_of_generated_imgs
+
+    @torch.inference_mode()
+    def process_openpose(
+        self,
+        image: np.ndarray,
+        prompt: str,
+        additional_prompt: str,
+        negative_prompt: str,
+        num_images: int,
+        image_resolution: int,
+        preprocess_resolution: int,
+        num_steps: int,
+        guidance_scale: float,
+        seed: int,
+        preprocessor_name: str,
+    ) -> list[PIL.Image.Image]:
+        if image is None:
+            raise ValueError
+        if image_resolution > MAX_IMAGE_RESOLUTION:
+            raise ValueError
+        if num_images > MAX_NUM_IMAGES:
+            raise ValueError
+
+        if preprocessor_name == "None":
+            image = HWC3(image)
+            image = resize_image(image, resolution=image_resolution)
+            control_image = PIL.Image.fromarray(image)
+        else:
+            self.preprocessor.load("Openpose")
+            control_image = self.preprocessor(
+                image=image,
+                image_resolution=image_resolution,
+                detect_resolution=preprocess_resolution,
+                hand_and_face=True,
+            )
+        self.load_controlnet_weight("Openpose")
+        results = self.run_pipe(
+            prompt=self.get_prompt(prompt, additional_prompt),
+            negative_prompt=negative_prompt,
+            control_image=control_image,
+            num_images=num_images,
+            num_steps=num_steps,
+            guidance_scale=guidance_scale,
+            seed=seed,
+        )
+        return [control_image] + results
+
+    @torch.inference_mode()
+    def process_segmentation(
+        self,
+        image: np.ndarray,
+        prompt: str,
+        additional_prompt: str,
+        negative_prompt: str,
+        num_images: int,
+        image_resolution: int,
+        preprocess_resolution: int,
+        num_steps: int,
+        guidance_scale: float,
+        seed: int,
+        preprocessor_name: str,
+    ) -> list[PIL.Image.Image]:
+        if image is None:
+            raise ValueError
+        if image_resolution > MAX_IMAGE_RESOLUTION:
+            raise ValueError
+        if num_images > MAX_NUM_IMAGES:
+            raise ValueError
+
+        if preprocessor_name == "None":
+            image = HWC3(image)
+            image = resize_image(image, resolution=image_resolution)
+            control_image = PIL.Image.fromarray(image)
+        else:
+            self.preprocessor.load(preprocessor_name)
+            control_image = self.preprocessor(
+                image=image,
+                image_resolution=image_resolution,
+                detect_resolution=preprocess_resolution,
+            )
+        self.load_controlnet_weight("segmentation")
+        results = self.run_pipe(
+            prompt=self.get_prompt(prompt, additional_prompt),
+            negative_prompt=negative_prompt,
+            control_image=control_image,
+            num_images=num_images,
+            num_steps=num_steps,
+            guidance_scale=guidance_scale,
+            seed=seed,
+        )
+        self.preprocessor.load('UPerNet')
+        conditions_of_generated_imgs = [
+            self.preprocessor(
+                image=np.array(x),
+                image_resolution=image_resolution,
+                detect_resolution=preprocess_resolution,
+            ) for x in results
+        ]
+        return [control_image] * num_images + results + conditions_of_generated_imgs
+
+    @torch.inference_mode()
+    def process_depth(
+        self,
+        image: np.ndarray,
+        prompt: str,
+        additional_prompt: str,
+        negative_prompt: str,
+        num_images: int,
+        image_resolution: int,
+        preprocess_resolution: int,
+        num_steps: int,
+        guidance_scale: float,
+        seed: int,
+        preprocessor_name: str,
+    ) -> list[PIL.Image.Image]:
+        if image is None:
+            raise ValueError
+        if image_resolution > MAX_IMAGE_RESOLUTION:
+            raise ValueError
+        if num_images > MAX_NUM_IMAGES:
+            raise ValueError
+
+        if preprocessor_name == "None":
+            image = HWC3(image)
+            image = resize_image(image, resolution=image_resolution)
+            control_image = PIL.Image.fromarray(image)
+        else:
+            self.preprocessor.load(preprocessor_name)
+            control_image = self.preprocessor(
+                image=image,
+                image_resolution=image_resolution,
+                detect_resolution=preprocess_resolution,
+            )
+        self.load_controlnet_weight("depth")
+        results = self.run_pipe(
+            prompt=self.get_prompt(prompt, additional_prompt),
+            negative_prompt=negative_prompt,
+            control_image=control_image,
+            num_images=num_images,
+            num_steps=num_steps,
+            guidance_scale=guidance_scale,
+            seed=seed,
+        )
+        conditions_of_generated_imgs = [
+            self.preprocessor(
+                image=x,
+                image_resolution=image_resolution,
+                detect_resolution=preprocess_resolution,
+            ) for x in results
+        ]
+        return [control_image] * num_images + results + conditions_of_generated_imgs
+
+    @torch.inference_mode()
+    def process_normal(
+        self,
+        image: np.ndarray,
+        prompt: str,
+        additional_prompt: str,
+        negative_prompt: str,
+        num_images: int,
+        image_resolution: int,
+        preprocess_resolution: int,
+        num_steps: int,
+        guidance_scale: float,
+        seed: int,
+        preprocessor_name: str,
+    ) -> list[PIL.Image.Image]:
+        if image is None:
+            raise ValueError
+        if image_resolution > MAX_IMAGE_RESOLUTION:
+            raise ValueError
+        if num_images > MAX_NUM_IMAGES:
+            raise ValueError
+
+        if preprocessor_name == "None":
+            image = HWC3(image)
+            image = resize_image(image, resolution=image_resolution)
+            control_image = PIL.Image.fromarray(image)
+        else:
+            self.preprocessor.load("NormalBae")
+            control_image = self.preprocessor(
+                image=image,
+                image_resolution=image_resolution,
+                detect_resolution=preprocess_resolution,
+            )
+        self.load_controlnet_weight("NormalBae")
+        results = self.run_pipe(
+            prompt=self.get_prompt(prompt, additional_prompt),
+            negative_prompt=negative_prompt,
+            control_image=control_image,
+            num_images=num_images,
+            num_steps=num_steps,
+            guidance_scale=guidance_scale,
+            seed=seed,
+        )
+        return [control_image] + results
+
+    @torch.inference_mode()
+    def process_lineart(
+        self,
+        image: np.ndarray,
+        prompt: str,
+        additional_prompt: str,
+        negative_prompt: str,
+        num_images: int,
+        image_resolution: int,
+        preprocess_resolution: int,
+        num_steps: int,
+        guidance_scale: float,
+        seed: int,
+        preprocessor_name: str,
+    ) -> list[PIL.Image.Image]:
+        if image is None:
+            raise ValueError
+        if image_resolution > MAX_IMAGE_RESOLUTION:
+            raise ValueError
+        if num_images > MAX_NUM_IMAGES:
+            raise ValueError
+
+        if preprocessor_name in ["None", "None (anime)"]:
+            image = 255 - HWC3(image)
+            image = resize_image(image, resolution=image_resolution)
+            control_image = PIL.Image.fromarray(image)
+        elif preprocessor_name in ["Lineart", "Lineart coarse"]:
+            coarse = "coarse" in preprocessor_name
+            self.preprocessor.load("Lineart")
+            control_image = self.preprocessor(
+                image=image,
+                image_resolution=image_resolution,
+                detect_resolution=preprocess_resolution,
+                coarse=coarse,
+            )
+        elif preprocessor_name == "Lineart (anime)":
+            self.preprocessor.load("LineartAnime")
+            control_image = self.preprocessor(
+                image=image,
+                image_resolution=image_resolution,
+                detect_resolution=preprocess_resolution,
+            )
+        # NOTE: We still use the general lineart model
+        if "anime" in preprocessor_name:
+            self.load_controlnet_weight("lineart_anime")
+        else:
+            self.load_controlnet_weight("lineart")
+        results = self.run_pipe(
+            prompt=self.get_prompt(prompt, additional_prompt),
+            negative_prompt=negative_prompt,
+            control_image=control_image,
+            num_images=num_images,
+            num_steps=num_steps,
+            guidance_scale=guidance_scale,
+            seed=seed,
+        )
+        self.preprocessor.load("Lineart")
+        conditions_of_generated_imgs = [
+            self.preprocessor(
+                image=x,
+                image_resolution=image_resolution,
+                detect_resolution=preprocess_resolution,
+            ) for x in results
+        ]
+
+        control_image = PIL.Image.fromarray((255 - np.array(control_image)).astype(np.uint8))
+        conditions_of_generated_imgs = [PIL.Image.fromarray((255 - np.array(x)).astype(np.uint8)) for x in conditions_of_generated_imgs]
+
+        return [control_image] * num_images + results + conditions_of_generated_imgs
+
+    @torch.inference_mode()
+    def process_shuffle(
+        self,
+        image: np.ndarray,
+        prompt: str,
+        additional_prompt: str,
+        negative_prompt: str,
+        num_images: int,
+        image_resolution: int,
+        num_steps: int,
+        guidance_scale: float,
+        seed: int,
+        preprocessor_name: str,
+    ) -> list[PIL.Image.Image]:
+        if image is None:
+            raise ValueError
+        if image_resolution > MAX_IMAGE_RESOLUTION:
+            raise ValueError
+        if num_images > MAX_NUM_IMAGES:
+            raise ValueError
+
+        if preprocessor_name == "None":
+            image = HWC3(image)
+            image = resize_image(image, resolution=image_resolution)
+            control_image = PIL.Image.fromarray(image)
+        else:
+            self.preprocessor.load(preprocessor_name)
+            control_image = self.preprocessor(
+                image=image,
+                image_resolution=image_resolution,
+            )
+        self.load_controlnet_weight("shuffle")
+        results = self.run_pipe(
+            prompt=self.get_prompt(prompt, additional_prompt),
+            negative_prompt=negative_prompt,
+            control_image=control_image,
+            num_images=num_images,
+            num_steps=num_steps,
+            guidance_scale=guidance_scale,
+            seed=seed,
+        )
+        return [control_image] + results
+
+    @torch.inference_mode()
+    def process_ip2p(
+        self,
+        image: np.ndarray,
+        prompt: str,
+        additional_prompt: str,
+        negative_prompt: str,
+        num_images: int,
+        image_resolution: int,
+        num_steps: int,
+        guidance_scale: float,
+        seed: int,
+    ) -> list[PIL.Image.Image]:
+        if image is None:
+            raise ValueError
+        if image_resolution > MAX_IMAGE_RESOLUTION:
+            raise ValueError
+        if num_images > MAX_NUM_IMAGES:
+            raise ValueError
+
+        image = HWC3(image)
+        image = resize_image(image, resolution=image_resolution)
+        control_image = PIL.Image.fromarray(image)
+        self.load_controlnet_weight("ip2p")
+        results = self.run_pipe(
+            prompt=self.get_prompt(prompt, additional_prompt),
+            negative_prompt=negative_prompt,
+            control_image=control_image,
+            num_images=num_images,
+            num_steps=num_steps,
+            guidance_scale=guidance_scale,
+            seed=seed,
+        )
+        return [control_image] + results

preprocessor.py

@@ -0,0 +1,92 @@
+import gc
+
+import numpy as np
+import PIL.Image
+import torch
+import torchvision
+from controlnet_aux import (
+    CannyDetector,
+    ContentShuffleDetector,
+    HEDdetector,
+    LineartAnimeDetector,
+    LineartDetector,
+    MidasDetector,
+    MLSDdetector,
+    NormalBaeDetector,
+    OpenposeDetector,
+    PidiNetDetector,
+)
+from controlnet_aux.util import HWC3
+
+from cv_utils import resize_image
+from depth_estimator import DepthEstimator
+from image_segmentor import ImageSegmentor
+
+from kornia.core import Tensor
+from kornia.filters import canny
+
+
+class Canny:
+
+    def __call__(
+        self,
+        images: np.array,
+        low_threshold: float = 0.1,
+        high_threshold: float = 0.2,
+        kernel_size: tuple[int, int] | int = (5, 5),
+        sigma: tuple[float, float] | Tensor = (1, 1),
+        hysteresis: bool = True,
+        eps: float = 1e-6
+    ) -> torch.Tensor:
+
+        assert low_threshold is not None, "low_threshold must be provided"
+        assert high_threshold is not None, "high_threshold must be provided"
+
+        images = torch.from_numpy(images).permute(2, 0, 1).unsqueeze(0) / 255.0
+
+        images_tensor = canny(images, low_threshold, high_threshold, kernel_size, sigma, hysteresis, eps)[1]
+        images_tensor = (images_tensor[0][0].numpy() * 255).astype(np.uint8)
+        return images_tensor
+
+
+class Preprocessor:
+    MODEL_ID = "lllyasviel/Annotators"
+
+    def __init__(self):
+        self.model = None
+        self.name = ""
+
+    def load(self, name: str) -> None:
+        if name == self.name:
+            return
+        if name == "Canny":
+            self.model = Canny()
+        elif name == "DPT":
+            self.model = DepthEstimator()
+        else:
+            raise ValueError
+        torch.cuda.empty_cache()
+        gc.collect()
+        self.name = name
+
+    def __call__(self, image: PIL.Image.Image, **kwargs) -> PIL.Image.Image:
+        if self.name == "Canny":
+            if "detect_resolution" in kwargs:
+                detect_resolution = kwargs.pop("detect_resolution")
+                image = np.array(image)
+                image = HWC3(image)
+                image = resize_image(image, resolution=detect_resolution)
+            image = self.model(image, **kwargs)
+            return PIL.Image.fromarray(image).convert('RGB')
+        elif self.name == "Midas":
+            detect_resolution = kwargs.pop("detect_resolution", 512)
+            image_resolution = kwargs.pop("image_resolution", 512)
+            image = np.array(image)
+            image = HWC3(image)
+            image = resize_image(image, resolution=detect_resolution)
+            image = self.model(image, **kwargs)
+            image = HWC3(image)
+            image = resize_image(image, resolution=image_resolution)
+            return PIL.Image.fromarray(image)
+        else:
+            return self.model(image, **kwargs)

requirements.txt

@@ -0,0 +1,117 @@
+absl-py==2.1.0
+accelerate==0.21.0
+aiofiles==23.2.1
+altair==5.3.0
+annotated-types==0.6.0
+anyio==4.3.0
+attrs==23.2.0
+certifi==2024.2.2
+cffi==1.16.0
+charset-normalizer==3.3.2
+click==8.1.7
+cmake==3.29.2
+contourpy==1.2.1
+controlnet-aux==0.0.6
+cycler==0.12.1
+diffusers==0.26.3
+einops==0.6.1
+exceptiongroup==1.2.0
+fastapi==0.110.1
+ffmpy==0.3.2
+filelock==3.13.4
+flatbuffers==24.3.25
+fonttools==4.51.0
+fsspec==2024.3.1
+gradio==4.26.0
+gradio_client==0.15.1
+h11==0.14.0
+httpcore==1.0.5
+httpx==0.27.0
+huggingface-hub==0.21.2
+idna==3.7
+imageio==2.34.0
+importlib_metadata==7.1.0
+importlib_resources==6.4.0
+Jinja2==3.1.3
+jsonschema==4.21.1
+jsonschema-specifications==2023.12.1
+kiwisolver==1.4.5
+kornia==0.7.0
+lazy_loader==0.4
+lit==18.1.3
+markdown-it-py==3.0.0
+MarkupSafe==2.1.5
+matplotlib==3.8.4
+mdurl==0.1.2
+mediapipe==0.10.1
+mpmath==1.3.0
+mypy-extensions==1.0.0
+networkx==3.3
+numpy==1.26.4
+nvidia-cublas-cu11==11.10.3.66
+nvidia-cuda-cupti-cu11==11.7.101
+nvidia-cuda-nvrtc-cu11==11.7.99
+nvidia-cuda-runtime-cu11==11.7.99
+nvidia-cudnn-cu11==8.5.0.96
+nvidia-cufft-cu11==10.9.0.58
+nvidia-curand-cu11==10.2.10.91
+nvidia-cusolver-cu11==11.4.0.1
+nvidia-cusparse-cu11==11.7.4.91
+nvidia-nccl-cu11==2.14.3
+nvidia-nvtx-cu11==11.7.91
+opencv-contrib-python==4.9.0.80
+opencv-python==4.9.0.80
+opencv-python-headless==4.8.0.74
+orjson==3.10.0
+packaging==24.0
+pandas==2.2.2
+pillow==10.3.0
+protobuf==3.20.3
+psutil==5.9.8
+pycparser==2.22
+pydantic==2.7.0
+pydantic_core==2.18.1
+pydub==0.25.1
+Pygments==2.17.2
+pyparsing==3.1.2
+pyre-extensions==0.0.29
+python-dateutil==2.9.0.post0
+python-multipart==0.0.9
+pytz==2024.1
+PyYAML==6.0.1
+referencing==0.34.0
+regex==2023.12.25
+requests==2.31.0
+rich==13.7.1
+rpds-py==0.18.0
+ruff==0.3.7
+safetensors==0.4.1
+scikit-image==0.23.1
+scipy==1.13.0
+semantic-version==2.10.0
+shellingham==1.5.4
+six==1.16.0
+sniffio==1.3.1
+sounddevice==0.4.6
+spaces==0.26.0
+starlette==0.37.2
+sympy==1.12
+tifffile==2024.2.12
+timm==0.9.16
+tokenizers==0.15.2
+tomlkit==0.12.0
+toolz==0.12.1
+torch==2.0.1
+torchvision==0.15.2
+tqdm==4.66.2
+transformers==4.38.1
+triton==2.0.0
+typer==0.12.3
+typing-inspect==0.9.0
+typing_extensions==4.11.0
+tzdata==2024.1
+urllib3==2.2.1
+uvicorn==0.29.0
+websockets==11.0.3
+xformers==0.0.20
+zipp==3.18.1

settings.py

@@ -0,0 +1,15 @@
+import os
+
+import numpy as np
+
+DEFAULT_MODEL_ID = os.getenv("DEFAULT_MODEL_ID", "runwayml/stable-diffusion-v1-5")
+
+MAX_NUM_IMAGES = int(os.getenv("MAX_NUM_IMAGES", "4"))
+DEFAULT_NUM_IMAGES = min(MAX_NUM_IMAGES, int(os.getenv("DEFAULT_NUM_IMAGES", "2")))
+MAX_IMAGE_RESOLUTION = int(os.getenv("MAX_IMAGE_RESOLUTION", "768"))
+DEFAULT_IMAGE_RESOLUTION = min(MAX_IMAGE_RESOLUTION, int(os.getenv("DEFAULT_IMAGE_RESOLUTION", "512")))
+
+ALLOW_CHANGING_BASE_MODEL = os.getenv("SPACE_ID") != "hysts/ControlNet-v1-1"
+SHOW_DUPLICATE_BUTTON = os.getenv("SHOW_DUPLICATE_BUTTON") == "1"
+
+MAX_SEED = np.iinfo(np.int32).max

style.css

@@ -0,0 +1,10 @@
+h1 {
+  text-align: center;
+}
+
+#duplicate-button {
+  margin: auto;
+  color: #fff;
+  background: #1565c0;
+  border-radius: 100vh;
+}

utils.py

@@ -0,0 +1,9 @@
+import random
+
+from settings import MAX_SEED
+
+
+def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
+    if randomize_seed:
+        seed = random.randint(0, MAX_SEED)
+    return seed