First commit

app.py

@@ -0,0 +1,195 @@
+import gradio as gr
+
+from io import BytesIO
+import requests
+import PIL
+from PIL import Image
+import numpy as np
+import os
+import uuid
+import torch
+from torch import autocast
+import cv2
+from matplotlib import pyplot as plt
+from inpainting import StableDiffusionInpaintingPipeline
+from torchvision import transforms
+from clipseg.models.clipseg import CLIPDensePredT
+
+auth_token = os.environ.get("API_TOKEN") or True
+
+def download_image(url):
+    response = requests.get(url)
+    return PIL.Image.open(BytesIO(response.content)).convert("RGB")
+
+img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
+mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+pipe = StableDiffusionInpaintingPipeline.from_pretrained(
+    "CompVis/stable-diffusion-v1-4",
+    revision="fp16", 
+    torch_dtype=torch.float16,
+    use_auth_token=auth_token,
+).to(device)
+
+model = CLIPDensePredT(version='ViT-B/16', reduce_dim=64)
+model.eval()
+model.load_state_dict(torch.load('./clipseg/weights/rd64-uni.pth', map_location=torch.device('cuda')), strict=False)
+
+transform = transforms.Compose([
+      transforms.ToTensor(),
+      transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
+      transforms.Resize((512, 512)),
+])
+
+def predict(radio, dict, word_mask, prompt=""):
+    if(radio == "draw a mask above"):
+        with autocast("cuda"):
+            init_image = dict["image"].convert("RGB").resize((512, 512))
+            mask = dict["mask"].convert("RGB").resize((512, 512))
+    else:
+        img = transform(dict["image"]).unsqueeze(0)
+        word_masks = [word_mask]
+        with torch.no_grad():
+            preds = model(img.repeat(len(word_masks),1,1,1), word_masks)[0]
+        init_image = dict['image'].convert('RGB').resize((512, 512))
+        filename = f"{uuid.uuid4()}.png"
+        plt.imsave(filename,torch.sigmoid(preds[0][0]))
+        img2 = cv2.imread(filename)
+        gray_image = cv2.cvtColor(img2, cv2.COLOR_BGR2GRAY)
+        (thresh, bw_image) = cv2.threshold(gray_image, 100, 255, cv2.THRESH_BINARY)
+        cv2.cvtColor(bw_image, cv2.COLOR_BGR2RGB)
+        mask = Image.fromarray(np.uint8(bw_image)).convert('RGB')
+        os.remove(filename)
+    with autocast("cuda"):
+        images = pipe(prompt = prompt, init_image=init_image, mask_image=mask, strength=0.8)["sample"]
+    return images[0]
+
+# examples = [[dict(image="init_image.png", mask="mask_image.png"), "A panda sitting on a bench"]]
+css = '''
+#image_upload{min-height:400px}
+#image_upload [data-testid="image"], #image_upload [data-testid="image"] > div{min-height: 400px}
+#mask_radio .gr-form{background:transparent; border: none}
+#word_mask{margin-top: .75em !important}
+#word_mask textarea:disabled{opacity: 0.3}
+.footer {
+            margin-bottom: 45px;
+            margin-top: 35px;
+            text-align: center;
+            border-bottom: 1px solid #e5e5e5;
+        }
+        .footer>p {
+            font-size: .8rem;
+            display: inline-block;
+            padding: 0 10px;
+            transform: translateY(10px);
+            background: white;
+        }
+        .dark .footer {
+            border-color: #303030;
+        }
+        .dark .footer>p {
+            background: #0b0f19;
+        }
+        .acknowledgments h4{
+            margin: 1.25em 0 .25em 0;
+            font-weight: bold;
+            font-size: 115%;
+        }
+#image_upload .touch-none{display: flex}
+'''
+def swap_word_mask(radio_option):
+    if(radio_option == "type what to mask below"):
+        return gr.update(interactive=True, placeholder="A cat")
+    else:
+        return gr.update(interactive=False, placeholder="Disabled")
+
+image_blocks = gr.Blocks(css=css)
+with image_blocks as demo:
+    gr.HTML(
+        """
+            <div style="text-align: center; max-width: 650px; margin: 0 auto;">
+              <div
+                style="
+                  display: inline-flex;
+                  align-items: center;
+                  gap: 0.8rem;
+                  font-size: 1.75rem;
+                "
+              >
+                <svg
+                  width="0.65em"
+                  height="0.65em"
+                  viewBox="0 0 115 115"
+                  fill="none"
+                  xmlns="http://www.w3.org/2000/svg"
+                >
+                  <rect width="23" height="23" fill="white"></rect>
+                  <rect y="69" width="23" height="23" fill="white"></rect>
+                  <rect x="23" width="23" height="23" fill="#AEAEAE"></rect>
+                  <rect x="23" y="69" width="23" height="23" fill="#AEAEAE"></rect>
+                  <rect x="46" width="23" height="23" fill="white"></rect>
+                  <rect x="46" y="69" width="23" height="23" fill="white"></rect>
+                  <rect x="69" width="23" height="23" fill="black"></rect>
+                  <rect x="69" y="69" width="23" height="23" fill="black"></rect>
+                  <rect x="92" width="23" height="23" fill="#D9D9D9"></rect>
+                  <rect x="92" y="69" width="23" height="23" fill="#AEAEAE"></rect>
+                  <rect x="115" y="46" width="23" height="23" fill="white"></rect>
+                  <rect x="115" y="115" width="23" height="23" fill="white"></rect>
+                  <rect x="115" y="69" width="23" height="23" fill="#D9D9D9"></rect>
+                  <rect x="92" y="46" width="23" height="23" fill="#AEAEAE"></rect>
+                  <rect x="92" y="115" width="23" height="23" fill="#AEAEAE"></rect>
+                  <rect x="92" y="69" width="23" height="23" fill="white"></rect>
+                  <rect x="69" y="46" width="23" height="23" fill="white"></rect>
+                  <rect x="69" y="115" width="23" height="23" fill="white"></rect>
+                  <rect x="69" y="69" width="23" height="23" fill="#D9D9D9"></rect>
+                  <rect x="46" y="46" width="23" height="23" fill="black"></rect>
+                  <rect x="46" y="115" width="23" height="23" fill="black"></rect>
+                  <rect x="46" y="69" width="23" height="23" fill="black"></rect>
+                  <rect x="23" y="46" width="23" height="23" fill="#D9D9D9"></rect>
+                  <rect x="23" y="115" width="23" height="23" fill="#AEAEAE"></rect>
+                  <rect x="23" y="69" width="23" height="23" fill="black"></rect>
+                </svg>
+                <h1 style="font-weight: 900; margin-bottom: 7px;">
+                  Stable Diffusion Inpainting
+                </h1>
+              </div>
+              <p style="margin-bottom: 10px; font-size: 94%">
+                Inpaint Stable Diffusion by either drawing a mask or typing what to replace
+              </p>
+            </div>
+        """
+    )
+    with gr.Row():
+        with gr.Column():
+            image = gr.Image(source='upload', tool='sketch', elem_id="image_upload", type="pil").style(height=400)
+            with gr.Box(elem_id="mask_radio").style(border=False):
+                radio = gr.Radio(["draw a mask above", "type what to mask below"], value="draw a mask above", show_label=False, interactive=True).style(container=False)
+                word_mask = gr.Textbox(label = "What to find in your image", interactive=False, elem_id="word_mask", placeholder="Disabled").style(container=False)
+            prompt = gr.Textbox(label = 'Your prompt (what you want to add in place of what you are removing)')
+            radio.change(fn=swap_word_mask, inputs=radio, outputs=word_mask)
+            radio.change(None, inputs=[], outputs=image_blocks, _js = """
+            () => {
+                css_style = document.querySelector('gradio-app').shadowRoot.styleSheets[1]
+                last_item = css_style.cssRules[css_style.cssRules.length - 1]
+                last_item.style.display = ["flex", ""].includes(last_item.style.display) ? "none" : "flex";
+            }""")
+            btn = gr.Button("Run")
+        with gr.Column():
+            result = gr.Image()
+        btn.click(fn=predict, inputs=[radio, image, word_mask, prompt], outputs=result)
+    gr.HTML(
+            """
+                <div class="footer">
+                    <p>Model by <a href="https://huggingface.co/CompVis" style="text-decoration: underline;" target="_blank">CompVis</a> and <a href="https://huggingface.co/stabilityai" style="text-decoration: underline;" target="_blank">Stability AI</a> - Inpainting by nagolinc and patil-suraj, inpainting with words by @yvrjsharma and @1littlecoder - Gradio Demo by 🤗 Hugging Face
+                    </p>
+                </div>
+                <div class="acknowledgments">
+                    <p><h4>LICENSE</h4>
+The model is licensed with a <a href="https://huggingface.co/spaces/CompVis/stable-diffusion-license" style="text-decoration: underline;" target="_blank">CreativeML Open RAIL-M</a> license. The authors claim no rights on the outputs you generate, you are free to use them and are accountable for their use which must not go against the provisions set in this license. The license forbids you from sharing any content that violates any laws, produce any harm to a person, disseminate any personal information that would be meant for harm, spread misinformation and target vulnerable groups. For the full list of restrictions please <a href="https://huggingface.co/spaces/CompVis/stable-diffusion-license" target="_blank" style="text-decoration: underline;" target="_blank">read the license</a></p>
+                    <p><h4>Biases and content acknowledgment</h4>
+Despite how impressive being able to turn text into image is, beware to the fact that this model may output content that reinforces or exacerbates societal biases, as well as realistic faces, pornography and violence. The model was trained on the <a href="https://laion.ai/blog/laion-5b/" style="text-decoration: underline;" target="_blank">LAION-5B dataset</a>, which scraped non-curated image-text-pairs from the internet (the exception being the removal of illegal content) and is meant for research purposes. You can read more in the <a href="https://huggingface.co/CompVis/stable-diffusion-v1-4" style="text-decoration: underline;" target="_blank">model card</a></p>
+               </div>
+           """
+        )
+demo.launch()

inpainting.py

@@ -0,0 +1,194 @@
+import inspect
+from typing import List, Optional, Union
+
+import numpy as np
+import torch
+
+import PIL
+from diffusers import AutoencoderKL, DDIMScheduler, DiffusionPipeline, PNDMScheduler, UNet2DConditionModel
+from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
+from tqdm.auto import tqdm
+from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+
+
+def preprocess_image(image):
+    w, h = image.size
+    w, h = map(lambda x: x - x % 32, (w, h))  # resize to integer multiple of 32
+    image = image.resize((w, h), resample=PIL.Image.LANCZOS)
+    image = np.array(image).astype(np.float32) / 255.0
+    image = image[None].transpose(0, 3, 1, 2)
+    image = torch.from_numpy(image)
+    return 2.0 * image - 1.0
+
+
+def preprocess_mask(mask):
+    mask = mask.convert("L")
+    w, h = mask.size
+    w, h = map(lambda x: x - x % 32, (w, h))  # resize to integer multiple of 32
+    mask = mask.resize((w // 8, h // 8), resample=PIL.Image.NEAREST)
+    mask = np.array(mask).astype(np.float32) / 255.0
+    mask = np.tile(mask, (4, 1, 1))
+    mask = mask[None].transpose(0, 1, 2, 3)  # what does this step do?
+    mask = 1 - mask  # repaint white, keep black
+    mask = torch.from_numpy(mask)
+    return mask
+
+class StableDiffusionInpaintingPipeline(DiffusionPipeline):
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: Union[DDIMScheduler, PNDMScheduler],
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPFeatureExtractor,
+    ):
+        super().__init__()
+        scheduler = scheduler.set_format("pt")
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+        )
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]],
+        init_image: torch.FloatTensor,
+        mask_image: torch.FloatTensor,
+        strength: float = 0.8,
+        num_inference_steps: Optional[int] = 50,
+        guidance_scale: Optional[float] = 7.5,
+        eta: Optional[float] = 0.0,
+        generator: Optional[torch.Generator] = None,
+        output_type: Optional[str] = "pil",
+    ):
+
+        if isinstance(prompt, str):
+            batch_size = 1
+        elif isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if strength < 0 or strength > 1:
+            raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+        # set timesteps
+        accepts_offset = "offset" in set(inspect.signature(self.scheduler.set_timesteps).parameters.keys())
+        extra_set_kwargs = {}
+        offset = 0
+        if accepts_offset:
+            offset = 1
+            extra_set_kwargs["offset"] = 1
+
+        self.scheduler.set_timesteps(num_inference_steps, **extra_set_kwargs)
+
+        # preprocess image
+        init_image = preprocess_image(init_image).to(self.device)
+
+        # encode the init image into latents and scale the latents
+        init_latent_dist = self.vae.encode(init_image).latent_dist
+        init_latents = init_latent_dist.sample(generator=generator)
+        init_latents = 0.18215 * init_latents
+
+        # prepare init_latents noise to latents
+        init_latents = torch.cat([init_latents] * batch_size)
+        init_latents_orig = init_latents
+
+        # preprocess mask
+        mask = preprocess_mask(mask_image).to(self.device)
+        mask = torch.cat([mask] * batch_size)
+
+        # check sizes
+        if not mask.shape == init_latents.shape:
+            raise ValueError(f"The mask and init_image should be the same size!")
+
+        # get the original timestep using init_timestep
+        init_timestep = int(num_inference_steps * strength) + offset
+        init_timestep = min(init_timestep, num_inference_steps)
+        timesteps = self.scheduler.timesteps[-init_timestep]
+        timesteps = torch.tensor([timesteps] * batch_size, dtype=torch.long, device=self.device)
+
+        # add noise to latents using the timesteps
+        noise = torch.randn(init_latents.shape, generator=generator, device=self.device)
+        init_latents = self.scheduler.add_noise(init_latents, noise, timesteps)
+
+        # get prompt text embeddings
+        text_input = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        text_embeddings = self.text_encoder(text_input.input_ids.to(self.device))[0]
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            max_length = text_input.input_ids.shape[-1]
+            uncond_input = self.tokenizer(
+                [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt"
+            )
+            uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        latents = init_latents
+        t_start = max(num_inference_steps - init_timestep + offset, 0)
+        for i, t in tqdm(enumerate(self.scheduler.timesteps[t_start:])):
+            # expand the latents if we are doing classifier free guidance
+            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+
+            # predict the noise residual
+            noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings)["sample"]
+
+            # perform guidance
+            if do_classifier_free_guidance:
+                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs)["prev_sample"]
+
+            # masking
+            init_latents_proper = self.scheduler.add_noise(init_latents_orig, noise, t)
+            latents = (init_latents_proper * mask) + (latents * (1 - mask))
+
+        # scale and decode the image latents with vae
+        latents = 1 / 0.18215 * latents
+        image = self.vae.decode(latents).sample
+
+        image = (image / 2 + 0.5).clamp(0, 1)
+        image = image.cpu().permute(0, 2, 3, 1).numpy()
+
+        # run safety checker
+        safety_cheker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(self.device)
+        image, has_nsfw_concept = self.safety_checker(images=image, clip_input=safety_cheker_input.pixel_values)
+
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+
+        return {"sample": image, "nsfw_content_detected": has_nsfw_concept}

requirements.txt

@@ -0,0 +1,9 @@
+torch
+torchvision
+diffusers
+transformers
+ftfy
+numpy
+matplotlib
+uuid
+opencv-python