from diffusers import StableDiffusionControlNetPipeline, ControlNetModel from segment_anything import sam_model_registry, SamAutomaticMaskGenerator from PIL import Image import gradio as gr import numpy as np import requests import torch import gc device = "cuda" if torch.cuda.is_available() else "cpu" # Download and Create SAM Model print("[Downloading SAM Weights]") SAM_URL = "https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth" r = requests.get(SAM_URL, allow_redirects=True) print("[Writing SAM Weights]") with open("./sam_vit_h_4b8939.pth", "wb") as sam_weights: sam_weights.write(r.content) del r gc.collect() sam = sam_model_registry["vit_h"](checkpoint="./sam_vit_h_4b8939.pth").to(device) mask_generator = SamAutomaticMaskGenerator(sam) gc.collect() # Create ControlNet Pipeline print("Creating ControlNet Pipeline") controlnet = ControlNetModel.from_pretrained( "mfidabel/controlnet-segment-anything", torch_dtype=torch.float16 ).to(device) pipe = StableDiffusionControlNetPipeline.from_pretrained( "runwayml/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=torch.float16, safety_check=None ).to(device) # Description title = "# 🧨 ControlNet on Segment Anything 🤗" description = """This is a demo on 🧨 ControlNet based on Meta's [Segment Anything Model](https://segment-anything.com/). Upload an Image, Segment it with Segment Anything, write a prompt, and generate images 🤗 ⌛️ It takes about 20~ seconds to generate 4 samples, to get faster results, don't forget to reduce the Nº Samples to 1. You can obtain the Segmentation Map of any Image through this Colab: [![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/mfidabel/JAX_SPRINT_2023/blob/main/Segment_Anything_JAX_SPRINT.ipynb) A huge thanks goes out to @GoogleCloud, for providing us with powerful TPUs that enabled us to train this model; and to the @HuggingFace Team for organizing the sprint. Check out our [Model Card 🧨](https://huggingface.co/mfidabel/controlnet-segment-anything) """ about = """ # 👨‍💻 About the model This [model](https://huggingface.co/mfidabel/controlnet-segment-anything) is based on the [ControlNet Model](https://huggingface.co/blog/controlnet), which allow us to generate Images using some sort of condition image. For this model, we selected the segmentation maps produced by Meta's new segmentation model called [Segment Anything Model](https://github.com/facebookresearch/segment-anything) as the condition image. We then trained the model to generate images based on the structure of the segmentation maps and the text prompts given. # 💾 About the dataset For the training, we generated a segmented dataset based on the [COYO-700M](https://huggingface.co/datasets/kakaobrain/coyo-700m) dataset. The dataset provided us with the images, and the text prompts. For the segmented images, we used [Segment Anything Model](https://github.com/facebookresearch/segment-anything). We then created 8k samples to train our model on, which isn't a lot, but as a team, we have been very busy with many other responsibilities and time constraints, which made it challenging to dedicate a lot of time to generating a larger dataset. Despite the constraints we faced, we have still managed to achieve some nice results 🙌 You can check the generated datasets below ⬇️ - [sam-coyo-2k](https://huggingface.co/datasets/mfidabel/sam-coyo-2k) - [sam-coyo-2.5k](https://huggingface.co/datasets/mfidabel/sam-coyo-2.5k) - [sam-coyo-3k](https://huggingface.co/datasets/mfidabel/sam-coyo-3k) """ gif_html = """ “” """ examples = [["photo of a futuristic dining table, high quality, tricolor", "low quality, deformed, blurry, points", "examples/condition_image_1.jpeg"], ["a monochrome photo of henry cavil using a shirt, high quality", "low quality, low res, deformed", "examples/condition_image_2.jpeg"], ["photo of a japanese living room, high quality, coherent", "low quality, colors, saturation, extreme brightness, blurry, low res", "examples/condition_image_3.jpeg"], ["living room, detailed, high quality", "low quality, low resolution, render, oversaturated, low contrast", "examples/condition_image_4.jpeg"], ["painting of the bodiam castle, Vicent Van Gogh style, Starry Night", "low quality, low resolution, render, oversaturated, low contrast", "examples/condition_image_5.jpeg"], ["painting of food, olive oil can, purple wine, green cabbage, chili peppers, pablo picasso style, high quality", "low quality, low resolution, render, oversaturated, low contrast, realistic", "examples/condition_image_6.jpeg"], ["Katsushika Hokusai painting of mountains, a sky and desert landscape, The Great Wave off Kanagawa style, colorful", "low quality, low resolution, render, oversaturated, low contrast, realistic", "examples/condition_image_7.jpeg"]] default_example = examples[4] examples = examples[::-1] css = "h1 { text-align: center } .about { text-align: justify; padding-left: 10%; padding-right: 10%; }" # Inference Function def show_anns(anns): if len(anns) == 0: return sorted_anns = sorted(anns, key=(lambda x: x['area']), reverse=True) h, w = anns[0]['segmentation'].shape final_img = Image.fromarray(np.zeros((h, w, 3), dtype=np.uint8), mode="RGB") for ann in sorted_anns: m = ann['segmentation'] img = np.empty((m.shape[0], m.shape[1], 3), dtype=np.uint8) for i in range(3): img[:,:,i] = np.random.randint(255, dtype=np.uint8) final_img.paste(Image.fromarray(img, mode="RGB"), (0, 0), Image.fromarray(np.uint8(m*255))) return final_img def segment_image(image, seed = 0): # Generate Masks np.random.seed(int(seed)) masks = mask_generator.generate(image) torch.cuda.empty_cache() # Create map map = show_anns(masks) del masks gc.collect() torch.cuda.empty_cache() return map def infer(prompts, negative_prompts, image, num_inference_steps = 50, seed = 4, num_samples = 4): try: # Segment Image print("Segmenting Everything") segmented_map = segment_image(image, seed) yield segmented_map, [Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8))] * num_samples # Generate rng = torch.Generator(device="cpu").manual_seed(seed) num_inference_steps = int(num_inference_steps) print(f"Generating Prompt: {prompts} \nNegative Prompt: {negative_prompts} \nSamples:{num_samples}") output = pipe([prompts] * num_samples, [segmented_map] * num_samples, negative_prompt = [negative_prompts] * num_samples, generator = rng, num_inference_steps = num_inference_steps) final_image = output.images del output except Exception as e: print("Error: " + str(e)) final_image = segmented_map = [np.zeros((512, 512, 3), dtype=np.uint8)] * num_samples finally: gc.collect() torch.cuda.empty_cache() yield segmented_map, final_image cond_img = gr.Image(label="Input", shape=(512, 512), value=default_example[2])\ .style(height=400) segm_img = gr.Image(label="Segmented Image", shape=(512, 512), interactive=False)\ .style(height=400) output = gr.Gallery(label="Generated images")\ .style(height=200, rows=[2], columns=[2], object_fit="contain") prompt = gr.Textbox(lines=1, label="Prompt", value=default_example[0]) negative_prompt = gr.Textbox(lines=1, label="Negative Prompt", value=default_example[1]) with gr.Blocks(css=css) as demo: with gr.Row(): with gr.Column(): # Title gr.Markdown(title) # Description gr.Markdown(description) with gr.Column(): # Examples gr.Markdown(gif_html) # Images with gr.Row(variant="panel"): with gr.Column(scale=1): cond_img.render() with gr.Column(scale=1): segm_img.render() with gr.Column(scale=1): output.render() # Submit & Clear with gr.Row(): with gr.Column(): prompt.render() negative_prompt.render() with gr.Column(): with gr.Accordion("Advanced options", open=False): num_steps = gr.Slider(10, 60, 50, step=1, label="Steps") seed = gr.Slider(0, 1024, 4, step=1, label="Seed") num_samples = gr.Slider(1, 4, 4, step=1, label="Nº Samples") segment_btn = gr.Button("Segment") submit = gr.Button("Segment & Generate Images") # TODO: Download Button with gr.Row(): with gr.Column(): gr.Markdown("Try some of the examples below ⬇️") gr.Examples(examples=examples, inputs=[prompt, negative_prompt, cond_img], outputs=output, fn=infer, examples_per_page=4) with gr.Column(): gr.Markdown(about, elem_classes="about") submit.click(infer, inputs=[prompt, negative_prompt, cond_img, num_steps, seed, num_samples], outputs = [segm_img, output]) segment_btn.click(segment_image, inputs=[cond_img, seed], outputs=segm_img) demo.queue() demo.launch()