import cv2 import gradio as gr import numpy as np import torch from diffusers import StableDiffusionControlNetPipeline, StableDiffusionLatentUpscalePipeline, ControlNetModel, AutoencoderKL from diffusers import UniPCMultistepScheduler from PIL import Image from lpw import _encode_prompt controlnet_ColorCanny = ControlNetModel.from_pretrained("ghoskno/Color-Canny-Controlnet-model", torch_dtype=torch.float16) vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-mse", torch_dtype=torch.float16) pipe = StableDiffusionControlNetPipeline.from_pretrained("Lykon/DreamShaper", vae=vae, controlnet=controlnet_ColorCanny, torch_dtype=torch.float16) pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config) pipe.enable_model_cpu_offload() pipe.enable_xformers_memory_efficient_attention() pipe.enable_attention_slicing() # Generator seed generator = torch.manual_seed(0) def HWC3(x): assert x.dtype == np.uint8 if x.ndim == 2: x = x[:, :, None] assert x.ndim == 3 H, W, C = x.shape assert C == 1 or C == 3 or C == 4 if C == 3: return x if C == 1: return np.concatenate([x, x, x], axis=2) if C == 4: color = x[:, :, 0:3].astype(np.float32) alpha = x[:, :, 3:4].astype(np.float32) / 255.0 y = color * alpha + 255.0 * (1.0 - alpha) y = y.clip(0, 255).astype(np.uint8) return y def resize_image(input_image, resolution, max_edge=False, edge_limit=False): H, W, C = input_image.shape H = float(H) W = float(W) if max_edge: k = float(resolution) / max(H, W) else: k = float(resolution) / min(H, W) H *= k W *= k H, W = int(H), int(W) img = cv2.resize(input_image, (W, H), interpolation=cv2.INTER_LANCZOS4 if k > 1 else cv2.INTER_AREA) if not edge_limit: return img pH = int(np.round(H / 64.0)) * 64 pW = int(np.round(W / 64.0)) * 64 pimg = np.zeros((pH, pW, 3), dtype=img.dtype) oH, oW = (pH-H)//2, (pW-W)//2 pimg[oH:oH+H, oW:oW+W] = img return pimg def get_canny_filter(image, low_threshold=100, high_threshold=200): image = cv2.Canny(image, low_threshold, high_threshold) image = image[:, :, None] image = np.concatenate([image, image, image], axis=2) return image def get_color_filter(cond_image, mask_size=64): H, W = cond_image.shape[:2] cond_image = cv2.resize(cond_image, (W // mask_size, H // mask_size), interpolation=cv2.INTER_CUBIC) color = cv2.resize(cond_image, (W, H), interpolation=cv2.INTER_NEAREST) return color def get_colorcanny(image, mask_size): canny_img = get_canny_filter(image) color_img = get_color_filter(image, int(mask_size)) color_img[np.where(canny_img > 128)] = 255 return color_img def process(input_image, prompt, n_prompt, strength=1.0, color_mask_size=96, size=512, scale=6.0, ddim_steps=20): prompt_embeds, negative_prompt_embeds = _encode_prompt(pipe, prompt, pipe.device, 1, True, n_prompt, 3) input_image = resize_image(input_image, size, max_edge=True, edge_limit=True) cond_img = get_colorcanny(input_image, color_mask_size) cond_img = Image.fromarray(cond_img) output = pipe( prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, image=cond_img, generator=generator, num_images_per_prompt=1, num_inference_steps=ddim_steps, guidance_scale=scale, controlnet_conditioning_scale=float(strength) ) return [output.images[0], cond_img] def inpaint_process(inpaint_image, input_image, prompt, n_prompt, strength=1.0, color_mask_size=96, size=512, scale=6.0, ddim_steps=20): if inpaint_image is None: return process(input_image, prompt, n_prompt, strength, color_mask_size, size, scale, ddim_steps) prompt_embeds, negative_prompt_embeds = _encode_prompt(pipe, prompt, pipe.device, 1, True, n_prompt, 3) input_image = resize_image(input_image, size, max_edge=True, edge_limit=True) inpaint_image = resize_image(inpaint_image, size, max_edge=True, edge_limit=True) canny_img = get_canny_filter(input_image) color_img = get_color_filter(inpaint_image, int(color_mask_size)) color_img[np.where(canny_img > 128)] = 255 cond_img = Image.fromarray(color_img) output = pipe( prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, image=cond_img, generator=generator, num_images_per_prompt=1, num_inference_steps=ddim_steps, guidance_scale=scale, controlnet_conditioning_scale=float(strength) ) return [output.images[0], cond_img] block = gr.Blocks().queue() with block: gr.Markdown(""" # 🧨 Color-Canny-ControlNet This is an extended model of ControlNet that not only utilizes the Canny edge of images but also incorporates the color features. We trained this model on the cleaned laion-art dataset that contains 2.6 million images with 2 epochs, using the Canny edge and color mosaic of the images as input. The processed dataset and pretrained model can be found in [ghoskno/laion-art-en-colorcanny](https://huggingface.co/datasets/ghoskno/laion-art-en-colorcanny) and [ghoskno/Color-Canny-Controlnet-model](https://huggingface.co/ghoskno/Color-Canny-Controlnet-model). This allows generated images to maintain the same color composition as the original images. If you are looking to control both the contours and colors of the original image while using ControlNet to generate images, then this is the best option for you! You can try out this model or test the examples provided below 🤗. ## Update Hi, everyone, We have added a Color-Canny-ControlNet accelerated version of our implementation based on Nvidia Triton and operator optimization. This faster ControlNet is deployed on a Nvidia A10 machine. For a 512-pixel image, the inference takes about 1.2s, which is more faster than general implementation with accelerate PyTorch2.0 about 40%. We provide detailed test results as shown below. | Method | Infomation | Inference Times | Speed-up Ratio | | --------------------- | --------------------- | --------------------- | -------------- | | Benchmark | [Huggingface implement](https://huggingface.co/blog/controlnet?spm=ata.21736010.0.0.422d24288Kj7zm) | 3.00s(V100)/5.00s(T4) | / | | Accelerate PyTorch2.0 | xFormers | 2.03s(A10) | 0 | | | SDPA | 2.02s(A10) | 0.5% | | Ours | TRT & OP optimize | 1.20s(A10) | 40.4% | Welcome to try this [faster Color-Canny-ControlNet](http://121.40.118.209:7860/). """) with gr.Row(): with gr.Column(): input_image = gr.Image(source='upload', type="numpy") color_image = gr.ImagePaint(type="numpy") prompt = gr.Textbox(label="Prompt", value='') n_prompt = gr.Textbox(label="Negative Prompt", value='') with gr.Row(): run_button = gr.Button(label="Run") run_edit_button = gr.Button(value='Run with inpaint color', label="Run with inpaint color") with gr.Accordion('Advanced', open=False): strength = gr.Slider(label="Control Strength", minimum=0.0, maximum=2.0, value=1.0, step=0.01) color_mask_size = gr.Slider(label="Color Mask Size", minimum=32, maximum=256, value=96, step=16) size = gr.Slider(label="Size", minimum=256, maximum=768, value=512, step=128) scale = gr.Slider(label="Guidance Scale", minimum=0.1, maximum=30.0, value=6.0, step=0.1) ddim_steps = gr.Slider(label="Steps", minimum=1, maximum=50, value=20, step=1) with gr.Column(): result_gallery = gr.Gallery(label='Output', show_label=False, elem_id="gallery").style(grid=2, height='auto') ips = [input_image, prompt, n_prompt, strength, color_mask_size, size, scale, ddim_steps] run_button.click(fn=process, inputs=ips, outputs=[result_gallery]) run_edit_button.click(fn=inpaint_process, inputs=[color_image] + ips, outputs=[result_gallery]) gr.Examples( examples=[ ["./asserts/1.png", "a concept art of by Makoto Shinkai, a girl is standing in the middle of the sea", "text, bad anatomy, blurry, (low quality, blurry)"], ["./asserts/2.png", "a concept art with vivid ocean by Makoto Shinkai", "text, bad anatomy, blurry, (low quality, blurry)"], ["./asserts/3.png", "sky city on the sea, with waves churning and wind power plants on the island", "text, bad anatomy, blurry, (low quality, blurry)"], ], inputs=[ input_image, prompt, n_prompt ], outputs=[result_gallery], fn=process, cache_examples=True, ) block.launch(debug = True, server_name='0.0.0.0')