import functools import os import shutil import sys import git import gradio as gr import numpy as np import torch as torch from PIL import Image from gradio_imageslider import ImageSlider from bilateral_normal_integration.bilateral_normal_integration_cupy import bilateral_normal_integration_function import spaces import fire import argparse import os import logging import numpy as np import torch from PIL import Image from tqdm.auto import tqdm import glob import json import cv2 from rembg import remove from segment_anything import sam_model_registry, SamPredictor from datetime import datetime import time import sys sys.path.append("../") from models.geowizard_pipeline import DepthNormalEstimationPipeline from utils.seed_all import seed_all import matplotlib.pyplot as plt from utils.de_normalized import align_scale_shift from utils.depth2normal import * from diffusers import DiffusionPipeline, DDIMScheduler, AutoencoderKL from models.unet_2d_condition import UNet2DConditionModel from transformers import CLIPTextModel, CLIPTokenizer from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection import torchvision.transforms.functional as TF from torchvision.transforms import InterpolationMode device = torch.device("cuda" if torch.cuda.is_available() else "cpu") stable_diffusion_repo_path = "stabilityai/stable-diffusion-2-1-unclip" vae = AutoencoderKL.from_pretrained(stable_diffusion_repo_path, subfolder='vae') scheduler = DDIMScheduler.from_pretrained(stable_diffusion_repo_path, subfolder='scheduler') sd_image_variations_diffusers_path = 'lambdalabs/sd-image-variations-diffusers' image_encoder = CLIPVisionModelWithProjection.from_pretrained(sd_image_variations_diffusers_path, subfolder="image_encoder") feature_extractor = CLIPImageProcessor.from_pretrained(sd_image_variations_diffusers_path, subfolder="feature_extractor") unet = UNet2DConditionModel.from_pretrained('.', subfolder="unet") pipe = DepthNormalEstimationPipeline(vae=vae, image_encoder=image_encoder, feature_extractor=feature_extractor, unet=unet, scheduler=scheduler) try: import xformers pipe.enable_xformers_memory_efficient_attention() except: pass # run without xformers pipe = pipe.to(device) def sam_init(): sam_checkpoint = os.path.join(os.path.dirname(__file__), "sam_pt", "sam_vit_l_0b3195.pth") model_type = "vit_l" sam = sam_model_registry[model_type](checkpoint=sam_checkpoint).to(device=f"cuda") predictor = SamPredictor(sam) return predictor sam_predictor = sam_init() def sam_segment(predictor, input_image, *bbox_coords): bbox = np.array(bbox_coords) image = np.asarray(input_image) start_time = time.time() predictor.set_image(image) masks_bbox, scores_bbox, logits_bbox = predictor.predict( box=bbox, multimask_output=True ) print(f"SAM Time: {time.time() - start_time:.3f}s") out_image = np.zeros((image.shape[0], image.shape[1], 4), dtype=np.uint8) out_image[:, :, :3] = image out_image_bbox = out_image.copy() out_image_bbox[:, :, 3] = masks_bbox[-1].astype(np.uint8) * 255 torch.cuda.empty_cache() return Image.fromarray(out_image_bbox, mode='RGBA'), masks_bbox @spaces.GPU def depth_normal(img_path, denoising_steps, ensemble_size, processing_res, seed, domain): seed = int(seed) if seed >= 0: torch.manual_seed(seed) img = Image.open(img_path) pipe_out = pipe( img, denoising_steps=denoising_steps, ensemble_size=ensemble_size, processing_res=processing_res, batch_size=0, domain=domain, show_progress_bar=True, ) depth_colored = pipe_out.depth_colored normal_colored = pipe_out.normal_colored depth_np = pipe_out.depth_np normal_np = pipe_out.normal_np path_output_dir = os.path.splitext(os.path.basename(img_path))[0] + datetime.now().strftime('%Y%m%d-%H%M%S') os.makedirs(path_output_dir, exist_ok=True) name_base = os.path.splitext(os.path.basename(img_path))[0] depth_path = os.path.join(path_output_dir, f"{name_base}_depth.npy") normal_path = os.path.join(path_output_dir, f"{name_base}_normal.npy") np.save(normal_path, normal_np) np.save(depth_path, depth_np) return depth_colored, normal_colored, [depth_path, normal_path] def reconstruction(image, files): torch.cuda.empty_cache() img = Image.open(image) image_rem = img.convert('RGBA') image_nobg = remove(image_rem, alpha_matting=True) arr = np.asarray(image_nobg)[:,:,-1] x_nonzero = np.nonzero(arr.sum(axis=0)) y_nonzero = np.nonzero(arr.sum(axis=1)) x_min = int(x_nonzero[0].min()) y_min = int(y_nonzero[0].min()) x_max = int(x_nonzero[0].max()) y_max = int(y_nonzero[0].max()) masked_image, mask = sam_segment(sam_predictor, img.convert('RGB'), x_min, y_min, x_max, y_max) depth_np = np.load(files[0]) normal_np = np.load(files[1]) dir_name = os.path.dirname(os.path.realpath(files[0])) mask_output_temp = mask[-1] name_base = os.path.splitext(os.path.basename(files[0]))[0][:-6] normal_np[:, :, 0] *= -1 _, surface, _, _, _ = bilateral_normal_integration_function(normal_np, mask_output_temp, k=2, K=None, max_iter=100, tol=1e-4, cg_max_iter=5000, cg_tol=1e-3) ply_path = os.path.join(dir_name, f"{name_base}_mask.ply") surface.save(ply_path, binary=False) return ply_path def run_demo(): custom_theme = gr.themes.Soft(primary_hue="blue").set( button_secondary_background_fill="*neutral_100", button_secondary_background_fill_hover="*neutral_200") custom_css = '''#disp_image { text-align: center; /* Horizontally center the content */ }''' _TITLE = '''GeoWizard: Unleashing the Diffusion Priors for 3D Geometry Estimation from a Single Image''' _DESCRIPTION = '''
Generate consistent depth and normal from single image. High quality and rich details. (PS: We find the demo running on ZeroGPU output slightly inferior results compared to A100 or 3060 with everything exactly the same.)
''' _GPU_ID = 0 with gr.Blocks(title=_TITLE, theme=custom_theme, css=custom_css) as demo: with gr.Row(): with gr.Column(scale=1): gr.Markdown('# ' + _TITLE) gr.Markdown(_DESCRIPTION) with gr.Row(variant='panel'): with gr.Column(scale=1): input_image = gr.Image(type='filepath', height=320, label='Input image') example_folder = os.path.join(os.path.dirname(__file__), "./files") example_fns = [os.path.join(example_folder, example) for example in os.listdir(example_folder)] gr.Examples( examples=example_fns, inputs=[input_image], cache_examples=False, label='Examples (click one of the images below to start)', examples_per_page=30 ) with gr.Column(scale=1): with gr.Accordion('Advanced options', open=True): with gr.Column(): domain = gr.Radio( [ ("Outdoor", "outdoor"), ("Indoor", "indoor"), ("Object", "object"), ], label="Data Type (Must Select One matches your image)", value="indoor", ) denoising_steps = gr.Slider( label="Number of denoising steps (More steps, better quality)", minimum=1, maximum=50, step=1, value=10, ) ensemble_size = gr.Slider( label="Ensemble size (More steps, higher accuracy)", minimum=1, maximum=15, step=1, value=3, ) seed = gr.Number(0, label='Random Seed. Negative values for not specifying') processing_res = gr.Radio( [ ("Native", 0), ("Recommended", 768), ], label="Processing resolution", value=768, ) run_btn = gr.Button('Generate', variant='primary', interactive=True) with gr.Row(): with gr.Column(): depth = gr.Image(interactive=False, show_label=False) with gr.Column(): normal = gr.Image(interactive=False, show_label=False) with gr.Row(): files = gr.Files( label = "Depth and Normal (numpy)", elem_id = "download", interactive=False, ) with gr.Row(): recon_btn = gr.Button('Is there a salient foreground object? If yes, Click here to Reconstruct its 3D model.', variant='primary', interactive=True) with gr.Row(): reconstructed_3d = gr.Model3D( label = 'Bini post-processed 3D model', height=320, interactive=False, ) run_btn.click(fn=depth_normal, inputs=[input_image, denoising_steps, ensemble_size, processing_res, seed, domain], outputs=[depth, normal, files] ) recon_btn.click(fn=reconstruction, inputs=[input_image, files], outputs=[reconstructed_3d] ) demo.queue().launch(share=True, max_threads=80) if __name__ == '__main__': fire.Fire(run_demo)