import os import gradio as gr from PIL import Image import torch import matplotlib.pyplot as plt import imageio import numpy as np import argparse from point_e.diffusion.configs import DIFFUSION_CONFIGS, diffusion_from_config from point_e.diffusion.sampler import PointCloudSampler from point_e.models.download import load_checkpoint from point_e.models.configs import MODEL_CONFIGS, model_from_config from point_e.util.plotting import plot_point_cloud from point_e.util.ply_util import write_ply from diffusers import StableDiffusionPipeline import trimesh import uuid state = "" device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') css = ''' .instruction{position: absolute; top: 0;right: 0;margin-top: 0px !important} .arrow{position: absolute;top: 0;right: -110px;margin-top: -8px !important} #component-4, #component-3, #component-10{min-height: 0} .duplicate-button img{margin: 0} ''' def set_state(s): print(s) global state state = s def get_state(): return state def load_img2mesh_model(model_name): set_state(f'Creating img2mesh model {model_name}...') i2m_name = model_name i2m_model = model_from_config(MODEL_CONFIGS[i2m_name], device) i2m_model.eval() base_diffusion_i2m = diffusion_from_config(DIFFUSION_CONFIGS[i2m_name]) set_state(f'Downloading img2mesh checkpoint {model_name}...') i2m_model.load_state_dict(load_checkpoint(i2m_name, device)) return i2m_model, base_diffusion_i2m def get_sampler(model_name, txt2obj, guidance_scale): if txt2obj: set_state('Creating txt2mesh model...') t2m_name = 'base40M-textvec' t2m_model = model_from_config(MODEL_CONFIGS[t2m_name], device) t2m_model.eval() base_diffusion_t2m = diffusion_from_config(DIFFUSION_CONFIGS[t2m_name]) set_state('Downloading txt2mesh checkpoint...') t2m_model.load_state_dict(load_checkpoint(t2m_name, device)) else: i2m_model, base_diffusion_i2m = load_img2mesh_model(model_name) set_state('Creating upsample model...') upsampler_model = model_from_config(MODEL_CONFIGS['upsample'], device) upsampler_model.eval() upsampler_diffusion = diffusion_from_config(DIFFUSION_CONFIGS['upsample']) set_state('Downloading upsampler checkpoint...') upsampler_model.load_state_dict(load_checkpoint('upsample', device)) return PointCloudSampler( device=device, models=[t2m_model if txt2obj else i2m_model, upsampler_model], diffusions=[base_diffusion_t2m if txt2obj else base_diffusion_i2m, upsampler_diffusion], num_points=[1024, 4096 - 1024], aux_channels=['R', 'G', 'B'], guidance_scale=[guidance_scale, 0.0 if txt2obj else guidance_scale], model_kwargs_key_filter=('texts', '') if txt2obj else ("*",) ) def generate_txt2img(prompt): pipe = StableDiffusionPipeline.from_pretrained("point_e_model_cache/stable-diffusion-2-1", torch_dtype=torch.float16) pipe = pipe.to("cuda") image = pipe(prompt).images[0] return image def generate_3D(input, model_name='base1B', guidance_scale=3.0, grid_size=128): set_state('Entered generate function...') # try: # input = Image.fromarray(input) # except: # img = generate_txt2img(input) # img.save('/tmp/img.png') # input = Image.open('/tmp/img.png') if isinstance(input, Image.Image): input = prepare_img(input) # if input is a string, it's a text prompt sampler = get_sampler(model_name, txt2obj=True if isinstance(input, str) else False, guidance_scale=guidance_scale) # Produce a sample from the model. set_state('Sampling...') samples = None kw_args = dict(texts=[input]) if isinstance(input, str) else dict(images=[input]) for x in sampler.sample_batch_progressive(batch_size=1, model_kwargs=kw_args): samples = x set_state('Converting to point cloud...') pc = sampler.output_to_point_clouds(samples)[0] set_state('Converting to mesh...') uniqid = uuid.uuid4() file_path = f'/tmp/{uniqid}.ply' save_ply(pc, file_path) set_state('') return file_path, create_gif(pc), gr.update(value=[file_path], visible=True) def prepare_img(img): w, h = img.size if w > h: img = img.crop((w - h) / 2, 0, w - (w - h) / 2, h) else: img = img.crop((0, (h - w) / 2, w, h - (h - w) / 2)) # resize to 256x256 img = img.resize((256, 256)) return img def ply_to_glb(ply_file, glb_file): mesh = trimesh.load(ply_file) # Save the mesh as a glb file using Trimesh mesh.export(glb_file, file_type='glb') return glb_file def save_ply(pc, file_name): # Produce a mesh (with vertex colors) with open(file_name, 'wb') as f: pc.write_ply(f) def create_gif(pc): fig = plt.figure(facecolor='black', figsize=(4, 4)) ax = fig.add_subplot(111, projection='3d', facecolor='black') fixed_bounds=((-0.75, -0.75, -0.75),(0.75, 0.75, 0.75)) # Create an empty list to store the frames frames = [] # Create a loop to generate the frames for the GIF for angle in range(0, 360, 4): # Clear the plot and plot the point cloud ax.clear() color_args = np.stack( [pc.channels["R"], pc.channels["G"], pc.channels["B"]], axis=-1 ) c = pc.coords ax.scatter(c[:, 0], c[:, 1], c[:, 2], c=color_args) # Set the viewpoint for the plot ax.view_init(elev=10, azim=angle) # Turn off the axis labels and ticks ax.axis('off') ax.set_xlim3d(fixed_bounds[0][0], fixed_bounds[1][0]) ax.set_ylim3d(fixed_bounds[0][1], fixed_bounds[1][1]) ax.set_zlim3d(fixed_bounds[0][2], fixed_bounds[1][2]) # Draw the figure to update the image data fig.canvas.draw() # Save the plot as a frame for the GIF frame = np.array(fig.canvas.renderer.buffer_rgba()) w, h = frame.shape[0], frame.shape[1] i = int(round((h - int(h*0.6)) / 2.)) frame = frame[i:i + int(h*0.6),i:i + int(h*0.6)] frames.append(frame) # Save the GIF using imageio imageio.mimsave('/tmp/pointcloud.mp4', frames, fps=30) return '/tmp/pointcloud.mp4' block = gr.Blocks().queue(max_size=250, concurrency_count=6) with block: with gr.Box(): if(not torch.cuda.is_available()): top_description = gr.HTML(f'''

Point-E Web UI


Duplicate Space

Generate 3D Assets in 2 minutes with a prompt or image! Based on the Point-E implementation


There's only one step left before you can train your model: attribute a T4 GPU to it (via the Settings tab) and run the training below. Other GPUs are not compatible for now. You will be billed by the minute from when you activate the GPU until when it is turned it off.

''') else: top_description = gr.HTML(f'''

Point-E Web UI


Duplicate Space

Generate 3D Assets in 2 minutes with a prompt or image! Based on the Point-E implementation

''') with gr.Row(): with gr.Column(): with gr.Tab("Image to 3D"): gr.Markdown("Best results with images of objects on an empty background.") input_image = gr.Image(label="Image") img_button = gr.Button(label="Generate") with gr.Tab("Text to 3D"): gr.Markdown("Uses Stable Diffusion to create an image from the prompt.") prompt = gr.Textbox(label="Prompt", placeholder="A HD photo of a Corgi") text_button = gr.Button(label="Generate") with gr.Accordion("Advanced options", open=False): model = gr.Radio(["base40M", "base300M", "base1B"], label="Model", value="base1B") scale = gr.Slider( label="Guidance Scale", minimum=1.0, maximum=10.0, value=3.0, step=0.1 ) with gr.Column(): model_gif = gr.Video(label="3D Model GIF") model_3d = gr.Model3D(value=None) # btn_pc_to_obj = gr.Button(value="Convert to OBJ", visible=False) file_out = gr.File(label="Files", visible=False) if torch.cuda.is_available(): gr.Examples( examples=[ ["images/pumpkin.png"], ["images/fantasy_world.png"], ], inputs=[input_image], outputs=[model_3d, model_gif, file_out], fn=generate_3D, cache_examples=True ) img_button.click(fn=generate_3D, inputs=[input_image, model, scale], outputs=[model_3d, model_gif, file_out]) text_button.click(fn=generate_3D, inputs=[prompt, model, scale], outputs=[model_3d, model_gif, file_out]) block.launch(show_api=True, share=True)