Razikus/point-e

1import gradio as gr2import plotly.graph_objects as go3 4import torch5from tqdm.auto import tqdm6 7from point_e.diffusion.configs import DIFFUSION_CONFIGS, diffusion_from_config8from point_e.diffusion.sampler import PointCloudSampler9from point_e.models.download import load_checkpoint10from point_e.models.configs import MODEL_CONFIGS, model_from_config11from point_e.util.plotting import plot_point_cloud12 13device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')14 15print('creating base model...')16base_name = 'base40M-textvec'17base_model = model_from_config(MODEL_CONFIGS[base_name], device)18base_model.eval()19base_diffusion = diffusion_from_config(DIFFUSION_CONFIGS[base_name])20 21print('creating upsample model...')22upsampler_model = model_from_config(MODEL_CONFIGS['upsample'], device)23upsampler_model.eval()24upsampler_diffusion = diffusion_from_config(DIFFUSION_CONFIGS['upsample'])25 26print('downloading base checkpoint...')27base_model.load_state_dict(load_checkpoint(base_name, device))28 29print('downloading upsampler checkpoint...')30upsampler_model.load_state_dict(load_checkpoint('upsample', device))31 32sampler = PointCloudSampler(33    device=device,34    models=[base_model, upsampler_model],35    diffusions=[base_diffusion, upsampler_diffusion],36    num_points=[1024, 4096 - 1024],37    aux_channels=['R', 'G', 'B'],38    guidance_scale=[3.0, 0.0],39    model_kwargs_key_filter=('texts', ''), # Do not condition the upsampler at all40)41 42def inference(prompt):43    samples = None44    for x in sampler.sample_batch_progressive(batch_size=1, model_kwargs=dict(texts=[prompt])):45        samples = x46    pc = sampler.output_to_point_clouds(samples)[0]47    pc = sampler.output_to_point_clouds(samples)[0]48    colors=(238, 75, 43)49    fig = go.Figure(50        data=[51            go.Scatter3d(52                x=pc.coords[:,0], y=pc.coords[:,1], z=pc.coords[:,2], 53                mode='markers',54                marker=dict(55                  size=2,56                  color=['rgb({},{},{})'.format(r,g,b) for r,g,b in zip(pc.channels["R"], pc.channels["G"], pc.channels["B"])],57              )58            )59        ],60        layout=dict(61            scene=dict(62                xaxis=dict(visible=False),63                yaxis=dict(visible=False),64                zaxis=dict(visible=False)65            )66        ),67    )68    return fig69 70demo = gr.Interface(71    fn=inference,72    inputs="text",73    outputs=gr.Plot(),74    examples=[75        ["a red motorcycle"],76        ["a RED pumpkin"],77        ["a yellow rubber duck"]78    ],79    title="Point-E demo: text to 3D",80    description="""Generated 3D Point Cloiuds with [Point-E](https://github.com/openai/point-e/tree/main). This demo uses a small, worse quality text-to-3D model to produce 3D point clouds directly from text descriptions.81    Skip the queue by duplicating this space and upgrading to GPU in settings82<a href="https://huggingface.co/spaces/openai/point-e?duplicate=true"><img src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>83"""84)85demo.queue(max_size=30)86demo.launch(debug=True)87