app.py

import gradio as gr
import spaces
import torch
import torchvision.transforms
import numpy as np
from transformers import AutoModel
from theia.decoding import load_feature_stats, prepare_depth_decoder, prepare_mask_generator, decode_everything

@spaces.GPU
def run_theia(image):
    theia_model = AutoModel.from_pretrained("theaiinstitute/theia-base-patch16-224-cdiv", trust_remote_code=True)
    theia_model = theia_model.to('cuda')
    target_model_names = [
        "google/vit-huge-patch14-224-in21k",
        "facebook/dinov2-large",
        "openai/clip-vit-large-patch14",
        "facebook/sam-vit-huge",
        "LiheYoung/depth-anything-large-hf",
    ]
    feature_means, feature_vars = load_feature_stats(target_model_names, stat_file_root="feature_stats")
    
    mask_generator, sam_model = prepare_mask_generator('cuda')
    depth_anything_model_name = "LiheYoung/depth-anything-large-hf"
    depth_anything_decoder, _ = prepare_depth_decoder(depth_anything_model_name, 'cuda')

    image = torch.transforms.v2.Resize(size=(224, 224))(image)
    images = [image]
    
    theia_decode_results, gt_decode_results = decode_everything(
        theia_model=theia_model,
        feature_means=feature_means,
        feature_vars=feature_vars,
        images=images,
        mask_generator=mask_generator,
        sam_model=sam_model,
        depth_anything_decoder=depth_anything_decoder,
        pred_iou_thresh=0.5,
        stability_score_thresh=0.7,
        gt=True,
        device='cuda',
    )
    
    vis_video = np.stack(
        [np.vstack([tr, gtr]) for tr, gtr in zip(theia_decode_results, gt_decode_results, strict=False)]
    )

    return vis_video

demo = gr.Interface(fn=run_theia, inputs=gr.Image(type="pil"), outputs=gr.Image(type="numpy"))
demo.launch()