Update app.py

app.py

@@ -18,7 +18,7 @@ except Exception as e:
 if torch.cuda.is_available():
     device = "cuda"
     gpu_type = torch.cuda.get_device_name(torch.cuda.current_device())
-    power_device = f"{gpu_type} GPU"
+    power_device = f"{gpu_type}"
     torch.cuda.max_memory_allocated(device=device)
 else:
     device = "cpu"
@@ -36,7 +36,7 @@ torch.backends.cudnn.allow_tf32 = True
 
 MAX_SEED = np.iinfo(np.int32).max
 
-FM_MODEL_ID = 'EPFL-VILAB/4M-21_B'
+FM_MODEL_ID = 'EPFL-VILAB/4M-21_XL'
 MODEL_NAME = FM_MODEL_ID.split('/')[1].replace('_', ' ')
 
 # Human poses visualization is disabled, since it needs SMPL weights. To enable human pose prediction and rendering:
@@ -111,9 +111,7 @@ with gr.Blocks(css=css, theme=gr.themes.Base()) as demo:
                 [`Website`](https://4m.epfl.ch) | [`GitHub`](https://github.com/apple/ml-4m) <br>[`4M Paper (NeurIPS'23)`](https://arxiv.org/abs/2312.06647) | [`4M-21 Paper (arXiv'24)`](https://arxiv.org/abs/2406.09406)
 
                 This demo predicts all modalities from a given RGB input, using [{FM_MODEL_ID}](https://huggingface.co/{FM_MODEL_ID}), running on *{power_device}*.
-                For more generative examples, and to enable human pose visualizations, please see our [GitHub repo](https://github.com/apple/ml-4m).
-                
-                (Disclaimer: The demo is a work in progress. We will switch it to using 4M-21 XL when running on GPU. Until then, this space runs on CPU and takes several minutes for inference.)
+                For more generative examples, and to enable human pose visualizations, please see our [GitHub repo](https://github.com/apple/ml-4m).                
                 """)
                 
                 img_path = gr.Image(label='RGB input image', type='filepath')