diff --git a/README.md b/README.md
index 995810e..8babee0 100644
--- a/README.md
+++ b/README.md
@@ -101,6 +101,75 @@ pip install -e .
 
 For details, see [PromptHMR README](third_party/PromptHMR/README.md).
 
+## Data Download and Setup
+
+After installing the environments, you need to download the required model files and data.
+
+### Download PromptHMR Data
+
+From the `third_party/PromptHMR` directory:
+
+```bash
+cd third_party/PromptHMR
+bash scripts/fetch_data.sh
+```
+
+This script downloads:
+- PromptHMR checkpoints
+- Third-party checkpoints (camcalib, droidcalib, vitpose, SAM, SAM2, YOLO, Detectron2)
+- Dataset annotations (for evaluation)
+- Example images
+
+### Download SMPL-X and SMPL Body Models
+
+From the `third_party/PromptHMR` directory:
+
+```bash
+cd third_party/PromptHMR
+bash scripts/fetch_smplx.sh
+```
+
+This script requires registration at:
+- [SMPL-X](https://smpl-x.is.tue.mpg.de) - for SMPL-X models
+- [SMPL](https://smpl.is.tue.mpg.de) - for SMPL models
+
+The script automatically downloads, extracts, and reorganizes the models into the correct directory structure:
+- SMPL-X models → `data/body_models/smplx/`
+- SMPL models → `data/body_models/smpl/`
+- Accessory files → `data/body_models/accessory/`
+
+### Convert SMPL-X Models to Float32 (Required)
+
+**Important:** The downloaded SMPL-X models use `float64` (double precision) which can cause compatibility issues with PyTorch and other frameworks. **You must convert them to `float32` (single precision) after downloading.** This also reduces file size by ~50%.
+
+Convert downloaded models to float32:
+
+```bash
+# From the project root directory
+python scripts/convert_smplx_to_float32.py \
+    --input_dir third_party/PromptHMR/data/body_models/smplx \
+    --output_dir third_party/PromptHMR/data/body_models/smplx
+```
+
+Or convert in-place (overwrites originals - backup recommended):
+
+```bash
+# Backup before converting
+cp -r third_party/PromptHMR/data/body_models/smplx third_party/PromptHMR/data/body_models/smplx_backup
+
+# Convert in-place
+python scripts/convert_smplx_to_float32.py \
+    --input_dir third_party/PromptHMR/data/body_models/smplx \
+    --in_place
+```
+
+The conversion script:
+- Converts all `.npz` files from `float64` to `float32`
+- Converts `int64` to `int32` and `uint64` to `uint32` for consistency
+- Handles object arrays and nested structures
+- Creates compressed output files
+- Preserves original files (unless using `--in_place`)
+
 ## Usage
 
 > **Note**: Scripts automatically switch to the appropriate conda environment (`gmr` or `phmr`) as needed. Just ensure both environments are installed - no need to manually activate them.
diff --git a/scripts/convert_smplx_to_float32.py b/scripts/convert_smplx_to_float32.py
new file mode 100755
index 0000000..8075240
--- /dev/null
+++ b/scripts/convert_smplx_to_float32.py
@@ -0,0 +1,225 @@
+#!/usr/bin/env python3
+"""
+Convert SMPL-X models from float64 to float32 format.
+
+This script converts SMPL-X .npz model files from float64 (double precision) to 
+float32 (single precision), which:
+- Reduces file size by approximately 50%
+- Improves compatibility with PyTorch and other frameworks
+- Maintains sufficient precision for most use cases
+
+Usage:
+    python scripts/convert_smplx_to_float32.py \
+        --input_dir third_party/PromptHMR/data/body_models/smplx \
+        --output_dir third_party/PromptHMR/data/body_models/smplx
+
+    # Convert in-place (overwrites original files)
+    python scripts/convert_smplx_to_float32.py \
+        --input_dir third_party/PromptHMR/data/body_models/smplx \
+        --in_place
+
+    # Backup originals before converting
+    python scripts/convert_smplx_to_float32.py \
+        --input_dir third_party/PromptHMR/data/body_models/smplx \
+        --output_dir third_party/PromptHMR/data/body_models/smplx \
+        --backup
+"""
+import os
+import os.path as osp
+import argparse
+import numpy as np
+import shutil
+from pathlib import Path
+
+
+def convert_npz_to_float32(in_path, out_path):
+    """
+    Convert an NPZ file from float64 to float32 format.
+    
+    Args:
+        in_path: Input NPZ file path
+        out_path: Output NPZ file path
+    """
+    if not osp.exists(in_path):
+        print(f"Warning: Input file {in_path} does not exist, skipping")
+        return False
+    
+    print(f"Processing {osp.basename(in_path)} -> {osp.basename(out_path)}")
+    in_npz = np.load(in_path, allow_pickle=True)
+    out_dict = {}
+    converted_count = 0
+    
+    for key, val in in_npz.items():
+        # Handle special cases for UV and face texture indices
+        if key in ["vt", "ft"]:
+            if isinstance(val, np.ndarray):
+                if val.dtype == object:
+                    # Try to convert object arrays
+                    try:
+                        val = np.array(val.tolist(), dtype=np.uint32)
+                        converted_count += 1
+                    except Exception as e:
+                        print(f"  Warning: Could not convert {key} from object: {e}")
+                        continue
+                elif val.dtype == np.float64:
+                    val = np.float32(val)
+                    converted_count += 1
+                elif val.dtype == np.int64:
+                    val = np.int32(val)
+                    converted_count += 1
+            out_dict[key] = val
+            continue
+        
+        # Convert numpy arrays
+        if isinstance(val, np.ndarray):
+            original_dtype = val.dtype
+            if val.dtype == np.float64:
+                val = np.float32(val)
+                converted_count += 1
+            elif val.dtype == np.int64:
+                val = np.int32(val)
+                converted_count += 1
+            elif val.dtype == np.uint64:
+                val = np.uint32(val)
+                converted_count += 1
+            elif val.dtype == object:
+                # Try to handle object arrays
+                try:
+                    # For most object arrays in SMPLX, they're usually lists/arrays
+                    val = np.array(val.tolist())
+                    if val.dtype == np.float64:
+                        val = np.float32(val)
+                        converted_count += 1
+                    elif val.dtype == np.int64:
+                        val = np.int32(val)
+                        converted_count += 1
+                except Exception as e:
+                    print(f"  Warning: Could not convert {key} from object: {e}")
+                    # Skip problematic keys
+                    continue
+            out_dict[key] = val
+        else:
+            # Handle non-array values (scalars, strings, etc.)
+            if isinstance(val, (float, np.floating)) and val.dtype == np.float64:
+                out_dict[key] = np.float32(val)
+                converted_count += 1
+            else:
+                out_dict[key] = val
+    
+    # Save as compressed NPZ
+    np.savez_compressed(out_path, **out_dict)
+    
+    # Get file sizes
+    in_size = osp.getsize(in_path) / (1024 * 1024)  # MB
+    out_size = osp.getsize(out_path) / (1024 * 1024)  # MB
+    reduction = ((in_size - out_size) / in_size * 100) if in_size > 0 else 0
+    
+    print(f"  ✓ Converted {converted_count} arrays/values")
+    print(f"  ✓ Size: {in_size:.2f} MB -> {out_size:.2f} MB ({reduction:.1f}% reduction)")
+    print(f"  ✓ Saved to {out_path}\n")
+    return True
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Convert SMPL-X models from float64 to float32 format",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog=__doc__
+    )
+    parser.add_argument(
+        "--input_dir",
+        type=str,
+        required=True,
+        help="Input directory containing SMPL-X .npz files"
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default=None,
+        help="Output directory (default: same as input_dir if --in_place not set)"
+    )
+    parser.add_argument(
+        "--in_place",
+        action="store_true",
+        help="Convert files in-place (overwrites original files)"
+    )
+    parser.add_argument(
+        "--backup",
+        action="store_true",
+        help="Create backup of original files before converting (only with --in_place)"
+    )
+    parser.add_argument(
+        "--pattern",
+        type=str,
+        default="*.npz",
+        help="File pattern to match (default: *.npz)"
+    )
+    
+    args = parser.parse_args()
+    
+    input_dir = Path(args.input_dir).resolve()
+    if not input_dir.exists():
+        print(f"Error: Input directory does not exist: {input_dir}")
+        return 1
+    
+    # Determine output directory
+    if args.in_place:
+        output_dir = input_dir
+        if args.backup:
+            backup_dir = input_dir / "backup_float64"
+            backup_dir.mkdir(exist_ok=True)
+            print(f"Backup directory: {backup_dir}")
+    else:
+        output_dir = Path(args.output_dir).resolve() if args.output_dir else input_dir
+        output_dir.mkdir(parents=True, exist_ok=True)
+    
+    # Find all matching files
+    npz_files = list(input_dir.glob(args.pattern))
+    if not npz_files:
+        print(f"No {args.pattern} files found in {input_dir}")
+        return 1
+    
+    print("=" * 60)
+    print("SMPL-X Float32 Conversion")
+    print("=" * 60)
+    print(f"Input directory: {input_dir}")
+    print(f"Output directory: {output_dir}")
+    print(f"Files to convert: {len(npz_files)}")
+    if args.in_place and args.backup:
+        print(f"Backup directory: {backup_dir}")
+    print("=" * 60)
+    print()
+    
+    converted = 0
+    for npz_file in npz_files:
+        if args.in_place:
+            # Create backup if requested
+            if args.backup:
+                backup_path = backup_dir / npz_file.name
+                shutil.copy2(npz_file, backup_path)
+                print(f"Backed up: {npz_file.name} -> backup_float64/{npz_file.name}")
+            
+            # Convert in-place (use temporary file to avoid corruption)
+            temp_path = npz_file.with_suffix('.npz.tmp')
+            if convert_npz_to_float32(npz_file, temp_path):
+                shutil.move(temp_path, npz_file)
+                converted += 1
+        else:
+            # Convert to output directory
+            output_path = output_dir / npz_file.name
+            if convert_npz_to_float32(npz_file, output_path):
+                converted += 1
+    
+    print("=" * 60)
+    print(f"Conversion complete! Converted {converted}/{len(npz_files)} files")
+    print("=" * 60)
+    
+    if args.in_place and args.backup:
+        print(f"\nOriginal files backed up to: {backup_dir}")
+    
+    return 0
+
+
+if __name__ == "__main__":
+    exit(main())
+