A SegFormer-B4 model fine-tuned for human parsing with 18 semantic classes, optimized for fashion and virtual try-on applications.

This package provides the exact preprocessing used during model training for maximum accuracy. For quick experimentation, you can also use the model directly via HuggingFace.

pip install fashn-human-parser

On first use, the model weights (~244 MB) are automatically downloaded from HuggingFace Hub and cached locally.

Default cache location:

• Linux/macOS: ~/.cache/huggingface/hub/
• Windows: C:\Users\<username>\.cache\huggingface\hub\

To customize the cache location, set the HF_HOME environment variable before running your script:

export HF_HOME=/path/to/custom/cache  # Linux/macOS
set HF_HOME=C:\path\to\custom\cache   # Windows

from fashn_human_parser import FashnHumanParser

# Initialize parser (automatically uses GPU if available)
parser = FashnHumanParser()

# Run prediction on an image
segmentation = parser.predict("path/to/image.jpg")

# segmentation is a numpy array of shape (H, W) with values 0-17
# representing the 18 semantic classes

The predict method accepts:

• File path (string)
• PIL Image
• NumPy array (RGB, uint8 or float32)

Note: NumPy arrays must be in RGB format. If using cv2.imread() (which returns BGR), convert first:

img = cv2.cvtColor(cv2.imread("image.jpg"), cv2.COLOR_BGR2RGB)

# Process multiple images in a single forward pass
results = parser.predict([image1, image2, image3])
# results is a list of numpy arrays, one per image

logits = parser.predict(image, return_logits=True)
# logits shape: (1, 18, H, W)

from fashn_human_parser import IDS_TO_LABELS, LABELS_TO_IDS

# Get label name from ID
print(IDS_TO_LABELS[3])  # "top"

# Get ID from label name
print(LABELS_TO_IDS["top"])  # 3

• Architecture: SegFormer-B4 (MIT-B4 encoder + MLP decoder)
• Input Size: 384 x 576 (width x height)
• Output: 18-class semantic segmentation mask
• Base Model: nvidia/mit-b4

Ease of use:

• Simple one-line prediction: parser.predict(image)
• Output matches input dimensions - no manual resizing needed
• Returns parsed label IDs directly (not raw logits)
• Accepts multiple input formats (file path, PIL, numpy)
• Batch processing with single forward pass: parser.predict([img1, img2, img3])

Utility exports:

• IDS_TO_LABELS / LABELS_TO_IDS for label mapping
• IDENTITY_LABELS for labels preserved in virtual try-on
• CATEGORY_TO_BODY_COVERAGE for clothing category mappings

Exact preprocessing:

• Uses cv2.INTER_AREA resize (optimal for downsampling)
• Matches the exact preprocessing from model training

The HuggingFace Hub version uses PIL LANCZOS resampling for broader compatibility, which may result in slightly different outputs.

You can also use the model directly via the HuggingFace pipeline:

from transformers import pipeline

pipe = pipeline("image-segmentation", model="fashn-ai/fashn-human-parser")
result = pipe("image.jpg")
# result is a list of dicts with 'label', 'score', 'mask' for each detected class

Note: The pipeline returns per-class masks. This package returns a single segmentation map with class IDs.

This model inherits the NVIDIA Source Code License for SegFormer. Please review the license terms before use.

• FASHN AI
• HuggingFace Model
• HuggingFace Space (Demo)
• GitHub Repository