aoiandroid/resemble-enhance-fp16-quantized

Resemble Enhance FP16 Quantized

FP16 (half-precision) quantized version of Resemble Enhance for mobile deployment.

Model Information

• Original Model: ResembleAI/resemble-enhance
• Quantization: FP16 (half precision)
• Size Reduction: 50% (from FP32)
• Parameters: 356,414,076
• Model Size: 679.81 MB

Usage

This FP16 quantized model is optimized for:

• iOS devices: Compatible with Apple Neural Engine (ANE)
• Mobile deployment: Reduced memory footprint
• Faster inference: 2-3x faster than FP32 on supported hardware

Loading the Model

import torch

# Load FP16 state dict
state_dict = torch.load("mp_rank_00_model_states_fp16.pt", map_location="cpu")

# Load into model (model must be converted to FP16 first)
model = YourResembleEnhanceModel()
model = model.half()  # Convert to FP16
model.load_state_dict(state_dict)
model.eval()

Conversion to CoreML (iOS)

For iOS deployment, convert to CoreML:

import coremltools as ct

# Convert PyTorch model to CoreML
mlmodel = ct.convert(
    model,
    inputs=[ct.TensorType(name="input", shape=input_shape)],
    minimum_deployment_target=ct.target.iOS16
)

# Save as .mlmodel
mlmodel.save("ResembleEnhanceFP16.mlmodel")

Performance

• Size: 679.81 MB (50% reduction from FP32)
• Inference Speed: 2-3x faster on Apple Neural Engine
• Quality: Minimal perceptual loss compared to FP32

Original Model

This is a quantized version of ResembleAI/resemble-enhance.

For more information about the original model, please refer to the original repository.

License

This model follows the same license as the original Resemble Enhance model.