athirajan/bitmind-v6-convnext-0.9748

BitMind V6 ConvNeXt-Large Image Detector

This model is optimized for the BitMind Subnet (SN34) Gasbench V6 competition.

Performance

• V6 Score: 0.9748
• Validation MCC: 0.9942
• Binary MCC: 0.9915 (Real vs Fake)

Architecture

• Base Model: ConvNeXt-Large-384
• Training: Robust training pipeline with class balancing and V6-specific augmentations
• Dataset: 80k+ images (Real, Synthetic, Semi-synthetic)

Usage

This model accepts images as uint8 tensors [batch_size, 3, 384, 384] with values in [0-255] and outputs logits for 3 classes: 0. Real 1. Synthetic 2. Semi-synthetic--- {}