360 AI Lab's FG-CLIP: The Vision Model That Sees What Others Miss

Dual-Stage Training Protocol

Unlike traditional CLIP's single-phase training, FG-CLIP adopts a two-stage strategy:              

1) Global Contrast Learning for initial image-text alignment              

2) Region-Text Contrast Learning using RoIAlign-extracted features              

This hybrid approach reduces false matches in complex scenes by 38%, as validated in MIT's Open Vocabulary Detection Benchmark.

Hard Negative Sample Mining

The model introduces semantic-boundary negative samples - text descriptions with subtle attribute changes (e.g., "light brown stool" vs "dark brown chair"). Trained on 12 million synthetic negative pairs, FG-CLIP achieves 89% precision in distinguishing visually similar objects, outperforming Google's SIGLIP by 15%.

?? Long-Text Comprehension

FG-CLIP processes 512-token descriptions (6.6x CLIP's capacity), enabling analysis of complex prompts like: 

"A Ming-style porcelain vase with crackled glaze, 32cm tall, displayed beside Renaissance oil paintings" 

In ArtGen-2025 test, it achieved 91% accuracy vs CLIP's 63% in multi-element scene understanding.

?? Microscopic Feature Matching

The OmniParser-v2 module combines visual saliency maps with text semantics, detecting sub-millimeter defects in industrial inspections. Partnering with BOE Technology, 360 reduced LCD panel quality control errors by 72% in pilot deployments.