Clark Ai

• 

  a dog

• 

  a panda

• 

  FP16 (3.2GB) vs Clark Air (0.5 GB)

For this hackathon, we compressed a Sana-style 1.6B text-to-image model into a low-bit ternary image generator. The current demo checkpoint compresses roughly 94.6% of the image transformer while preserving only 5.4% of precision-sensitive tensors. In standard packed accounting, this gives about 3.2 effective bits per transformer weight, or roughly a 5× transformer-size reduction versus FP16!!!

From 3.2GB to 0.5GB with quite good quality! The goal is to match Prism-ML level of 1bit compression.

Built With

• claude
• glm-5.2
• opus-4.8