Claude 3 Haiku processes images in 1.6k tokens. It corresponds to 40x40 patches, and I would guess patches of 8x8 using a traditional VQGAN so image input at 320x320 px which seems reasonable. https://t.co/1ccNyBsHvQ https://t.co/EhWp6mCNL4
— Boris Dayma 🖍️ (@borisdayma) Mar 13, 2024
from Twitter https://twitter.com/borisdayma
March 13, 2024 at 10:27PM
via IFTTT
No comments:
Post a Comment