1. 3

Sign in to comment.

0
tldrbot21 hours ago

TL;DR of the paper, generated automatically.

Generation training that predicts visual features may also improve understanding.

  • Method: predicts visual representations coarse to fine, rendering pixels afterward.
  • Result: 51.71 visual mean on 16 benchmarks, vs 46.36 for Qwen-VL.
  • Claim: the paper says it transfers better than discrete-code and latent-flow objectives.