Machine Heart

Jul 13, 2026 · Artificial Intelligence

PRA Beats 1.9B Baseline Without a Visual Tokenizer: 135M Model Surpasses Large-Scale Pixel‑Space AR

The paper introduces Parallel Rollout Approximation (PRA), a pixel‑space autoregressive image generation approach that eliminates visual tokenizers, reduces high‑dimensional prediction difficulty with a low‑dimensional intermediate state, and uses parallel rollout approximation to mitigate training‑inference mismatch, achieving FID 2.58 with 135 M parameters—outperforming a 1.9 B‑parameter baseline—and demonstrating strong visual representation learning.