Prologue · interactive demo

Prologue: a small set of token positions at the front of an autoregressive sequence that are trained only with cross-entropy loss; the rest of the sequence is the usual reconstruction-trained visual tokens.

This demo shows the qualitative property that motivates the method: the prologue prefix carries class identity and global layout, while the visual suffix carries texture and fine detail.

1. Pick an ImageNet class and click Resample all to draw a fresh sample.
2. Click Resample visual only to keep the prologue tokens from step 1 and redraw only the visual tokens. Class and layout stay, texture varies.

Backbone: Prologue-L-XL (685M AR + L-tokenizer); paper gFID = 1.46 with CFG.