Autoregressive Visual Generation Needs a Prologue
📰 ArXiv cs.AI
Learn how Prologue bridges the reconstruction-generation gap in autoregressive image generation, enabling better visual token sequence generation
Action Steps
- Implement Prologue by generating a small set of prologue tokens prepended to the visual token sequence
- Train prologue tokens exclusively with the AR cross-entropy loss
- Keep visual tokens dedicated to reconstruction
- Evaluate the performance of Prologue in bridging the reconstruction-generation gap
- Compare the results with existing autoregressive image generation methods
Who Needs to Know This
Computer vision engineers and researchers working on autoregressive image generation models can benefit from this approach to improve their model's performance and efficiency
Key Insight
💡 Prologue generates a small set of prologue tokens to bridge the reconstruction-generation gap, improving autoregressive image generation
Share This
🔍 Introducing Prologue: a novel approach to bridge the reconstruction-generation gap in autoregressive image generation #autoregressive #imagegeneration #Prologue
Key Takeaways
Learn how Prologue bridges the reconstruction-generation gap in autoregressive image generation, enabling better visual token sequence generation
Full Article
Title: Autoregressive Visual Generation Needs a Prologue
Abstract:
arXiv:2605.06137v1 Announce Type: cross Abstract: In this work, we propose Prologue, an approach to bridging the reconstruction-generation gap in autoregressive (AR) image generation. Instead of modifying visual tokens to satisfy both reconstruction and generation, Prologue generates a small set of prologue tokens prepended to the visual token sequence. These prologue tokens are trained exclusively with the AR cross-entropy (CE) loss, while visual tokens remain dedicated to reconstruction. This
Abstract:
arXiv:2605.06137v1 Announce Type: cross Abstract: In this work, we propose Prologue, an approach to bridging the reconstruction-generation gap in autoregressive (AR) image generation. Instead of modifying visual tokens to satisfy both reconstruction and generation, Prologue generates a small set of prologue tokens prepended to the visual token sequence. These prologue tokens are trained exclusively with the AR cross-entropy (CE) loss, while visual tokens remain dedicated to reconstruction. This
DeepCamp AI