Summary of Lantern: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding, by Doohyuk Jang et al.
LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding
by Doohyuk Jang, Sihwan Park, June Yong Yang, Yeonsung Jung, Jihun Yun, Souvik Kundu, Sung-Yub Kim, Eunho Yang
First submitted to arxiv on: 4 Oct 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper proposes a novel approach to accelerate auto-regressive (AR) models in image generation. AR models have gained popularity, often matching or surpassing diffusion models’ performance. However, their sequential nature slows down generation compared to other methods like GANs or diffusion-based approaches that operate more efficiently. To address this challenge, the authors identify “token selection ambiguity” where visual AR models frequently assign low probabilities to tokens, hampering speculative decoding’s effectiveness. They propose a relaxed acceptance condition called LANTERN that leverages token interchangeability in latent space, enabling more flexible use of candidate tokens and restoring speculative decoding’s efficiency. Additionally, they incorporate a total variation distance bound to ensure image quality and semantic coherence are not compromised. Experimental results demonstrate the efficacy of their method, providing significant speed-ups over state-of-the-art speculative decoding. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper makes it possible for computer models to create images more quickly. Right now, these models can only add one pixel at a time, which is slow. The authors found that this slow process was caused by the model having trouble choosing what to draw next. They created a new way to help the model make decisions faster and better. This new method, called LANTERN, allows the model to look at different possibilities for what to draw next and choose the best one more quickly. The results show that this new method works well and can speed up the process by 1.75 times compared to other methods. |
Keywords
» Artificial intelligence » Diffusion » Image generation » Latent space » Token
