Summary of Stretching Each Dollar: Diffusion Training From Scratch on a Micro-budget, by Vikash Sehwag et al.
Stretching Each Dollar: Diffusion Training from Scratch on a Micro-Budget
by Vikash Sehwag, Xianghao Kong, Jingtao Li, Michael Spranger, Lingjuan Lyu
First submitted to arxiv on: 22 Jul 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
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 The paper proposes a low-cost method for training large-scale text-to-image (T2I) generative transformer models. To reduce computational costs, the authors introduce a deferred masking strategy that preprocesses image patches before randomly masking up to 75% of them during training. This approach improves model performance and reduces the need for model downscaling. The authors also incorporate recent advancements in transformer architecture, such as mixture-of-experts layers, to enhance performance. Notably, they train a 1.16 billion parameter sparse transformer using only publicly available real and synthetic images at an economical cost of $1,890, achieving a FID score of 12.7 on the COCO dataset. This is comparable to state-of-the-art models trained with significantly more resources. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper shows how to make large-scale text-to-image models work better and cheaper. It uses a new way of training called “deferred masking” that helps reduce the cost of making these models. The authors also use the latest improvements in transformer technology, like mixture-of-experts layers, to make their model better. They show that they can train a big model using only public images at a low cost of $1,890 and get good results. |
Keywords
* Artificial intelligence * Mixture of experts * Transformer
