Text Diffusion
summary
Claims that text diffusion cuts latency by doing block denoising (example: 24 denoising steps to generate 256 tokens = roughly 10x fewer memory transfers than autoregression), at the cost of lower throughput for large batches. Concrete behavior: a Gemini Diffusion demo yields 60 then 49 then corrects to 39 on a math chain, while GPT-4o and Gemini 2.5 Flash misreport and one sticks to 42. Low-latency demos work: a fake Wikipedia, Reddit clone, OS UI, and a todo app built in 15 seconds by voice.
AI-written summary · about
chapters (11)
- 0:00Introduction to Text Diffusion
- 1:02How Text Diffusion Works (Training and Inference)
- 2:06Gemini Diffusion Research Preview
- 3:04Difference Between Autoregressive and Diffusion Models
- 4:02Pros and Cons of Text Diffusion
- 6:13Hardware Efficiency: Why Text Diffusion is Faster
- 8:47Bidirectional Reasoning and Self-Correction
- 12:00Dynamic and Adaptive Computation
- 14:26In-place Text Editing
- 16:09Low Latency Applications and Demos
- 20:05Q&A Session
related talks
01h 02mAdvancing Diffusion Models for Text Generation
45mDiffusion to Prod: vLLM-Omni Unifies Multimodal Inference
20mRealtime Video AI with Diffusion Models
40mDiffusion Large Language Models
37mShort Industry Talks pt.2
21mBuilding at the Speed of Thought
53mGenerative AI: Overview and use cases
39mDemocratizing Diffusion Models with Diffusers