Move fast and make things: New HART AI generates images 5 times quicker than DALL-E, Imagen 3

New AI text-to-image generator HART delivers unprecedented speed, creating images in just 1.8 seconds—over 5 times faster than leading diffusion models like DALL-E and Imagen 3. Developed by MIT, Nvidia, and Tsinghua University, HART uses an innovative autoregressive approach that generates images step-by-step rather than through the diffusion process used by most popular generators. This breakthrough demonstrates how different AI training methodologies can dramatically impact performance, potentially setting a new standard for real-time image generation applications.

The big picture: HART (Hybrid Autoregressive Transformer) represents a significant leap forward in AI image generation speed while maintaining comparable image quality to slower competitors.

• The model can generate images with 3.1 to 5.9 times lower latency than state-of-the-art diffusion models, completing generations in about 1.8 seconds—roughly the time it takes to say “Mississippi.”
• When tested against OpenAI‘s GPT-4o and Google‘s Imagen 3 using identical prompts, HART generated images 58 times faster than GPT-4o (which took 1 minute 45 seconds) and 5.5 times faster than Imagen 3 (which took about 10 seconds).

How it works: Unlike most popular text-to-image generators that use diffusion models, HART employs an autoregressive (AR) approach similar to OpenAI’s recently released GPT-4o image generator.

• AR models offer more precise control by generating images sequentially, but traditionally face challenges with training costs and quality at higher resolutions.
• Researchers overcame these limitations by developing a hybrid tokenizer that processes different parts of the image more efficiently, resulting in both increased speed and higher throughput.

Quality comparison: While HART’s primary advantage is speed, the quality of its outputs remains competitive with leading image generators.

• In side-by-side comparisons, Google’s Imagen 3 delivered the best balance of speed and quality, though it still took approximately 10 times longer than HART to generate comparable images.
• According to the article’s author, who has tested most text-to-image models on the market, HART is definitively the quickest available option.

Why this matters: As AI image generation becomes more integrated into creative workflows and applications, the dramatic reduction in generation time could enable new real-time use cases and significantly improve user experience.

Open access: The model is available for free public use and its inference code has been open-sourced through a public GitHub repository, making it accessible to developers, researchers, and AI enthusiasts for further experimentation.