While the generator market text-to-image is gradually becoming denser, one question remains: which tool offers the best performance? The Artificial Analysis organization, which provides “benchmarking and objective insights to help developers, consumers, researchers and users of AI models”offers an answer to this question with Text to Image Arena. A ranking, similar to the Chatbot Arena, that objectively evaluates AI-powered image generation models based on user contributions. On the organization's website, visitors are invited to compare two models whose identity is hidden from them, evaluating the one that most accurately answers the query.
Top 10 Best Performing Image Generation Models in August 2024
In August 2024, surprise: the ranking established by Artificial Analysis is dominated by a new player. Black Forest Labs, a startup founded by the creators of Stable Diffusion, directly takes first place thanks to the most advanced version of FLUX.1, a generative model unveiled on August 1, 2024. In less than a month, this technology, which powers the controversial generator of the Grok conversational agent designed by X (Twitter), has firmly established itself in this ranking, since two other iterations are placed in the top 10 (4th and 7th).
The rest of the podium is occupied by the latest model deployed by Ideogram and version 6.1 of Midjourney, which recently broke away from Discord by launching its own website. In the rest of the ranking, Stability.ai remains in the top 10 (6th), as does Playground AI (8th). A leader in text generation, OpenAI is significantly behind the competition in image generation, with its technologies powering DALL-E only ranking 9th and 10th.
Here are the top 10 performing image generation models as of August 2024:
- FLUX.1 (pro): 1,151 (Elo score)
- Ideogram v2: 1,130
- Midjourney v6.1: 1,126
- FLUX.1 (dev): 1,118
- Midjourney v6: 1 112
- Stable Diffusion 3 Large: 1,098
- FLUX.1 (fast): 1,091
- Playground v2.5: 1,046
- SLAB 3 HD: 1,041
- SLAB 3: 1,023
The Text to Image Arena ranking criteria
Like Chatbot Arena, Text to Image Arena is based on a duel system. The principle is as follows: users are invited to decide between two anonymized models, by designating the one that best responds to the initial request. Each model then receives an Elo score, which evolves according to its performance. For example, if model A suffers a defeat against model B whose Elo score is lower, it loses points, and vice versa. According to Artificial Analysis, the ranking is updated every hour.
