Arena.ai
@arena
New provider HiDream-01-Image by @HiDream_AI ranks #27 overall in the Text-to-Image Arena, making it the #4 open source model. Congrats to the @HiDream_AI team on the release! https://t.co/9ptkOh9cQu
4retweets30likes
View on XHiDream-01-Image Ranks as Top Four Open Source Model in Arena
Arena.ai, a community-driven platform for blind AI model evaluation, ranked the new HiDream-01-Image model at #27 overall in its Text-to-Image Arena. Developed by HiDream.ai, the model secured the #4 position among open-source entries. This follows the fal launch of HiDream-01-Image, which introduced a unified transformer architecture that processes pixels and text in one space.
- Overall rank
- #27
- Open source rank
- #4
- Total models in Arena
- 60
- Total community votes
- 5,169,154
- Elo score
- 1121±9
The results highlight a trend where open-source models challenge proprietary systems. While the leaderboard is dominated by closed-source models like gpt-image-2 and gemini-3.1-flash, HiDream's entry provides a high-performance alternative for users who prioritize open weights. This performance mirrors Arena's ranking of Kimi K2.6 as a top-tier open model.
You can now access HiDream-01-Image through inference providers, matching the pattern of fal's FLUX Erase integration for specialized image editing. The model is effective for tasks requiring high subject consistency and complex layouts. As the leaderboard expands with 60 models, these community rankings offer a verified way to select the most capable generator for creative workflows.
Still wondering? A few quick answers below.
HiDream-01-Image is a text-to-image foundation model developed by the AI lab HiDream.ai. It uses a pixel-level unified transformer architecture, which allows the system to process text and image data within a single token space. This design is intended to improve subject consistency and layout accuracy compared to traditional image generation methods.
The model debuted at number 27 overall out of 60 models on the Arena Text-to-Image leaderboard. Among open-source and open-weight models, it secured the number four position. These rankings are determined by over five million community votes through blind human preference testing, where users compare generated images without knowing which model produced them.
Yes, HiDream-01-Image is an open-weight model, meaning its internal parameters are publicly available for download and local deployment. This allows developers and researchers to run the model on their own hardware or integrate it into custom applications without relying on a proprietary API, distinguishing it from closed systems like those from OpenAI.
The top of the Arena leaderboard is currently dominated by proprietary models, including OpenAI's gpt-image-2 and Google's gemini-3.1-flash-image-preview. While HiDream-01-Image is the fourth-highest open model, it still trails these closed-source frontier systems in overall human preference scores, though it remains highly competitive within the broader field of 60 evaluated models.
Users can access the model through various inference providers, such as the fal platform, which hosted its initial release. Because it is an open-weight model, technical users can also download the weights from repositories like Hugging Face to run the system locally for high-fidelity image generation, editing, and subject personalization tasks.
