88Z AI, 20 May — HiDream has open-sourced its flagship image generation model, HiDream-O1-Image-Pro. Built on the UiT native multimodal architecture, it maps pixels, text, and task conditions into a shared token space, eliminating traditional VAEs and separate encoders. The model achieves state-of-the-art results on benchmarks like GenEval and DPG, ranking eighth in the Artificial Analysis Arena and surpassing open-source models such as FLUX.2[Dev].
At a glance
- Uses UiT architecture to unify multimodal data mapping, discarding traditional VAEs
- Sets new SOTA on benchmarks and outranks FLUX.2[Dev] in the arena
The update
HiDream has officially open-sourced its flagship image generation model, HiDream-O1-Image-Pro. Built on the innovative UiT native multimodal architecture, the model's core breakthrough lies in mapping image pixels, text information, and task conditions into a shared token space. This design completely eliminates traditional variational autoencoders (VAEs) and separate encoder structures, enabling more efficient modal fusion. In terms of performance, HiDream-O1-Image-Pro has set new state-of-the-art records on multiple authoritative benchmarks, including GenEval and DPG. Furthermore, the model ranks eighth in the Artificial Analysis Arena, demonstrating strong competitiveness. Its performance has surpassed well-known open-source models such as FLUX.2[Dev], marking a new advancement in the field of open-source image generation.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- HiDream智象未来Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.