HiDream

HiDream is an open-source, foundation-scale text-to-image model developed by HiDream.ai (Sparking Innovations Limited, Hong Kong) and served through the vivago.ai creative platform. The flagship release, HiDream-I1, is a 17B-parameter diffusion transformer released under the permissive MIT license, which makes it one of the very few frontier-quality image generators that can be self-hosted and used commercially without royalty or content-usage restrictions. It ships in three variants tuned for a quality-versus-speed trade-off: 'Full' targets maximum fidelity at around 50 denoising steps, 'Dev' balances quality and throughput at ~30 steps, and 'Fast' produces usable images in ~16 steps for interactive or high-volume workflows. In independent evaluations HiDream-I1 has posted top-tier GenEval and DPG-Bench scores, beating open-source peers like FLUX, AuraFlow and SD 3.5 and matching or exceeding several closed models such as Midjourney v7, Ideogram v3 and Reve. On vivago.ai, the model powers a browser-based creator suite with prompt-to-image generation, style presets, image editing, upscaling and a community feed; the same account also gives access to third-party video and image models Vivago aggregates. The weights and inference code are on GitHub and Hugging Face and are supported natively in the Hugging Face diffusers library, so teams that need on-prem or private-cloud deployment can run it themselves; those who don't want to manage GPUs can call it via Vivago's hosted API or through providers like Fal and Runware. Typical users are indie designers and marketers who want strong out-of-the-box results without Midjourney's guardrails, plus ML teams that need a commercially clean base model to fine-tune for product photography, concept art, editorial illustration or synthetic-data pipelines.