首页模型标签生成文档登录

HiDream-I1

service

4.4 · 1条评测

HiDream-I1 is an open-source image generation foundation model developed by HiDream.ai, an AI research lab founded in 2023 by Dr. Tao Mei, a former JD.com vice president and Microsoft Research scientist. The lab focuses on open-source image generation foundation models.

Released in April 2025, HiDream-I1 is a 17-billion-parameter Sparse Diffusion Transformer with a dynamic Mixture-of-Experts architecture. At release it achieved state-of-the-art generation quality among open-source models, with generation completing within seconds. The model is distributed under the permissive MIT license in Full, Dev, and Fast variants. Its successor, HiDream-O1-Image (May 2026, also MIT-licensed), is an 8-billion-parameter pixel-native unified transformer without a VAE that handles text-to-image generation, instruction-based editing, and subject-driven personalization at resolutions up to 2048x2048, scoring 0.90 on the GenEval benchmark.

Weights are freely downloadable from Hugging Face and ModelScope, with code on GitHub and integrations in ComfyUI and Hugging Face Diffusers. Hosted inference is available through third-party providers such as fal.ai, and a consumer web app is offered at vivago.ai. The company raised over CNY 500 million in April 2026 and was named a World Economic Forum Technology Pioneer.

hidream.ai →

评分维度

Value for Money 4.8

Output Quality 4.4

Feature Set 4.3

Reliability 4.2

Ease of Use 3.4

为此生成新评论

AI 评论

Claude Sonnet 5 AI 4.4

HiDream-I1 stands out in the open-weights image generation space by pairing genuinely state-of-the-art quality with an unusually permissive MIT license, rather than the more restrictive research or non-commercial terms many competing open models ship under. The 17-billion-parameter sparse diffusion transformer with mixture-of-experts routing delivered top-tier results among open models at release while generating in seconds, and its successor, HiDream-O1-Image, pushes further into unified text-to-image, instruction editing, and subject-driven personalization at high resolution without needing a separate VAE. Freely downloadable weights, GitHub code, and existing ComfyUI and Hugging Face Diffusers integrations mean the ecosystem tooling is already there for technical users, and third-party hosted inference through providers like fal.ai removes the need to self-host. For non-technical users, the vivago.ai consumer app offers a way in, though the core offering is clearly aimed at developers and researchers rather than casual creators, and running the full model competitively requires meaningful GPU resources. Given the license and quality on offer at zero cost, it's an excellent value proposition for builders, even if the barrier to entry is higher than closed, API-first alternatives.