中文EN RSS

MiniMax 开源 SOTA 级图像模型 H3-ImageMiniMax open-sources SOTA image model H3-Image


AI free token/AI免费积分
MiniMax 开源 SOTA 级图像模型 H3-Image

拓展智能体视觉创作能力。

本文为IPO早知道原创

作者|Stone Jin

据IPO早知道消息,MiniMax于11月10日发布图像生成模型 MiniMax H3-Image,并同步开放模型权重。

该模型由 MiniMax H3 衍生,兼容文生图、图像参考和图像编辑。在 H3 预训练赋予的广泛图像能力基础上,H3-Image 进一步强化了整个后训练链路,数据、强化学习、步数压缩、PE、部署优化等多维度提高使得 H3-image 同时具备了速度、画面质量、指令遵循三维度竞争力。

据介绍,MiniMax H3-Image 可以在 10s 内生成一张 4096 × 4096 像素,约 16MP(4K) 的高规格图像,且速度可达市面主流水平的 5 倍。画面质量上,模型给出了多元的画面细节和极强的真实感,AI 生成图像与真实视觉世界之间的距离进一步缩小。另外,H3-Image 具备开源模型中顶级的指令遵循能力并最多兼容 16 张参考图片。

MiniMax 同时借助开放平台、MiniMax Code 和 MiniMax 海螺给出 H3-Image 图像生成服务。H3-Image 标准模式的生成速度约为市面主流工具的 1.5 至 2 倍,Highspeed 模式可达 3 至 4 倍,服务定价约为同档位主流工具的 50% 至 70%。

发布首周,MiniMax 对 16 MP 及以上规格的生成请求给出官方定价五折优惠。

MiniMax H3 发布三个月后,MiniMax 将开放权重的策略从视频生成延伸到图像生成。在模型架构维度,H3 Image 移除了 H3 原有架构中的音频分支,针对图像生成和编辑开展了优化。Hugging Face公布的工艺资料表明,H3 Image使用331亿参数的稠密扩散Transformer(DiT),共有50层,沿用MiniMax H3的Qwen3-VL文本编码器,并结合图像VAE处理参考图片和生成图像。

企业和开发者可以基于模型权重探索新的工具和创作方式,也可以借助 MiniMax 开放平台等官方渠道直接使用高质量生成服务。借助 H3-Image 的官方 API,个人开发者可以在 MiniMax Code、WorkBuddy、Codex 等智能体中接入日常工作流,让图像生成成为各类 Agent 框架可调用的基础能力。

H3-Image 的发布将开源图像模型的能力上限提高到新的维度,即能否准确理解复杂的创作要求,以及能否把这些要求转化为有美感、有真实感的画面。一张包含文字和多层信息的图像,既要把内容和关系交代清楚,也要让构图、光影和细节服务于创作者的意向表达。H3-Image 将指令遵循、多图参考与高分辨率生成结合起来,延续了 MiniMax 的多模态融合路线,也为开源模型在复杂视觉创作中缩小与顶尖闭源模型的差距给出了新的可能。

开源首日,MiniMax 携手华为昇腾、平头哥、昆仑芯、摩尔线程、壁仞、海光、沐曦、天数、NVIDIA、AMD、英特尔、Tenstorrent等全球芯片企业与多家主流推理和云服务平台推进Day 0接入工作。MiniMax表示,“希望 H3-Image 可以给更多人的日常生活、工作带来便利,笔者也会持续构建更强有力的工具来协助普惠 AGI 的达成。”


郑重声明:使用者在社区发表的所有信息将由本网站记录保存,仅代表作者个人观点,与本网站立场无关,不对您构成任何投资建议,据此操作风险自担。请勿相信代客理财、免费荐股和炒股培训等宣传内容,远离非法证券活动。请勿添加发言使用者的手机号码、公众号、微博、微信及QQ等信息,谨防上当受骗!

本文来源:https://caifuhao.eastmoney.com/news/20261112002331589315520
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

Expanding the visual creation capabilities of AI agents.

This article is original to IPO Zhidao (IPO早知道).

Author: Stone Jin

According to IPO Zhidao, MiniMax released its image-generation model, MiniMax H3-Image, on November 10, and simultaneously made the model weights available.

The model is derived from MiniMax H3 and supports text-to-image, image reference, and image editing. Building on the broad image capabilities enabled by H3’s pretraining, H3-Image further strengthens the entire post-training pipeline. Improvements across multiple dimensions—including data, reinforcement learning, step compression, PE (positional encoding), and deployment optimization—give H3-Image competitive strength in three areas: speed, image quality, and instruction-following.

According to the company, MiniMax H3-Image can generate a high-specification image at 4096 × 4096 pixels, approximately 16MP (4K), within 10s, with speeds up to 5× that of mainstream solutions on the market. In terms of image quality, the model delivers diverse visual details and strong realism, further narrowing the gap between AI-generated images and the real visual world. In addition, H3-Image features top-tier instruction-following among open-source models and supports up to 16 reference images.

MiniMax is also making H3-Image image-generation services available through its open platform, MiniMax Code, and MiniMax Hailuo. The generation speed of H3-Image’s Standard Mode is about 1.5 to 2× that of mainstream tools on the market, while Highspeed Mode can reach 3 to 4×. Service pricing is approximately 50% to 70% of mainstream tools in the same tier.

During the first week after release, MiniMax is offering a 50% discount on official pricing for generation requests at 16MP and above.

Three months after the launch of MiniMax H3, MiniMax is extending its open-weights strategy from video generation to image generation. In terms of model architecture, H3-Image removes the audio branch in H3’s original architecture and has been optimized for image generation and editing. Technical materials published on Hugging Face indicate that H3-Image uses a 33.1-billion-parameter dense diffusion Transformer (DiT) with a total of 50 layers, retains MiniMax H3’s Qwen3-VL text encoder, and incorporates an image VAE to process reference images and generated images.

Enterprises and developers can explore new tools and creative approaches based on the model weights, or use high-quality generation services directly through official channels such as the MiniMax open platform. With H3-Image’s official API, individual developers can integrate it into everyday workflows within agents such as MiniMax Code, WorkBuddy, and Codex, making image generation a foundational capability that can be called by various Agent frameworks.

The release of H3-Image raises the capability ceiling of open-source image models to a new dimension: whether complex creative requirements can be accurately understood, and whether those requirements can be translated into visuals that are aesthetic and realistic. An image containing text and multiple layers of information must not only convey content and relationships clearly, but also ensure that composition, lighting, and details serve the creator’s intended expression. H3-Image combines instruction-following, multi-image reference, and high-resolution generation, continuing MiniMax’s multimodal-fusion approach and offering new possibilities for open-source models to narrow the gap with leading closed-source models in complex visual creation.

On the first day of open-sourcing, MiniMax worked with global chip companies—including Ascend (Huawei), T-Head, Kunlunxin, Moore Threads, Biren, Hygon, Muxi, Iluvatar, NVIDIA, AMD, Intel, and Tenstorrent—as well as multiple mainstream inference and cloud service platforms to advance Day 0 integration efforts. MiniMax stated, “We hope H3-Image can bring convenience to more people’s daily lives and work, and our team will continue to build more powerful tools to help achieve inclusive AGI.”

Important notice: All information posted by users in the community will be recorded and retained by this website. It represents only the personal views of the authors and does not reflect the position of this website. It does not constitute any investment advice, and any actions taken based on it are at your own risk. Please do not trust promotional content such as managing funds on behalf of others, free stock recommendations, or stock-trading training, and stay away from illegal securities activities. Please do not add personal contact information such as the phone numbers, official accounts, Weibo, WeChat, or QQ of commenters, to avoid being scammed.

MiniMax open-sources SOTA image model H3-Image

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry

←上一篇← Prev下一篇→Next →