今年 8 月,MiniMax 开放了 H3 视频生成模型的权重,这是有史以来第一个达到 SOTA 水平的开源视频模型。
此后几个月,社区基于它玩出的花样远超官方演示,也催生了 H3 max 等一批衍生模型,实实在在推着 AI 创意行业往前走了一步。

现在,MiniMax 连图像生成模型也不打算藏着了。就在刚刚,MiniMax H3 Image 正式发布,重点是图像生成与编辑,目标是开源图像模型中的 SOTA。
此前的 H3 已经借助统一的全模态预训练,具备了图像、视频、音频等多种内容的理解与生成能力,其中也涵盖文生图和图像编辑。
而这次的 H3 Image 是在 H3 已有能力的基础上,进一步针对图像生成与编辑开展后训练优化,让模型在复杂指令遵循、图像细节和视觉审美上有了更好的表现。

比较直观的感受,是分辨率和速度。在笔者的实测中,H3 Image 生成一张 4096×4096、4K 画质的高规格图片,只需要 10~20 秒。
这意味着,过去动辄需要等待几十秒甚至更久的高分辨率生图,如今已经可以在十几秒内完成。4K 级图像生成正在进入以十秒计的时代,高分辨率带来的细节长处,也终于可以用在更频繁的日常创作中。
笔者使用 H3 Image 生成了各种极具美学的图片,它在 CG、游戏情形、工业外观以及多种艺术风格上都有不错的表现。

它还重点强化了复杂指令遵循能力,最多兼容 16 张参考图片,可以同时理解不同的主体、材质和风格要求。
更关键的是,H3 Image 在发布当天就开放权重,开发者也可以借助 API 接入,或者在 MiniMax 开放平台、MCode 和 MiniMax 海螺平台中使用。

APPSO 也提前感受了这款新模型,除了画得好看,笔者更想知道,它到底能理解多复杂的需要,又能把图像细节做到什么程度。
复杂的需要,也能准确画出来
过去笔者测试图像模型,往往更关注它能不能画出提示词里浮现的物体。
但实际工作里,笔者交给外观师的需要经常复杂得多。例如,机型的零件应该如何组装,某个区域需要保留哪些元素,不同参考图之间又有什么联系。模型需要先理解这些条件,才可能画对图。

这次,笔者从一个工业机型的结构外观入手,让 H3 Image 生成一台将来主义咖啡机的机型爆炸图。它需要展示豆仓、研磨器、冲煮模块、水箱等结构,同时标明各部分的连接关系,并保持机型外观的一致性。

除了机械结构,文字处理也是图像生成模型经常遇到的难题。尤其是既要把字写对,又要兼顾字体外观、画面构图和总体视觉成效。
对于大字海报,模型需要同时处理文字内容、字体造型和视觉构图。而在包含大段小字的情形,文字准确性、信息层级和排版也会成为新的难题。
笔者先让 H3 Image 生成了一张以凤凰(Phoenix)为主题的新古典主义装饰艺术海报,融合了西方神话、复古雕版画、维多利亚时代装饰纹样和现代极简排版。
H3 Image 准确地渲染了画面标题 PHOENIX MYTH(凤凰神话)和副标题 Eternal Rebirth of the Fire Bird(火鸟的永恒重生)。

对于 H3 Image 在小字上的表现,笔者找到了这段时间沸沸扬扬的黄仁勋公开信,让 H3 Image 生成了一份放在桌面的信纸版。

图片里面的文字基本上都准确生成出来了,没有显著的错误,总体的样式也和图片周围的风格保持一致。如果你仔细看信纸周围的细节,都和老黄有关。
笔者还能生成怀旧风的情形:挂在墙上的经典皮衣,伴随着打字机上的独白「我已经被穿了 30 年了。」

笔者还尝试让 H3 Image 生成更复杂的游戏界面,图片里的文字渲染和布局都和生存游戏类似,表明生命、魔力和体力等级。

更进一步的是,看 Image 能否理解文字、图片之间的逻辑,这要求它具备语义理解的能力,并具备科学、数学、生物等知识,并能把逻辑准确、美观地展现出来。

从机械结构到文字排版,再到知识理解,这些案例实际上都指向同一个难题:模型能否真正理解使用者想表达的信息,并将其准确地转化为图像。
这次 H3 Image 兼容最多 16 张图片作为参考。这意味着,笔者可以一次给出机型的不同视角、材质参考、品牌元素和情形图片,再要求模型将这些信息整合到同一张作品中。

更值得期待的是,这次 H3 Image 还可以借助 API 接入 MiniMax Code、Claude Code、Codex 等 Agent Harness,直接参与更复杂的任务流程。
也就是说,笔者可以直接在 Agent 的对话框里提出创作需要,让 Agent 整理任务、调用 H3 Image 生成图片,再借助后续对话不断修改外观。


向左滑动查看更多内容
就像这个在 Codex 内使用 MiniMax H3 Image 生成 V8 引擎爆炸拆解图的案例,借助 Agent 对信息的搜集和整理,最后生成的几张拆解图都格外真实。




向左滑动查看更多内容
对于 Agent 来说,它需要持续管理上下文和使用者的修改意见;对于 H3 Image 来说,则需要理解 Agent 整理后的复杂指令,并尽可能准确地执行。这样的工作方式,对图像模型的指令遵循和复杂信息理解能力也提出了更高的要求。
而这款开源的图像生成模型 H3 Image 确实能做到理解复杂需要、执行精确修改,并在连续创作中保持一致性,这也它开始真正融入 Agent 工作流。
尤其是当下很多主流生图模型生成的图片还停留在百万像素或四百万像素,而 H3 Image 可以直接生成 4096×4096 的超高清画面,像素数量达到 4K 画质。
这就意味着在后续的工作流里,即便笔者把画面局部放大、裁切,仍旧有足够的细节可供使用。无论是机型外观、商业海报,还是需要在大屏上展示的视觉素材,都有了更多后期处理空间。
只靠一段提示词,就能画出真实的美感
理解力强不代表审美就一定好,笔者已经见过不少可以严格遵守提示词的 AI,画出来的东西却始终差点意思。
就像 GPT-Image,很多人对它的印象都是生成的信息图很强,不过在面对 Midjourney 这类风格化极强的图片生成上,GPT-Image 也会有点力不从心。
而对 AI 图像来说,尤其是 CG、游戏概念外观、插画这些情形,很多时候,笔者并没有办法把想要的构图、色彩和氛围全部描述清楚,只能给出一个大致的方向,基本上是靠模型自己的审美水平。
这次测 H3 Image 的美学,笔者从 Midjourney 上找了一些喜欢的参考图出发,先让 ChatGPT 把图里的构图、光线和材质拆成提示词,再用这些提示词给 H3 Image 出题。




向左滑动查看更多内容
例如这几张充满速度感的赛马图片,笔者在提示词里面明确了图片的核心风格是电影感动态摄影、慢门追焦摄影和复古运动胶片美学。
H3 Image 能准确地把这种赛马高速奔跑产生的横向拖影、深墨绿色的背景、金色夕阳照亮的马匹,以及画面中大面积幽暗、抽象的负空间全部描绘出来。
不用说没有 AI 味,当看到这些照片被生成时,甚至还有很强的大片电影感,是经过摄影摄像师花大量时间的修图。
笔者继续用这组提示词测试不同的运动情形,从赛车、骑行、滑板,到行驶的汽车、奔跑的人和动物,看看 H3 Image 能否展现速度之美,以及摄影中那种不完美、但动人的质感。



向左滑动查看更多内容
每一张都很有感觉,运动中的主体逐步消融成光线、色彩与时间的痕迹。



向左滑动查看更多内容
在人像的生成上,笔者让 H3 Image 生成了几张复古直闪胶片摄影 + 生活方式纪实摄影 + 独立时尚杂志美学的图片。
生成画面就像是朋友在一场私人派对上随手拍到的瞬间,没有精心安排的姿势,却比正式写真更有生命力。
尤其值得注意的是这些图片里所展示出来的人物真实感。皮肤保留了自然的纹理,头发、衣物和光影的细节也足够多元,没有很多 AI 生成人像常见的过度磨皮和油腻质感。




向左滑动查看更多内容
当把这种风格复制到其他图片上时,也能看到 H3 Image 对于风格的把控能力。把夜晚直闪带来的强烈明暗反差、人物自然大笑的失控瞬间、略微凌乱的构图、浓郁的花卉色彩,以及不加修饰的胶片快照质感,全都顺利生成了出来。




向左滑动查看更多内容
类似的人像照片还有笔者使用 H3 Image 生成的一张极具东方古典美学的电影剧照,融合中国古装电影摄影、自然主义光影和复古胶片视觉质感。

这些图片就像是把古代生活中一个平凡而亲密的瞬间给复现了。穿过树叶的午后阳光温柔照亮人物的脸庞,自然的神情与斑驳的树影,让画面变成了一段被时间珍藏的电影记忆。
而在复刻 GTA V 游戏实机画面的过程中,笔者也是在提示词里面指定了风格,以及第三人称的视角,要求 H3 Image 生成一张具有《Grand Theft Auto V》游戏实机视觉风格的开放世界 3D 游戏截图。




向左滑动查看更多内容
H3 Image 完全理解了笔者想要的是「GTA 游戏画面感」,并非简单的「GTA 风格」。几张图片都像玩家在 GTA 开放世界中随手截取的一帧画面,和其他 AI 生成图片不太一样,其他 AI 生成的更像是官方精心制作的宣传图。
笔者另外还生成几张极具风格的图片,而所有的图片都是直接一次性生成的文生图形式,没有额外的图片参考,只凭借一次性的提示词,H3 Image 就能在美学上做到完全没有 AI 味的同时,还把笔者要求的风格准确描绘出来。




对已经有参考图的创作者来说,笔者可以使用 GPT-Image 去提取对应的风格提示词,也可以直接使用 H3 Image 的参考能力,应用某张照片的追焦方式,或者一幅插画露出纸底的笔触,再把它放到另一个题材上。
当复杂指令理解与视觉审美可以同时兼顾时,图像模型也就有了更多实际用途。从游戏概念图、品牌海报,到机型结构示意和外观思路,使用者需要反复解释需要、调整提示词的次数,有机会进一步减少。
从视频到图像,MiniMax 正在把全模态能力开放出去
回到 H3 Image,这次发布最值得关注的地方应该就是开源。
无论是一开始的视频生成模型,还是现在的图像生成,都是长期以来对算力、训练数据和模型能力有很高要求的领域。
尤其是顶尖的视觉生成模型,很多时候,笔者只能借助官方机型或者 API 使用。模型能做什么、怎样运行、如何进一步改造,仍然取决于服务给出者。

MiniMax 的开源让模型从一家公司的机型,变成了更多使用者、开发者继续创造的起点:他们不仅可以使用现成的模型,也有机会依据自己的需要开展部署、微调和优化,围绕它开发新的创作工具,真正地把模型能力融入现有机型和工作流。
对内容和更广的行业来说,H3 Image 作为一款 SOTA 的开源图像模型,给了人们一种新的挑选。例如,游戏开发者可以尝试建立符合自身美术风格的素材生成流程,外观团队可以针对特定机型建立自动化工作流,个人开发者也有机会围绕开放模型构建自己的创作应用。

从视频到图像,再结合文本、语音等模型,MiniMax 的多模态布局正在逐步形成更完整的能力矩阵。H3 已经证明过一次,当权重开放之后,社区涌现的玩法、衍生模型和工作流远超官方演示的边界,也验证了开源模型同样可以支撑起扎实的商业化。
对笔者来说,最值得期待的事情,是第一梯队的视觉生成能力不再只能依赖少数封闭服务。
当足够聪明、足够好看的生成模型也能开放给开发者,更多人就有机会把自己的创意,变成真正属于自己的工具和作品。
*本文所有图片均用 MiniMax H3 Image 生成
本文来源:https://www.ifanr.com/1688790
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
In August this year, MiniMax released the weights of the H3 video generation model, the first open-source video model to ever reach SOTA level.
In the months since, the creative applications built by the community have far surpassed the official demos, also spawning a batch of derivative models such as H3 max, tangibly pushing the AI creative industry forward a step.
Now, MiniMax is not holding back on its image generation model either. Just moments ago, MiniMax H3 Image officially launched, focusing on image generation and editing, with the goal of being the SOTA among open-source image models.
The previous H3, through unified omni-modal pretraining, already possessed the ability to understand and generate various types of content including images, video, and audio, which also included text-to-image and image editing.
This H3 Image builds on H3's existing capabilities with further post-training optimization focused on image generation and editing, enabling the model to perform better in complex instruction following, image detail, and visual aesthetics.
The most immediately noticeable improvements are resolution and speed. In my hands-on testing, H3 Image generated a high-spec image at 4096×4096 with 4K quality in just 10 to 20 seconds.
This means that high-resolution image generation, which previously often required waiting dozens of seconds or even longer, can now be completed in just over ten seconds. 4K-level image generation is entering the era of ten-second timing, and the detail advantages brought by high resolution can finally be utilized in more frequent daily creative work.
I used H3 Image to generate a variety of highly aesthetic images, and it performed well in CG, game scenarios, industrial design, and various artistic styles.
It also places heavy emphasis on complex instruction-following capabilities, supporting up to 16 reference images and simultaneously understanding different subject, material, and style requirements.
More importantly, H3 Image opened its weights on the day of release, and developers can also access it via API or use it on the MiniMax Open Platform, MCode, and the MiniMax Hailuo Platform.
APPSO also got an early experience of this new model, and beyond just creating beautiful images, I wanted to know how complex the requirements it can understand are, and to what extent it can render image details.
Complex requirements can also be accurately rendered
In the past when I tested image models, I tended to focus more on whether they could draw the objects that emerged in the prompts.
But in real work, the requirements I give to designers are often far more complex. For example, how the parts of a machine should be assembled, which elements need to be preserved in a certain area, and what connections exist between different reference images. The model needs to understand these conditions first before it can generate the correct image.
This time, I started with the structural appearance of an industrial machine and asked H3 Image to generate an exploded diagram of a futuristic coffee machine. It needed to display structures such as the bean hopper, grinder, brewing module, and water tank, while indicating the connection relationships between each part and maintaining consistency in the machine's exterior design.
Besides mechanical structures, text processing is also a common challenge for image generation models, especially when they must not only write the text correctly but also account for font aesthetics, visual composition, and overall visual impact.
For large-text posters, the model needs to simultaneously handle text content, font styling, and visual composition. In cases involving large blocks of smaller text, text accuracy, information hierarchy, and layout also become new challenges.
The author first asked H3 Image to generate a Neoclassical decorative art poster themed around the Phoenix, blending Western mythology, vintage engravings, Victorian-era decorative motifs, and modern minimalist layouts.
H3 Image accurately rendered the title PHOENIX MYTH and the subtitle Eternal Rebirth of the Fire Bird.
To test H3 Image's performance with smaller text, the author took the widely discussed recent open letter from Jensen Huang and asked H3 Image to generate a desktop letterhead version.
The text within the image was generated accurately without any noticeable errors, and the overall style remained consistent with the surrounding elements. If you look closely at the details around the letterhead, they are all related to Jensen Huang.
The author can also generate nostalgic scenarios: classic leather jackets hanging on a wall, accompanied by a monologue on a typewriter stating, "I have been worn for 30 years."
The author also attempted to have H3 Image generate a more complex game interface, where the text rendering and layout resemble those of a survival game, indicating health, magic, and stamina levels.
Furthermore, it tests whether Image can understand the logic between text and images, which requires semantic comprehension capabilities as well as knowledge of science, mathematics, and biology to accurately and aesthetically present that logic.
From mechanical structures to text layout, and further to knowledge comprehension, these cases all point to the same challenge: whether the model can truly understand the information the user intends to convey and accurately convert it into an image.
This time, H3 Image supports up to 16 reference images. This means the author can provide different perspectives of a device, material references, brand elements, and scenario images simultaneously, and then request the model to integrate this information into a single piece of artwork.
Even more promising is that H3 Image can now connect to Agent Harnesses such as MiniMax Code, Claude Code, and Codex via API to directly participate in more complex workflows.
In other words, the author can directly specify creative requirements in the Agent's dialog box, allowing the Agent to organize the task, invoke H3 Image to generate the image, and then use subsequent conversations to continuously modify the appearance.
Swipe left to see more content
Just like this case of using MiniMax H3 Image within Codex to generate an exploded diagram of a V8 engine, the resulting several exploded-view images were exceptionally realistic thanks to the Agent’s gathering and organizing of information.
Swipe left to see more content
For the Agent, it needs to continuously manage context and user revision requests; for H3 Image, it must understand the complex instructions organized by the Agent and execute them as accurately as possible. This workflow also places higher demands on the image model’s instruction following and complex information understanding.
And this open-source image generation model H3 Image can indeed understand complex requirements, execute precise modifications, and maintain consistency across successive creations, which is why it has truly begun to integrate into Agent workflows.
Especially now that many mainstream image generation models still produce images at 1 million or 4 million pixels, H3 Image, however, can directly generate ultra-high-definition images at 4096×4096, with a pixel count reaching 4K quality.
This means that in subsequent workflows, even if I zoom in or crop parts of the image, there are still sufficient details available. Whether for vehicle exteriors, commercial posters, or visual assets that need to be displayed on large screens, there is more room for post-processing.
With just a single prompt, it can produce a realistic aesthetic.
Strong comprehension doesn’t necessarily mean good aesthetics. I’ve encountered many AIs that can strictly adhere to prompts, yet their outputs always feel somewhat lacking.
Take GPT-Image, for example; many people see it as very strong at generating infographics, but when faced with highly stylized image generation like Midjourney, GPT-Image can feel a bit out of its depth.
As for AI images, especially in cases like CG, game concept art, or illustration, I often can’t fully describe the composition, colors, and atmosphere I want; I can only give a general direction, so it basically comes down to the model’s own aesthetic level.
To test H3 Image’s aesthetics this time, I started by picking some reference images I like from Midjourney. I first asked ChatGPT to break down the composition, lighting, and materials in those images into prompts, then used those prompts as challenges for H3 Image.
Swipe left to see more content.
For example, with these horse racing images full of a sense of speed, the author specified in the prompt that the core style should be cinematic dynamic photography, slow-shutter panning, and a vintage sports film aesthetic.
H3 Image can accurately render the horizontal motion streaks from the horses running at high speed, the deep ink-green background, the horses lit by the golden sunset, and the large expanses of dark, abstract negative space in the frame.
Needless to say, they don’t look AI-generated; when these photos are generated, they even have a strong big-budget cinematic feel, as if a photographer spent a great deal of time retouching them.
The author continued to test this set of prompts across different motion scenarios—from racing cars, cycling, and skateboarding, to moving vehicles, running people, and animals—to see whether H3 Image could capture the beauty of speed and that imperfect yet moving texture found in photography.
Swipe left to see more content.
Every image is full of feeling, with the moving subject gradually dissolving into traces of light, color, and time.
Swipe left to see more content.
For portrait generation, the author asked H3 Image to create several images in the style of vintage direct-flash film photography + lifestyle documentary photography + an independent fashion magazine aesthetic.
The generated images feel like moments casually captured by a friend at a private party—no carefully arranged poses, yet more alive than formal portrait shoots.
What’s especially noteworthy is the realism of the people shown in these images. The skin retains natural texture, and the details of hair, clothing, and lighting are sufficiently varied, avoiding the over-smoothed skin and greasy look that many AI-generated portraits often have.
Swipe left to see more content
When applying this style to other images, H3 Image's ability to control the style is also evident. It successfully generated the intense contrast between light and shadow brought by direct flash at night, the unrestrained moments of natural laughter from the subject, slightly messy composition, rich floral colors, and the unadorned quality of a film snapshot.
Swipe left to see more content
Another similar portrait is a cinematic still generated by the author using H3 Image that possesses a strong sense of classical Eastern aesthetics, blending Chinese costume film photography, naturalistic lighting, and a vintage film visual texture.
These images are like a recreation of an ordinary yet intimate moment from ancient life. The afternoon sunlight filtering through the leaves gently illuminates the subject's face, while the natural expression and dappled shadows of the trees turn the scene into a cinematic memory preserved by time.
In the process of replicating actual gameplay footage from GTA V, the author specified the style and third-person perspective in the prompt, requesting H3 Image to generate an open-world 3D game screenshot with the visual style of "Grand Theft Auto V".
Swipe left to see more content
H3 Image fully understood that what the author wanted was the "feeling of a GTA game screen," not simply "GTA style." These images look like frames captured casually by a player in GTA's open world, unlike other AI-generated images that tend to resemble officially curated promotional shots.
I also generated several highly stylized images. All of them were created in a single pass using text-to-image generation, without any additional image references. Relying solely on one-shot prompts, H3 Image is able to achieve an aesthetic that completely avoids the "AI look," while accurately depicting the styles I requested.
For creators who already have reference images, I can use GPT-Image to extract corresponding style prompts, or I can directly utilize H3 Image's reference capabilities to apply the focus tracking style of a specific photo, or the paper-texture brushstrokes of an illustration, onto another subject.
When complex instruction comprehension and visual aesthetics can be balanced simultaneously, image models gain more practical applications. From game concept art and brand posters to structural diagrams and exterior design concepts, the number of times users need to repeatedly explain requirements and adjust prompts has the potential to be further reduced.
From Video to Image: MiniMax is Opening Up Its Omnimodal Capabilities
Returning to H3 Image, the most noteworthy aspect of this release is likely its open-source nature.
Whether it was the initial video generation models or the current image generation, these have long been domains with high demands for computing power, training data, and model capabilities.
Especially with top-tier visual generation models, I often had to rely on official models or APIs. What the model can do, how it runs, and how it can be further modified remained dependent on the service provider.
MiniMax's open-sourcing turns the model from a single company's product into a starting point for continued creation by more users and developers: they can not only use the ready-made model but also have the opportunity to deploy, fine-tune, and optimize it according to their own needs, develop new creative tools around it, and truly integrate model capabilities into existing models and workflows.
For content creation and the broader industry, H3 Image, as a SOTA open-source image model, offers a new choice. For example, game developers can attempt to establish asset generation pipelines that match their own art styles, exterior design teams can build automated workflows for specific models, and individual developers also have the opportunity to build their own creative applications around the open model.
Image | CG Scene: The Tidal Library on the Cliff
From video to image, and combined with text and voice models, MiniMax's multimodal layout is gradually forming a more complete capability matrix. H3 has already proven once that when weights are opened up, the playstyles, derivative models, and workflows emerging from the community far exceed the boundaries of official demos; it also validated that open-source models can likewise support robust commercialization.
For me, the most promising prospect is that first-tier visual generation capabilities will no longer be dependent solely on a few closed services.
When sufficiently intelligent and aesthetically pleasing generation models are also opened up to developers, more people will have the opportunity to turn their creativity into tools and works that truly belong to them.
*All images in this article were generated using MiniMax H3 Image.