从 Demo 好看到生产好用。
2026 年的 AI 3D 赛道，正站在一个微妙的分水岭上。
一级市场的热度仍在攀升，新模型、新创业公司层出不穷，单图生 3D 的正面渲染效果几乎已经做到了以假乱真。
但剥开炫目的 Demo 外壳：绝大多数生成结果依然停留在「转台展示」阶段：一个精致的角色或机械模型，在灯光下缓缓旋转，鳞片、衣褶和金属刻纹栩栩如生；可这些细节一旦进入游戏引擎装配、工业仿真以及 3D 打印的真实生产流程就立刻原形毕现：结构断裂、背面崩坏、细节粘连、网格不可编辑等问题开始集中爆发，大量人工修补成本吞噬了 AI 带来的效率红利。
正面视图保真度不等于 3D 质量成为新的行业共识。AI 3D 的竞赛，也从 能不能生成」的上半场，进入了能不能落地的下半场。
也正是在这个节点上，头部玩家 Hi3D 接连落下两枚棋子：正式开源经过生产验证的 Hi3D V1.0 底座，同时推出以创作者为核心的 Hi3D Agent，一头连接产业底层，一头深度扎根进入真实的创作全流程。
在此之前，这家成立不到三年的公司已经交出了一份扎实的成绩单：半年内连续完成两轮融资、总额近 5000 万美元，红杉中国、IDG 资本、美团龙珠等一线机构扎堆下注；旗下 Hi3D 平台覆盖120+ 国家与地区、200 万 + 创作者，与拓竹科技、xTool 等头部设备商深度集成。
从稀疏几何技术的学术突破，到生产级模型的开源开放，再到 Agent 创作系统的全链路打通，Hi3D 已经拿下了通往 3D 生成下半场的门票。
01
3D 生成，被高估的 Demo，
被低估的生产鸿沟
单图生 3D 的底层逻辑，天然带着视角偏见。随着ULIP、Uni3D 等指标被反复刷新，重正面、轻整体，重外观、轻内部逻辑的一度成为这个行业过去几年的共同写照。
但 3D 资产的本质，是可旋转、可编辑、可加工的立体对象。一张角色图片里，垂落的发丝可以只占几个像素。进入三维空间后，模型必须决定它的厚度、与肩膀之间的距离，以及它如何连接头部。更进一步，物体的光影与轮廓之外，背面、底部、被遮挡的内部结构，全部需要模型从训练数据中学到的三维先验进行推理。
否则，一个正面相似的 3D 模型，很可能转个角度就露出马脚——本该闭合的壳体出现破洞，对称的零件左右扭曲，纤细的结构要么融成一团，要么直接凭空消失。
更进一步，解决了多视角下的几何一致性问题后，一个模型能否被下游工具使用，还取决于它的精细结构保留度、网格拓扑合理性，甚至是壁厚、封闭性等工程属性。
有实测数据显示：同处 1536³ 分辨率档位，即使一些头部玩家的产品，3D 打印切片成功率仅为 7%-17%。看起来细节拉满的模型，实际走向实体生产就会变得困难重重。
这也解释了为什么资本市场热度不减，产业端却始终处于试点多、规模化少的状态。而当行业走向下半场，谁能先跨过 Demo 好看→生产好用这道坎，谁才能真正拿到 AI 3D 产业化的入场券。
02
Hi3D V1.0 开源：
把 200 万资产训练的生产级底座，
交还给全行业
Hi3D V1.0 的技术起点，可以追溯到 2025 年的 Sparc3D。
这项由 Math Magic、南洋理工大学及帝国理工学院研究者参与的研究，处理的是高分辨率三维建模中的一个基础问题：怎样保存复杂表面，又不让计算和存储成本随着分辨率失控。
三维网格背后的体素空间有立方级的规模变化。每个方向从 1024 提高到 1536，线性分辨率增加 50%，完整网格中的单元数量就会变成原来的 3.375 倍。大量空白区域如果也参与计算，成本会迅速上升。
Sparc3D 采用稀疏表示，将计算更多集中在与物体表面有关的区域，并通过稀疏卷积自编码器压缩和重建几何信息。论文使用约 50 万件资产进行训练，展示了 1024³级别的高分辨率重建能力，通过改善几何信息在表示、压缩和恢复过程中的保存方式，在高分辨率几何建模领域建立了技术标杆。
Hi3D V1.0 则在此基础上，将训练资产扩充到约 200 万件，覆盖角色、硬表面物件、日常产品、载具、建筑、有机形态等多元品类。并新增了 1536³的高分辨率输出层级，通过三级渐进式稀疏重建实现精度跃升：三级结构分别基于 64³、128³、192³的稀疏坐标网格，通过固定 8 倍的结构到表面解码器缩放，对应输出 512³、1024³、1536³的名义分辨率。
Hi3D V1.0 生成的矮人工匠
但比训练规模与分辨率数字更重要的是生产验证。在正式开源前，这套模型已经在面向用户的生产部署中完成了真实场景的打磨。
公开测试报告显示，在主观的专业 3D 算法从业者的人工偏好测试中，Hi3D V1.0 的综合偏好得分达到 85.9%，对多数主流模型的偏好胜率超过 90%。在客观指标上，其多视图保真度（MVF）达到 0.75，细节丰富度（DR）达到 0.841，两项核心几何指标均位列同期方法第一。
在 Hi3D 披露的一组 1536³档位测试中，其切片通过率为 90%，而同期同档位的竞品仅为 7% 至 17%。
十几倍的差距背后，是实验室模型和生产级模型的分界线。它打破了高分辨率必然牺牲工程可用性的行业刻板印象，证明了精细几何与生产兼容完全可以并行。
而通过开源，这套经过真实业务检验的高分辨率几何生成能力，又从此变成了全行业可复用的公共底座：企业可以基于它做私有化部署，把 AI 3D 能力嵌入自身的数字化产线；开发者可以在上面搭建垂直工具，不用从零重建高分辨率几何管线；研究团队可以站在生产级基线之上，去探索更前沿的方向。
03
从 Hi3D V1.0 到
Hi3D V3.0 与 Hi3D Agent，
一条清晰的生产级交付进化路径
一定程度上，Hi3D 的版本迭代路线，几乎就是一部 AI 3D 向生产端靠拢的浓缩史。
V1.0 阶段，核心突破是 1536³ 级别的高精度几何重建，把高分辨率几何重建做到了生产可用。这解决了细节准不准、结构完不完整的问题，让生成的模型至少能进下游工具、能做基础加工。
V1.5 与 V2.0 阶段，主要突破则在于几何与纹理一体化生成，智能遮挡结构补全、去光照预处理，覆盖了从输入图到完整资产的一站式输出，减少了用户在不同工具间的跳转成本，纹理质量与生成效率也同步完成提升。
到了 V3.0 阶段，几何分辨率进一步拉升至业界首个商用 2048³体素级，纹理输出上限达到 8K，并深度升级了跨视角结构推理与自研 UV 补全算法，上线智能拆件、多色打印、摆盘等功能。
使用 Hi3D 制作的游戏 Demo
随着版本不断升级，人工修补成本随之大幅削减，AI 生成的 3D 资产，也逐渐从能用真正走向可交付。
但模型能力再强，也只是创作流程中的一个环节。3D 创作的真实路径从来不是用一张图出一个模型就结束了：能打印的高密度三角网格，未必适合角色动画；适合概念验证的模型，也未必满足精密零件的尺寸要求。
Hi3D Agent 的出现，加速了模型初期设计到现实世界生产之前的打磨效率。
当前行业对 3D Agent 的主流探索，大多基于 MCP 等协议，由大语言模型调用外部 3D 工具，本质上是「连接层」的思路——大模型负责发号施令，工具负责执行单点操作。
可通用大语言模型擅长理解意图、拆解任务，但它不懂几何拓扑，不懂薄壁结构，不懂网格规范；专业 3D 模型擅长还原形态、生成高精度几何，但它听不懂复杂的创作目标，也不会主动推进流程。
为此，Hi3D Agent 选择了另一条路径：直接构建面向创作者的 3D 工作层，以专业 3D 模型为核心底座，把意图理解、任务规划、模型生成、持续修改、工程优化、场景适配串联成一条连续的工作流。
在 Hi3D 的设计逻辑里，通用大语言模型与专业 3D 模型有着清晰的分工：大模型擅长理解创作者想做什么，拆解目标、规划步骤；而专业的 Hi3D 模型负责怎么做好，把意图转化为几何准确、结构完整的三维结果。
两者各司其职，深度耦合，最终把生成一个模型的单点动作，延伸成完成一件作品的完整链路。
对比，原图 / ChatGPT6 模型生成 / Hi3D Agent 模型生成
而对创作者来说，完成这个过程，可能只需要从一张参考图、一个半成品模型，甚至一句模糊的想法表达比如「修复车门结构，优化硬表面边缘，输出可打印格式」开始，剩下的迭代、修改、适配工作由 Agent 推进。每一步调整都有可对比的前后版本，创作者随时可以叫停、修正方向、决定取舍。
它追求的是一种创作者主导的、可感知、可控制的连续创作过程，能够把创作者从重复劳动里解放出来，让他们把时间花在审美判断、风格打磨、创意表达上。
借助 Agent 完成了 demo 到生产级模型的跨越后，Hi3D 还试图跨越生产级模型到最终产品之间的鸿沟。
因为 Hi3D 发现，有时候同一处断裂，可能来自几何连接太细，也可能来自材料、摆放方式或加工参数。算法团队需要知道哪些失败应该进入模型改进，哪些应该交给工艺调整。诊断错了方向，投入更多训练资源也未必有效。
过去，要排查清楚问题的根源，往往需要在多个软件与生成流程之间反复跳转，AI 生成解决了「从 0 到 1」的第一步后，「从 1 到 100」的打磨依然高度依赖人工。
为了弥补这一生成与生成之间的鸿沟，Hi3D 甚至在珠三角自建了工厂，把 AI 模型研发和实体生产工艺直接打通。公司创始人任利锋甚至还会带着投资人在投前去工厂看看，因为在他看来，AI 的价值最终要落地到真实的制造、真实的交付、真实的商品上，一个模型企业还必须积累关于材料、设备和产品类别的更多 know how，才更更好的服务于真实生产。
04
从做好产品到做大蛋糕
从 V1.0 到 V3.0，从开源底座到 Agent 系统，Hi3D 的布局脉络已经不只做一款最好用的 3D 生成工具，它更想成为的，是 AI 3D 时代的基础设施。
向下，Hi3D 通过 V1.0 开源开放底层能力，让更多的开发者和企业，以更低成本站在成熟的生产级底座上做创新；
向上，Hi3D 通过 V3.0 持续推高商用模型的精度边界，2048³体素级几何、8K 纹理输出，为专业创作和上层应用不断抬高天花板；
横向，Hi3D 则通过 Agent 打通从想法到交付的全链路，把单点工具升级为完整创作工作流；并通过 MCP 接口向外部 Agent 开放能力，让更多工作流、更多工具都能接入 Hi3D 的 3D 生成能力。让 3D 内容生成从专业开发者群体拓展到对 3D 内容生成感兴趣的普通创作者。
三层能力叠在一起，构成了一个完整的生态闭环：研究者在这里做技术迭代，开发者在这里做场景创新，创作者在这里完成作品交付，企业在这里落地生产应用……蛋糕越做越大，或许才是 Hi3D 的真正野心所在。
*头图来源： Hi3D
本文为极客公园原创文章，转载请联系极客君微信 geekparkGO
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
It's good to see production from Demo.
2026 AI 3D, standing on a delicate watershed.
The heat of the first-level market is still rising, new models, new start-ups are emerging, and the positive projection of 3D is almost complete.
However, the demo shell: the vast majority of the results are still in the "turner display" phase: a fine role or mechanical model, which rotates slowly under the light, and the plasters, coats and metal are tattooed as well; but these details are immediately as soon as they enter the game engine assembly, industrial simulations and the real production process of 3D printing: structural cracks, broken backs, details glues, and the non-editing of the grids have begun to erupt, and the efficiency dividend of AI has been absorbed by extensive manual repair costs.
Positive view authenticity does not equal 3D quality to new industry consensus. The AI 3D contest, also from the first half of whether it can be generated, entered the second half of whether it can land.
It is also at this node that the head player Hi3D has successively dropped two pieces: the official open source is a production-certified Hi3D V1.0 base, while the production-centred Hi3D Agent is launched, with one head connected to the bottom of the industry and one deep-seated into the real creation process.
Prior to that, the company, which had been in existence for less than three years, had handed over a solid report: two consecutive rounds of financing, totalling nearly $50 million, had been completed in six months, with the redwood China, IDG capital, the United States Dragonball, and so on; the Hi3D platform under the flag covered 120+ countries and regions, 2 million + creators, and had been in-depthd with masonry, xTool, and so on.
From the academic breakthrough of thin geometry to the open source of production-level models, to the full link of the Agent creation system, Hi3D has taken the ticket to the second half of 3D.
01
3D generated, overestimated Demo,
Undervalued production divide
The bottom logic of 3D is based on a single image, with natural bias. As the ULIP, Uni3D and other indicators have been updated repeatedly, they have been more positive and more holistic, and they have once become a common picture of the industry over the past few years.
However, the essence of the 3D asset is a revolving, editable, processable 3D object. In a character picture, the drop of hair can only be a few pixels. After entering a three-dimensional space, the model must determine its thickness, the distance to its shoulder and how it connects to the head. Further, beyond the light and contours of the object, the back, the bottom, the hidden internal structure, all require a three-dimensional a priori of the model from the training data.
Otherwise, a 3-D model that looks similar to the one that's likely to show up at an angle -- The hull that should have been closed had holes broken, symmetrical parts twisted to the right and left, and the fine structure either melted into a lump or simply disappeared.
Further, when the problem of geometry in multiple perspectives is resolved, whether a model can be used in downstream tools also depends on its fine structural retention, the legitimacy of grid expansion, and even the engineering attributes of wall thickness and containment.
According to empirical data, the 15,363 resolution slot at the same location, even for some head player products, 3D print slice success rate is only 7-17%. It appears that a model full of details makes it difficult to actually move to physical production.
This explains why capital markets are not less hot, while the industrial sector is always in a situation of many pilots and smaller sizes. And when the industry moves to the second half of the game, who can get past Demo and make good use of it, who can actually get an AI 3D entry ticket.
02
Hi3D V1.0 Open source:
The production base that trained 2 million assets,
Give it back to the whole industry.
The technical starting point for Hi3D V1.0 can be traced back to Sparc3D in 2025.
This study, which involved researchers from Math Magic, Nanyang Polytechnic University and Imperial Institute of Technology, dealt with a fundamental issue in the three-dimensional modelling of high resolution: how to preserve complex surfaces without allowing the calculation and storage costs to run out of control with resolution.
The body space behind the 3D grid has a Cube scale change. Each direction increased from 1024 to 1536, and the linear resolution increased by 50%, and the number of units in the complete grid became 3.375 times the original. A large number of blank areas can rise rapidly if they are also involved in computing.
Sparc3D, with a thin representation, will calculate to concentrate more on the area associated with the surface of the object and to compress and reconstruct geometric information through the thin volume encoder. The paper, which was trained using approximately 500,000 assets, demonstrated the reconstruction capacity of high resolution at level 10243 and established technical poles in the field of high resolution geometric modelling by improving the way geometric information is stored in the expression, compression and recovery processes.
On this basis, Hi3D V1.0 expands training assets to approximately 2 million items, covering a wide variety of types such as role, hard surface items, everyday products, carriers, buildings, organic forms, etc. A high resolution output level of 15,363 was added to achieve precision leaps through three-tier gradual and thin reconstruction: The three-tier structure is based on a thin grid of coordinates of 643, 1283 and 1923, respectively, and corresponds to the nominal resolution output of 5123, 10243 and 15363 by a fixed 8-fold structure to the surface decoder scaling.
Hi3D V1.0 Shortsmith
But more important than the scale of training and the resolution figure is production validation. Prior to the official open source, the model had been able to sharpen the real scene in the user-oriented production deployment.
Public testing reports indicate that the combined Hi3D V1.0 score reached 85.9 per cent in the artificial preference test for subjective professional 3D algorithms practitioners, with a preference of over 90 per cent for most mainstream models. For objective indicators, its multi-view authenticity (MVF) is 0.75, its detail enrichment (DR) is 0.841, and both core geometry indicators are ranked first in the methodology for the same period.
In a set of 15363 slot tests disclosed by Hi3D, the slice pass rate was 90 per cent, compared to only 7 to 17 per cent for the same slot during the same period.
Behind the gap of a dozen times is the dividing line between laboratory and production-level models. It breaks the industry stereotypes that high resolution must have sacrificed the availability of the works, proving that fine geometry and production compatibility are perfectly compatible.
This high resolution geometry capability, which has been tested for real business, has since become an industry-wide public base: Enterprises can use it as the basis for privatization deployments, embedding AI 3D capabilities in their own digital line; developers can build vertical tools on them without having to rebuild high resolution geometry lines from zero; and research teams can be on the production-level baseline to explore the way forward.
03
From Hi3D V1.0 to
Hi3D V3.0 with Hi3D Agent,
A clear evolutionary path for production-level delivery
To a certain extent, the Hi3D version of the iterative route is almost an AI 3D enrichment history that converges to the production end.
V1.0 The core breakthrough was a high-precision geometry reconstruction at the level of 15363, making high resolution geometry productive. This solves the problem of inaccuracy of detail and incomplete structure, so that the resulting model can at least enter downstream tools and do basic processing.
V1.5 With the V2.0 phase, the main breakthrough is geometrically integrated with texture, the completion of the smart shield structure, photopre-processing, covering the all-in-one output from the input diagram to the full asset, which reduces the cost of jumps between the user and the different tools, and increases in texture quality and generation efficiency.
Up to stage V3.0, geometry resolution was further raised to the industry ' s first commercial 20483 vivo level, with texture output caps of 8K and a deep upgrade of cross-view structural reasoning and self-study UV completion algorithms, upline smart-dismantling, multi-colour printing, palleting, etc.
Game Demo with Hi3D
As the version has been upgraded, the manual repair costs have been significantly reduced, and the 3D assets generated by AI have gradually moved from real to deliverable.
But models are more powerful and are only part of the creative process. 3D The real path of creation never ends with a model: a high-density triangle capable of printing, not necessarily for role animation; a model suitable for conceptual validation, and not necessarily for the size of a precise part.
The emergence of Hi3D Agent has accelerated the efficiency of grinding before the early design of the model to real world production.
The current industry ' s mainstream exploration of 3D Agent is based mostly on agreements such as MCP, where the large-language model calls for an external 3D tool, which is essentially a "connection layer" idea - LLM is responsible for issuing orders, and the tool is responsible for implementing single-point operations.
The UML model is good at understanding intentions, dismantling missions, but it does not understand geometry, does not understand thin wall structures, does not understand grid norms; and the professional 3D model is good at retrofitting forms, generating high-precision geometry, but it does not understand complex creative goals and does not actively advance processes.
To this end, Hi3D Agent chooses another path: to build a 3D work layer directly for creators, using a professional 3D model as the core base, linking intent, mission planning, model generation, continuous modification, engineering optimization, and scene adaptation to a continuous workflow.
In the design logic of Hi3D, there is a clear division of labour between the Universal Large Language Model and the Professional 3D Model: LLM is good at understanding what creators want to do, dismantling objectives, planning steps; and the Professional Hi3D Model is responsible for how to do it, translating intent into geometrically accurate, structured three-dimensional results.
The two work in depth and in combination, and eventually extend the single-point action that generates a model to a complete link to the completion of a work.
contrast, original / ChatGPT6 model generation / Hi3D Agent model generation
In the case of creators, this process may only start with a reference map, a semi-finished model, or even a vague idea, such as "refurbish the door structure, optimize the hard surface edge, and output printable formats", with the remaining iterative, modified and adapted work being advanced by Agent. Each step of the adjustment is accompanied by a comparable version, and the creator can stop, correct and decide on the choice.
It pursues a continuous process led by creators, which is perceived and controlled, and which can free creators from repetitive work and allow them to devote their time to aesthetic judgement, style grinding and creative expression.
After Agent finished the demo to the production level model, Hi3D tried to bridge the gap between the production level model and the end product.
Because Hi3D found that sometimes the same site is fractured and may be derived from too thin a geometric connection or from materials, placement or processing parameters. The algorithm team needs to know which failures should go into model improvement and which should go to process adjustments. The diagnosis is in the wrong direction and it is not necessarily effective to invest more training resources.
In the past, to sort out the root causes of the problem, there was often a need for multiple leaps between the software and the generation process, and after AI generated the first step of solving "0 to 1", grinding "from 1 to 100" remained highly manual.
To bridge the gap between generation and generation, Hi3D has built its own plants even in the Pearl Triangle, directly connecting AI model development and physical production processes. The company’s founder, Litt, would even take the investors to the factory, because, in his view, the value of AI would end up in real manufacturing, real delivery, real goods, and a modeling enterprise would also have to accumulate more knowledge of materials, equipment and product categories, how to better serve real production.
04
From making good products to making big cakes.
From V1.0 to V3.0, from the open source base to the Agent system, Hi3D has become more than one of the best 3D generation tools, much more than an AI 3D infrastructure.
Downward, Hi3D allows more developers and firms to innovate at lower cost on mature production floors through V1.0 Open Bottom Capabilities;
Upward, Hi3D continuously pushes the precision boundary of the commercial model through V3.0, 20483 body-grade geometry, 8K texture output, and constantly lifts the ceiling for professional creation and upper-level applications;
Horizontally, Hi3D provides access to the entire chain of ideas to delivery, upgrades the single-point tool to complete workflow; and opens the external Agent capability through the MCP interface to allow more workflow and more tools to access the 3D generation capacity of Hi3D. Allows 3D content generation to extend from a group of professional developers to ordinary creators interested in 3D content generation.
Three layers of capacity piled together to form a complete ecological closed circle: the researchers here do the technological hierarchies, the developers here do the sceneal innovations, the creators here do the work, the firms here do the work and the applications here the bigger the cake is, perhaps the real ambition of Hi3D.
* Headchart source: Hi3D
This article is an original article from Polar Park.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.