原帖引用 · AI 翻译
🧵3/n Hi3D v1.0 是开放的基础模型，而 Hi3D 自有的闭源模型已达到 v3.0，正以每边 2048 个立方体的分辨率进行精细化处理。
新的 Hi3D Agent 将 3D 模型的生成、编辑和处理整合到一个流程中，模型上下文协议（MCP）服务器则让其他 AI 智能体能够直接调用 Hi3D。
Hi3D 表示，它已将用户的积分成本降低了 50% 以上。
从事 3D 算法工作的人更青睐它的几何形状。13 名具有专业 3D 算法经验的人，在 562 次判断中，将 Hi3D v1.0 与另外 8 个图像生成 3D 模型并排比较。他们有 83.5% 的时候选择了 Hi3D v1.0。
🧵2/n 大多数图像生成 3D 的测试只评估输入图像所展示的视角，因此团队还检查了每个结果的背面和底部。
例如，对于一个带翅膀的披甲人物，TRELLIS.2 塌成了一块平板，而 Hi3D v1.0 保留了相互连接的翅膀和四肢。
由于 1 张照片无法展示隐藏的侧面，这些部分仍是模型做出的最佳推测。
原文
🧵2/n Most image-to-3D tests only grade the view shown in the input image, so the team also checked the back and bottom of every result. On a winged armored figure for example, TRELLIS.2 collapsed into a flat panel, while Hi3D v1.0 kept the wings and limbs as connected parts. The hidden sides are still the model’s best guess, since 1 photo can’t show them.
这对开放式 3D 生成来说太棒了。去年的尖端技术，今年开源了
@Hi3D_AI 已将 Hi3D v1.0（原名 Sparc3D）开源。这款图像生成 3D 模型从各个角度看都能保持结构完整，而且在发布前就已用于 Hi3D 的线上产品。
> 任何人都可以在自己的服务器上私有运行它，将其接入内部系统，并在其基础上为自己的行业开发工具和应用。
> 代码和模型权重已发布在 GitHub 和 Hugging Face 上，包括商业用途在内均可免费使用，
大多数图像生成 3D 模型都按正面效果来评判，因此背面、底部以及翅膀、角和把手等细薄部分往往会变成模糊的块状、断裂或缺失。
Hi3D v1.0 接收图像以及一个粗略的形状草稿（来自另一个模型的“粗网格”），然后分 3 轮细化，分辨率依次为 512、1024 和 1536。
这里的分辨率是指物体周围的空间被划分成称为体素的微小立方体，基本上就是 3D 像素；1536 意味着每边有 1536 个立方体。
完整网格将包含超过 36 亿个立方体，因此模型采用稀疏方式运算，只处理表面附近的立方体，跳过空白空间和物体内部的实体部分。
每一轮中，“自剪枝”解码器都以精细 8 倍的尺度重建表面，并剔除最终判定为空的立方体。
训练数据从 50 万个增加到约 200 万个经过筛选的 3D 资产。
与包括 TRELLIS.2 和 Hunyuan3D-2.1 在内的另外 8 个图像生成 3D 模型相比，它在多视角保真度（0.75）和细节丰富度（0.841）上得分最高：前者检查形状从 6 个相机角度看是否保持一致且完整，后者评估精细的表面起伏和小部件。
它还保持了 90% 的切片成功率，即 3D 打印机软件能够将模型切分为打印层的比例；相比之下，唯一另一个达到 1536 分辨率的模型 Pixal3D 为 7%，TRELLIS.2 为 17%。
在 13 名具有专业 3D 经验的人进行的 562 次并排评分中，Hi3D v1.0 有 83.5% 的时候被选中。
它只输出形状（没有颜色或纹理），需要 NVIDIA GPU；在月用户数不超过 20 万或月收入不超过 100 万美元的情况下，代码和模型权重可免费用于商业用途。
Hi3D 1.0 现已开源
https://t.co/AaRHAfa3Bu
原文
This is so good for open 3D generation. last year’s cutting-edge technology, open-sourced this year @Hi3D_AI has open-sourced Hi3D v1.0 (formerly Sparc3D), an image-to-3D model that holds together from every angle and already ran in Hi3D’s live product before release. > Anyone can run it privately on their own servers, connect it to their internal systems, and build tools and apps for their industry on top of it. > Code and weights are on GitHub and Hugging Face, free to use including commercially, Most image-to-3D models get judged by how the front looks, so the back, the bottom and thin parts like wings, horns and handles often come out blobby, broken or missing. Hi3D v1.0 takes the image plus a rough draft shape (a “coarse mesh” from another model) and sharpens it in 3 passes, at 512, 1024 and 1536 resolution. Resolution here means the space around the object is cut into tiny cubes called voxels, basically 3D pixels, and 1536 means 1536 cubes along each side. That full grid would hold over 3.6B cubes, so the model works sparse, only touching cubes near the surface and skipping empty air and the solid inside. In each pass, a “self-pruning” decoder rebuilds the surface at 8x finer scale and drops cubes that turn out empty. Training data grew from 500K to about 2M curated 3D assets. Against 8 other image-to-3D models, including TRELLIS.2 and Hunyuan3D-2.1, it scored best on multi-view fidelity (0.75), which checks the shape stays consistent and unbroken from 6 camera angles, and detail richness (0.841), which grades fine surface relief and small parts. It also kept a 90% slicing success rate, the share of models a 3D printer’s software could cut into print layers, versus 7% for Pixal3D, the only other model at 1536, and 17% for TRELLIS.2. In 562 side-by-side ratings by 13 people with professional 3D experience, Hi3D v1.0 was picked 83.5% of the time. It outputs shape only (no colors or textures), needs an NVIDIA GPU, and the code and weights are free for commercial use up to 200K monthly users or $1M monthly revenue. And Hi3D 1.0 Is Now Open Source https://t.co/AaRHAfa3Bu