中文精读
Meshy的3D生成课：好看的转圈视频，距离可交付资产还有几步
42章经单期精读 · AI 3D与创作管线
嘉宾：胡渊鸣，Meshy.AI创始人，清华姚班本科、MIT计算机科学博士，身份依当期介绍｜主持：曲凯
发布：2023年12月17日（北京时间）｜时长：43分44秒｜整理：2026年10月18日
官方节目与收听入口
先说清“生成3D”到底交付什么
看见一张有立体感的图，或一段围着物体旋转的视频，未必得到可以修改、绑定动作并放进游戏的资产。胡渊鸣把本期从图形学基础讲到3D表示，再落到Meshy当时的产品与创业选择。核心线索是：输出要进入用户原有工作流，才能超出演示。
这是2023年底的技术与商业观察。市场规模、成本比例、领先地位和未来两三年的预测都来自当期访谈，未作独立验证，不是当前Meshy功能说明。转写中的市场比例有明显重复失真，本文不补数；清晰的机制与具体任务优先于宏大市场判断。
1. 同样一张画面，电影与游戏面对不同约束
节目线索：01:43–11:22。 胡渊鸣把图形学理解为受应用推动、合成数字内容的学科。影视可以离线渲染一张画面很久，最后输出固定视频；游戏必须快速回应玩家，不能把开放世界所有内容一次塞进内存，延迟与资源管理直接影响可玩性。
渲染在本期狭义上指把3D场景变成2D画面。灯光照到物体后继续传播，材质的反射、透明度、纹理共同影响像素。修改照片的光影可以改变视觉效果，却不同于从三维场景计算画面。这个区分帮助定位工具是在生成素材、模拟光，还是修改最终图片。
硬件、软件、算法与生态相互推动：更多算力让新效果可行，新的视觉要求又推动优化。不能只用“模型更大”解释每个进步。编辑推论：需求文档先写最终载体、是否交互、允许等待与资源预算；截图看起来相似的两项任务，验收可能完全不同。
2. 3D有多种表示，选择取决于后续要做什么
节目线索：11:30–21:35。 点云是空间中的点；体素是规则的小方块，可像切豆腐一样分格。NeRF（神经辐射场）用神经网络表示从不同角度看场景的样子；节目讨论的高斯表示以空间中带大小与方向等属性的元素组织场景。它们的便利与限制不同，不应排成从低级到高级的单一路线。
Mesh是网格，用相连面片表示表面，纹理贴图决定外观。胡渊鸣用立方体说明：六个正方形各切成两个三角形，可用十二个三角形表达。网格也可以有四边形等形式，动画师关心拓扑——面片怎样连接——是否便于编辑与变形，而不仅是最终看起来细腻。
当时主流游戏与影视流程广泛使用网格。嘉宾强调生成结果要与现有管线相容；这不是说新表示永远无用，而是研究里好看的形式不必然立刻适合现有生产。512×512的二维格子增加一维变成512³，元素量急增，也说明3D不只是给图像模型多加一个按钮。
编辑推论：问供应商输出格式之前，先问使用者要改哪里、怎样做动作、接什么工具、运行预算是什么。格式可导出只是入口，编辑和运行都通过才是完整交付。
3. 给白模涂色、生成物体、猜照片背面，是三类任务
节目线索：22:06–25:53。 文生贴图需要已有三维模型，像拿一座白色雕像再上颜色，主要改变表面。文生3D同时生成几何和贴图，既要知道形状也要决定外观。图生3D以照片约束正面，降低文字描述角色的困难，但仍要补出看不见的部分。
胡渊鸣明确说单张照片的背面要靠预测。训练中的多视角资料帮助模型学习正反关系，不代表真实测量到了某双鞋或某个人的背面。它与受控多照片重建，输入证据并不相同。
当时他认为中远景道具更容易用起来：远处的小物件、桌面苹果，不会被近距离审视；精细角色则还有差距。这是2023年的团队判断。编辑推论：先选允许近似的场景，再检查生成错误会不会影响故事、识别或交互。不能用一个远景苹果成功，证明近景人物已经能替代专业制作。
4. 数据和消费入口，分别限制技术与市场
节目线索：26:02–30:15、32:08–35:41。 嘉宾用数十亿图像与数百万3D模型作数量级对照，说明可得资料差异；同时强调网站模型有版权，看得见不代表能拿来训练。这里应分别问数据规模、质量、任务相关性与许可，不能把其中一项替代全部。
节目提供过游戏服装的用户生成内容试验：用户输入提示词，生成自己的外观。胡渊鸣认为这类玩家可能更重视“我能做出来”的乐趣，对精度要求不同于专业生产。未核的“全球第一”定位不作为事实，重要的是用户任务与质量线有区别。
普通消费者还面临另一道门槛：收到3D文件，可能连怎样打开都不知道。头显若普及，或许增加可交互内容的需求；Vision Pro会否像iPhone、2026–2028年会发生什么，是2023年的条件性预测，本篇不宣布已经兑现。
同理，视频已有大众消费场景，不代表故事生成已经解决。把两秒延到八秒、插入更多帧，也不自动得到叙事逻辑。技术成熟度和消费市场应分别检查，不能用其中一条的进步替另一条报喜。
5. 科研能力需要迁移到用户问题，原课题可以被推翻
节目线索：30:41–34:38、38:05–41:30。 胡渊鸣回顾自己从偏重物理方程，到接受数据驱动方法；传统图形与AI经验的交叉，使团队看到3D生成方向。他谈专业美术成本与市场空间，但数字是粗估，尚不能证明团队具体能取得多少收入。
更实际的创业反思是Why先于How：为什么做，比怎样做更早需要回答。问用户“你是不是有我认为的痛点”，容易得到礼貌同意。让用户讲已有任务、返工和付钱原因，才可能推翻创始人的设想。说愿付与真正支付，也不是同一层证据。
如果博士期间研究ABC，用户却要DEF，选择不一定是强推旧成果。胡渊鸣建议相信解决问题的能力可以迁移，而非与某个课题绑定。对实验客观，也要对自己的执念客观。编辑推论：保留专业能力与可复用系统，允许问题定义变化；转向前仍要判断新任务是否在能力与资源范围内，不能把“我能学”当自动成功。
CEO也需要做自己不舒服的事。他说自己内向，长时间交流后要恢复，仍需接触用户并为团队负责。这是个人经验，不是要求所有科研创业者模仿一种性格。创造虚拟世界是愿景，当前资产是否可用仍靠阶段结果证明。
从一件道具开始做整条交付测试
以下是编辑建议。让真实制作者带一个最近用过的道具任务，先讲旧流程，再试生成：
| 步骤 | 需要检查 | 失败可能说明 |
|---|---|---|
| 生成 | 正反面、比例、风格是否满足用途 | 输入约束或几何仍不足 |
| 导入 | 格式、贴图、尺寸能否进入目标工具 | 截图成功未完成资产交付 |
| 编辑与动作 | 能否局部修改，必要时正常变形 | 拓扑不适合后续工作 |
| 运行 | 面数、内存、加载与画面预算 | 好看但资源不合格 |
| 使用与付费 | 用户省了哪步，返工多少，是否复用 | 生成惊喜未转成持续价值 |
先把UGC娱乐与专业管线分开比较，再讨论替代比例。对于单图生成，看不见部分的不确定也应进入验收。这样的完整样例，比展示很多转圈视频更能定位下一次需要改进的步骤。
回听导航与关联阅读
- 01:42：图形学由什么应用推动。
- 12:36：点云、体素、NeRF与网格。
- 19:20：生成结果为什么要进管线。
- 22:25：三类3D生成任务。
- 29:49：消费入口的条件性未来。
- 38:48：Why先于How。
- 39:51：从原课题转向用户真正需要的问题。
相关主题：创意工具、资产交付、科研创业。可比较ACE Studio的可控修改与工作流和陈昱的技术实现证据。
来源边界：完整阅读所提供转写、旧精读及官方介绍后重组，未逐句复听。日期按北京时间，官方UTC为2023年12月16日16:00。术语以清晰语境解释，不将“高斯散射”等转写直接当严格定义；含混市场比例未采用。技术优劣、产品领先、成本和预测均归当期嘉宾，未扩成当前验收。
English edition
Meshy’s 3D Generation Lesson: A Good Turntable Video Is Still Several Steps from a Usable Asset
42 Zhangjing episode analysis · AI 3D and creative pipelines
Guest: Hu Yuanming, founder of Meshy.AI, Tsinghua Yao Class graduate and MIT computer science PhD, as introduced at the time | Host: Qu Kai
Published: December 17, 2023, Beijing time | Duration: 43 minutes 44 seconds | Prepared: October 18, 2026
Official episode and listening link
Define what generating 3D actually delivers
An image with depth or a video rotating around an object may not provide an asset that can be edited, animated and placed in a game. Hu moves from graphics fundamentals through representations to Meshy’s product and startup choices. The central thread is that outputs need to enter existing workflows to move beyond demonstrations.
These are technical and business observations from late 2023. Market sizes, cost shares, leadership claims and forecasts for the next few years remain contemporary statements without independent verification, not descriptions of current Meshy features. An obviously corrupted market ratio is omitted rather than reconstructed. Clear mechanisms and tasks take priority over large market narratives.
1. Similar images can face very different film and game constraints
Episode reference: 01:43–11:22. Hu describes graphics as an application-driven discipline for synthesizing digital content. Film can spend substantial time rendering a frame before delivering fixed video. Games must respond quickly to players and cannot load an entire open world into memory. Latency and resources affect playability directly.
Rendering in this discussion’s narrower sense converts a 3D scene into a 2D image. Light continues through a scene after hitting surfaces; reflections, transparency and textures affect pixels. Editing a photograph’s lighting changes appearance but differs from calculating an image from a three-dimensional scene. This locates whether a tool generates assets, simulates light or edits a final image.
Hardware, software, algorithms and ecosystems influence one another. More computing enables effects, while new expectations motivate optimization. Bigger models do not explain every advance. Editorial inference: Requirements should specify the final medium, interactivity, acceptable waits and resource limits. Similar screenshots can conceal very different acceptance conditions.
2. Representations should serve what happens afterward
Episode reference: 11:30–21:35. Point clouds represent spatial points. Voxels divide space into regular blocks, like diced tofu. NeRF, a neural radiance field, uses a network to represent a scene’s appearance from different directions. The Gaussian representation discussed organizes elements with attributes such as spatial size and direction. These approaches offer different tradeoffs rather than a simple ladder from primitive to advanced.
A mesh represents a surface through connected faces, with textures determining appearance. Hu uses a cube: divide six square sides into two triangles each, giving twelve triangles. Meshes can also contain quadrilaterals and other faces. Animators care about topology, how faces connect, because it affects editing and deformation beyond visual detail.
Contemporary game and film pipelines widely used meshes. Hu stresses compatibility with those pipelines. New representations are not permanently useless, but a striking research result may not immediately fit production. Moving from a 512-by-512 grid to 512 cubed also sharply expands element counts. Three-dimensional generation is more than another button on an image model.
Editorial inference: Ask what creators must edit, how motion works, which tools receive the output and what runtime budget applies before asking vendors about export formats. Export begins delivery; editing and execution must work too.
3. Coloring a model, generating an object and guessing a photograph’s back are different tasks
Episode reference: 22:06–25:53. Text-to-texture requires an existing model, like coloring a white statue. It changes the surface. Text-to-3D generates both geometry and texture. Image-to-3D constrains the front through a photograph, avoiding difficult verbal descriptions of a character, but must infer unseen parts.
Hu explicitly says a single photograph requires predicting the back. Multiview training can teach relationships between sides without measuring the actual back of a specific shoe or person. Controlled reconstruction from many photographs begins with different evidence.
He considered middle- and distant-view props easier to use at the time: a small object far away or an apple on a table may not receive close scrutiny, unlike a detailed character. This is a team assessment from 2023. Editorial inference: Start where approximation is acceptable, then inspect whether errors damage narrative, recognition or interaction. One successful distant apple cannot establish that close-up characters replace professional production.
4. Data and consumption entry points constrain technology and markets separately
Episode reference: 26:02–30:15 and 32:08–35:41. Hu contrasts billions of images with millions of 3D models to illustrate available scale, while stressing that website models have copyright. Visibility does not authorize training. Quantity, quality, task relevance and licensing require separate checks.
A game-clothing experiment lets users enter prompts to create their own appearance. Hu suggests these users may value being able to make something more than professional precision. An unverified “world first” claim is omitted. The useful distinction concerns different tasks and quality thresholds.
Ordinary consumers face another barrier: they may not know how to open a 3D file. More widespread headsets might increase demand for interactive assets. Whether Vision Pro would resemble the iPhone, and what might happen in 2026–2028, were conditional forecasts in 2023. This article does not declare them fulfilled.
Likewise, video already has mass consumption without having solved narrative generation. Extending two seconds to eight or inserting frames does not automatically create story logic. Technical maturity and market consumption need separate checks; progress in one cannot report success for the other.
5. Transfer research capability to user problems, even if the original topic is rejected
Episode reference: 30:41–34:38 and 38:05–41:30. Hu recalls moving from a strong preference for physical equations toward data-driven methods. The intersection of traditional graphics and AI helped his team identify 3D generation. Professional-art costs and market opportunities are rough estimates, not evidence of revenue the team can capture.
His practical startup lesson is to ask why before how. Asking users whether they have the founder’s assumed problem invites polite agreement. Existing tasks, rework and reasons for paying can instead overturn the premise. Saying one would pay differs from making a purchase.
If a PhD studied ABC while users need DEF, forcing the old result is not the only choice. Hu suggests trusting transferable problem-solving capability rather than binding identity to a topic. Objectivity about experiments should extend to personal attachments. Editorial inference: Preserve expertise and reusable systems while allowing problem definitions to change. Still assess whether the new task fits resources; willingness to learn does not guarantee success.
A CEO also performs uncomfortable work. Hu describes being introverted and needing recovery after long conversations while continuing to meet users and accept responsibility. This is personal experience, not a personality prescription for researchers. Creating virtual worlds remains a vision; usable assets require staged evidence.
Begin an end-to-end delivery test with one prop
This editorial exercise asks an actual creator to bring a recent prop task and describe the old process before trying generation:
| Step | Check | What failure may reveal |
|---|---|---|
| Generation | Sides, proportions and style fit the purpose | Insufficient constraints or geometry |
| Import | Format, textures and dimensions enter the target tool | A screenshot did not complete asset delivery |
| Editing and movement | Local modification and necessary deformation work | Topology unsuitable for later work |
| Runtime | Polygon, memory, loading and visual budgets | Attractive output with unacceptable resources |
| Use and payment | Steps saved, rework and repeated use | Novelty without continuing value |
Compare entertainment UGC separately from professional pipelines before discussing substitution. Uncertainty in unseen sides belongs in single-image acceptance too. A complete example locates the next improvement more effectively than numerous turntable videos.
Listening guide and related reading
- 01:42: Applications driving graphics.
- 12:36: Point clouds, voxels, NeRF and meshes.
- 19:20: Why outputs need production pipelines.
- 22:25: Three generation tasks.
- 29:49: Conditional futures for consumer entry points.
- 38:48: Why before how.
- 39:51: Moving from a research topic to users’ needs.
Related themes: creative tools, asset delivery and research entrepreneurship. Compare ACE Studio’s editable workflow and Chen Yu’s implementation evidence.
Source boundary: Reorganized after reading the complete supplied transcript, prior analysis and official introduction, without sentence-by-sentence audio checking. Publication uses Beijing time; the official UTC timestamp is December 16, 2023, at 16:00. Terms follow clear context rather than treating transcription variants as strict definitions. The corrupted market ratio is omitted. Technical comparisons, product leadership, costs and forecasts remain the guest’s contemporary statements, not present-day acceptance.