3D Generation

700,000 triangles down to 71,000, UV islands from 10,947 to 961: they squeezed the seven dirty steps after image-to-3D into one local PC

November 19, 20260 views阅读中文原文
700,000 triangles down to 71,000, UV islands from 10,947 to 961: they squeezed the seven dirty steps after image-to-3D into one local PC

The short version

MetAI, a Japanese AI company, released a desktop app called Studio3D on November 18. Drop in a photo or an illustration and it first generates a 3D model, then keeps going: part splitting, retopology, UVs, materials, lighting, camera previs and export — work that normally means switching between three or four tools — all happen on your own machine. Nothing goes over the network; models and photos never leave the computer. One pair of numbers the vendor publishes as a sanity check: after retopology, triangles drop from 699,999 to 71,290, and UV islands from 10,947 to 961.

What it actually is

Start with where image-to-3D gets stuck today.

Over the past year, plenty of models can turn a picture into a 3D object: Tripo, Meshy, Hunyuan3D, various open-source options. Speed is no longer the problem. The problem shows up when you try to use the output. A freshly generated model only looks right: hundreds of thousands of triangles, one fused solid with no separate parts, scrambled UVs, and a single texture pasted over the whole thing. The first thing that happens when it lands in Unity or Unreal is that somebody complains — it can't be rigged, its materials can't be swapped, it has no LODs.

Studio3D aims at exactly that gap. It is not another generation model; it is a local application that turns generated output into a deliverable asset. The company lays it out as eight stages, in order: Generate, Parts, Mesh, UV, Materials, Light & render, Previs, Export.

Why it matters

Three reasons.

First, it admits that generation is only a small slice of the pipeline. That is something much of the industry prefers not to say out loud: generation demos look great, but getting an asset into a game, a film shot or a product review still leaves seven manual stages.

Second, it treats "your data stays home" as the product, not a bonus feature. For teams working on unreleased car models, unannounced products or character designs nobody has seen yet, uploading artwork to somebody else's cloud API is a compliance line you cannot cross. Studio3D installs on your own PC or server, and the company states plainly that models and photos are never sent anywhere else.

Third, it offers engineering numbers you can check instead of "stunning results". Triangle counts, quad counts, UV islands, bake resolution and VRAM ranges are the kind of thing you can defend to a manager.

The eight stages, one by one

1. Generate. Four input types: single image, text, multi-view and batch. One photo is enough; feed it up to eight views and the sides and back come out more accurately. The app works out which view each photo represents. Several generation models are built in, oriented toward different goals — fast, top quality with PBR, faithful to the photo.

2. Parts. This is the closest thing to magic: you type part names such as hair, face or shoe, and the app looks at the model from 12 directions, finds each region and cuts the mesh. The published figures: text-driven splitting of 7 parts takes about 35 seconds; letting the AI decide produces a finer 19-part split in roughly 7–8 minutes. UVs and 4K textures survive the split, so you can swap materials per part or export GLB piece by piece.

3. Mesh. The generated triangle mesh is rebuilt with quads, and colour, roughness and normals are baked from the original at 4K. The vendor's comparison: 699,999 triangles become 71,290 triangles, rebuilt as 35,645 quads, with a look that is "almost the same". Clean-up, decimation, LOD chains, real-world scale (mm/cm/m) and pivots all live in the same panel.

4. UV. Choose a single atlas, per-part or UDIM layouts, set resolution, padding and texel density, then check the result in a checker view. Published numbers: UV islands go from 10,947 to 961. Texel density can be locked to your sector's target, for example 512 px/m.

5. Materials. About 30 presets in seven categories: automotive, metal, glass, organic, plastic, toon and special. Assign automatically per part, let the AI repaint textures, then bake and export the maps.

6. Light & render. Five studio presets; HDRI, lights, floor and background are adjustable. Lighting environments are generated inside the app, so there are no HDRI packs to download. It outputs PNG stills, turntable videos and final renders, with tone mapping, depth of field and anime outlines. Your own .hdr and .exr files can be loaded too.

7. Previs. Frame the shot, add camera keys, check the move on a timeline, then export colour and depth PNG sequences plus a camera path JSON — the base plate for a video generation pass. This stage exists for people doing AI video.

8. Export. Eight presets: Unreal (FBX), Unity (FBX), Godot (GLB), Web/three.js (GLB), USD/Omniverse, AR on iOS (USDZ), 3D printing (STL), and OBJ with MTL. A Blender add-on sends models back and forth directly. There are also four industry presets — automotive, games, anime, product/3D studio — that switch units, lighting and default materials.

How those numbers were measured

The test conditions matter: the company notes that the figures were measured on a workstation with an NVIDIA RTX 6000 Ada (48 GB of VRAM), and that results vary with input images and settings. That is a professional card, not a consumer GPU.

On hardware requirements: an NVIDIA GPU is mandatory, and depending on which generation model you use, plan for roughly 6 to 32 GB of VRAM. That is a wide range, which tells you "fast" and "top quality" sit far apart in memory cost.

How to use it

For an individual or a small team, the path looks like this:

  1. Install it on a machine with an NVIDIA card (the company handles installation and training, on site or remotely, with a manual).
  2. Bring in material: a photo, an illustration, or just text. If you already have a model, import GLB, OBJ, STL or PLY and enter mid-pipeline at part splitting, retopology or UVs instead of starting from generation.
  3. After generating, split the parts you need to touch individually.
  4. Run retopology and UVs, locking texel density to your target engine's convention.
  5. Pick material presets and render lighting stills for review.
  6. Export for the target platform; for Blender, use the add-on to send models back and forth.

Who it is for

  • Indie games and small studios that need a batch of props or character parts ready for the engine without standing up a modelling outsourcing pipeline.
  • Industrial design and product review teams that only have photos or sketches and want to judge materials and lighting at real-world scale.
  • Anime and film pre-production that needs quick previs shots plus colour and depth sequences for a downstream video model.
  • 3D printing, since STL is one of the export presets.
  • Teams with confidentiality constraints: if designs cannot leave the building, every cloud SaaS option is already out.

When it is not a good fit

  • Anyone on a Mac or an AMD card: the NVIDIA requirement is hard.
  • Anyone who wants pay-per-use and nothing installed: this is local deployment with a subscription or a one-time purchase, not an API meter.
  • Anyone expecting one-click finished assets: eight stages means you need some grasp of retopology and UVs, or the output is still unusable.
  • Anyone who needs large-scale cloud concurrency: it runs locally, so one machine is one machine's worth of compute.

Cost and time

No price has been published. The product page's FAQ says a proposal is made to fit how you want to work — subscription, one-time purchase, or buying their 3D production service — and your environment. So the billing shape is known; the number is not.

On timing, the published reference points are 35 seconds for a 7-part text-driven split and 7–8 minutes for a 19-part automatic split. Generation, retopology and UV times are not given as single figures because they depend heavily on input and settings. The product page's line is "from idea to usable 3D data the same day".

Things to watch

  • Do not treat vendor numbers as your numbers. 699,999 to 71,290 was measured on a 48 GB professional card; your results will differ, especially at the 6 GB end.
  • Verify "almost the same" yourself. Retopology plus 4K baking tends to lose detail in close-ups and normal maps; check in the target engine before delivery.
  • Local-only is both the advantage and the ceiling. Data stays in, but compute is not shared; several people working at once means several machines.
  • It does not replace a modeller. It removes repetitive labour; whether the topology suits later animation and whether the UV layout is sane still needs a human call.
  • Mind the language and support footprint. The interface ships in Japanese, English and Spanish — Chinese is not listed — and documentation and installation services currently centre on the Japanese market.

Next step: your first move

If you are evaluating this route, do not buy first. Take one real asset with real delivery requirements — not a statue you would only use in a demo — and run it through generate, split, retopology, UV and export into the engine you actually use. See whether the look and the polygon count clear your own acceptance bar. Then do the math: convert the hours that step used to cost into money and compare it with the subscription or purchase price.

If that first pass shows the topology it produces does not suit your animation needs, the route is not viable for you yet — and that has nothing to do with how strong the model is. It is a question of where the tool sits in the pipeline.

Sources

  • MetAI official blog (November 18, 2026): https://www.met-ai.tech/en/blog/26-11-18---studio3d-local-3d-app/
  • Studio3D official product page (measured figures and hardware requirements): https://www.met-ai.tech/en/studio3d/
  • Official sitemap (three-language pages, same date): https://www.met-ai.tech/sitemap.xml

All facts and figures above come from those official pages. Figures described as vendor-published are MetAI's own measurements, taken on a machine with an NVIDIA RTX 6000 Ada (48 GB).

Sources

Comments

0 published
Log inorsign up to leave a comment.

Loading comments...