This Week in 3D and AI: Atlas, Blender 5.3 and a Revit bridge
Made with AIThe clearest signal this week is that photo-based 3D is moving into tools designers already use: Blender now opens Gaussian splats natively, and World Labs released a world model that turns a few photos into a scene you can move through and export as a mesh. Research kept pushing toward geometry you can separate, measure and edit.
World Labs opens Atlas to everyone
World Labs released Atlas, its world model, to the public with four tools. Viewfinder rebuilds a scene from photos or generated images, Chisel lets you block out a graybox to steer the layout, Splats to Mesh converts the result to a mesh, and Relight changes the lighting. A developer API and a Blender integration come with it.
What it means: a few phone photos from a site visit could become a context model for early presentations, but the output is splats or a mesh, not walls and slabs you can dimension.
Sourcesworldlabs.airadiancefields.comNovember 13, 2026, as reported by Radiance Fields
Blender 5.3 reads Gaussian splats
Blender 5.3 shipped with native import of 3D Gaussian splats from PLY and SPZ files. Once inside, they can be processed with Geometry Nodes and rendered in Workbench, EEVEE and Cycles. The release also makes Vulkan the default on Windows and Linux and adds dispersion for glass in Cycles.
What it means: a splat capture of a room or a street can now sit behind your model in a free renderer, with no third-party add-on.
Sourceblender.orgNovember 10, 2026
SketchUp Importer for Revit announced
At its Dimensions conference, Trimble announced a SketchUp Importer for Revit. It reads component tags and classes and turns them into native Revit elements that can be edited and scheduled, instead of plain imported geometry. Trimble says it will be available in November 2026 for Pro Advanced Workflows and Studio subscribers.
What it means: tagging and classifying components in SketchUp now pays off later, because that is the information the importer uses to map a concept model into BIM.
Sourcenews.trimble.comNovember 10, 2026
OmniLatent generates objects already split into parts
A paper from the University of Hong Kong, Tencent, CUHK and Nanjing University describes a 1.9-billion-parameter model that generates shape, texture and part labels from one image in a single pass, with up to 64 parts per object. On a standard part benchmark it scores 74.18 mIoU with mask guidance, against 58.10 for the strongest earlier method in its table. The authors note it can still merge parts that should stay separate.
What it means: a generated chair whose legs, seat and back come out as separate pieces is much closer to something you can edit than one closed shell.
Sourcearxiv.orgNovember 6, 2026, posted to arXiv
Hugging Face shows how to fine-tune TRELLIS.2
Hugging Face published scripts and containers for fine-tuning TRELLIS.2, an open image-to-3D model, on your own meshes. Its test used 188 shoes from Google Scanned Objects and trained small LoRA adapters on the two geometry stages only. On 9 held-out shoes, F-score rose from 0.276 to 0.803 and Chamfer distance fell from 0.174 to 0.033. The two training runs took 7.2 and 3.6 hours on one A100 GPU.
What it means: a studio with a consistent product line could teach an open model its own shapes, though 9 test objects in one category is a small proof.
Sourcehuggingface.coNovember 12, 2026
Autodesk puts Flow Studio inside ChatGPT
Autodesk released Flow Studio MCP, a ChatGPT plugin that generates 3D models from text or an image, remeshes them to a target polygon count and retextures them. Text and image prompts return up to 4 draft variations. It requires a paid Flow Studio account.
What it means: concept objects can now start in a chat window, and Pro users can send a result to Fusion for refinement before printing.
Sourceblogs.autodesk.comNovember 9, 2026
One thing to try
Install Blender 5.3, import a splat of a space you know, and measure one door or wall against your site notes. A splat or a generated mesh often arrives without true units, so one known size is the fastest check before you model against it. If you need the object itself as editable geometry rather than a backdrop, Montea's models are built by AI and checked against your photo. Sizes come from the sizes you give, or from the maker's published dimensions when we find them; otherwise they're estimated from the photo, so check a size before you make anything. See how a photo becomes a SketchUp model, or read the page for visualization artists to see where it fits in a render workflow.
Have a photo of the piece you need? Model a Photo