이미지: X — 미디어·생성AI
Summary
- Intangible has added Gaussian splat capture support to its 3D scene builder
- Users can capture a real kitchen countertop, import it into Build mode, and assemble a set by snapping and parenting objects to it
- Only the camera changes to produce multiple angles of the same space, reordering the traditional production workflow
- 기능
- 가우시안 스플랫(Gaussian Splat) 캡처 지원 추가
- 도구
- Intangible 3D 씬 빌더 Build 모드
- 테스트 사례
- 부엌 아일랜드 조리대 캡처
- 오브젝트 처리
- 스냅(조리대→사과)·부모화(parent) 지원
- 메타데이터
- Details 패널에 이름·설명·참조 이미지 추가 가능
- 발표
- Umesh, X, 2026년 8월 15일
Shooting a kitchen countertop and using it directly as a 3D set
Intangible, a 3D content creation tool, has added a feature that lets users bring Gaussian splat captures directly into its scene builder for editing. According to a thread posted on X on August 15, 2026 by AI creator Umesh, a capture file scanned from a real space can be imported into Intangible's Build mode, where objects can be placed on top of it and multiple scenes can be produced simply by changing the camera angle. He walked through the entire process using a kitchen island countertop as an example.
Gaussian splatting is a 3D capture method that stitches together multiple photographs to reconstruct a space as a dense cloud of colored points. From photos taken at different angles alone, the shape, texture, and even lighting of a space can be reconstructed in three dimensions. Unlike traditional 3D modeling, which requires polygons to be drawn by hand, this technique produces results from actual footage, giving it a level of realism that has recently drawn attention across the gaming and video industries.
How it works in practice
The thread posted by Umesh outlines the following workflow:
- Import the splat — After importing a captured 3D splat file, Intangible processes the capture and shows a screen for checking its orientation and scale.
- Place it in Build mode — Once confirmed, the splat is placed directly into the Build mode scene. From this point, the splat is no longer just a reference image but an object that makes up the scene. It can be selected, moved, viewed from different angles, and built upon.
- Add and snap objects — Placing an apple on the countertop causes the object to automatically snap to the splat's surface. The tool recognizes the countertop as a surface and positions the apple precisely on top of it.
- Parent objects — Parenting added objects to the splat keeps the entire set organized as a single unit without things drifting out of place.
- Add context in the Details panel — Opening the splat's Details panel allows users to add a name, description, and reference images. This step goes beyond simply displaying the captured space visually, telling the tool what the space is actually for.
- Move the camera and shoot — With the same kitchen, the same arrangement, and the same spatial reference maintained, multiple scenes can be produced by changing only the camera angle.
In the bottom right of the interface shown, a persistent input field reads "Describe a scene, watch AI build it in 3D," suggesting that even after importing a splat, users can layer additional elements on top via text prompts. However, the sign-up requirements, pricing plans, and supported capture formats for this feature were not addressed in the thread.
How this differs from prompt-only generation
| Category | Traditional generative 3D workflow | Splat-based workflow |
|---|---|---|
| Starting point | Reimagine a world from scratch via text prompt | Import a real space captured as-is |
| Changing angles | Must hope the next scene resembles the same space | Camera moves while the same spatial reference is preserved |
| Object placement | Generated fresh each time | Placed to match the existing set via snapping and parenting |
| Spatial information | Relies solely on prompt text | Recorded separately via name, description, and reference images |
Umesh described this shift as a new 3D production sequence running from capture to import to set dressing to direction to shooting, stating: "Reality is now an editable material" (Umesh, X).
Editor's take
What makes this feature interesting isn't generation—it's the "reference point" it provides. Text-to-3D tools up to now have focused on reimagining a space from scratch with a single prompt. The problem has always been what comes after: change the angle, and wall thickness shifts, lighting changes, and the kitchen you built yesterday becomes a different kitchen today. Embedding a Gaussian splat as an actual object within the scene structurally eliminates this problem, since all that's needed is to rotate the camera. This should be read as a signal that generative 3D is taking a step from being a "creation tool" toward becoming a "capture tool."
Anyone who has actually used prompt-based 3D generation tools in production work knows this frustration. When a client needs to see three different angles of the same space, regenerating each one from scratch and then spending more time in Photoshop matching colors and proportions than an actual photo shoot would have taken is a common experience. Using a splat as a fixed reference eliminates this mismatch entirely. It's less a generational leap than a shift in the nature of the problem itself.
Teams producing interior, furniture, or real estate content domestically should take note of this workflow. Once a store or showroom has been scanned with a smartphone, seasonal prop swaps or new product placement shots can be produced by simply changing objects within the same space, rather than reshooting from scratch every time. That said, this disclosure is at the level of a single user's personal thread, and pricing plans, smartphone scan support, and commercial usage terms have not yet been announced.
In the coming weeks, other 3D generation tools are likely to follow suit with features that place real-world captures directly into scenes. Gaussian splatting itself is already a proven technology in the gaming and VFX industries, so the remaining competition is likely to narrow down to how seamlessly generated objects and real-world captures can be blended within a single scene.
