360 panoramas and 3D tours both let visitors look around a space on the web, but the data behind each and the way visitors move through the space differ. A 360 panorama is one full-sphere image per camera position: the visitor looks around from that position, then jumps to the next. A 3D tour is a single spatial model: the visitor moves the viewpoint freely and can section the model by floor or by unit. This article compares the two by data size, mobile behavior, production workflow and fit, and extends the spatial-application section of the 3D web design guide.
Scan-based Gaussian splatting is a third approach, with data captured from a real location; details are in 3D Gaussian Splatting on the web: bottlenecks and tiering. The three approaches can also be layered. The 3D property showroom website uses a 3D model to pick a floor and unit, then switches to 360 panoramas once the visitor is inside.
Comparing the three approaches
| Item | 360 panorama | 3D tour | Scan-based (Gaussian splatting) |
|---|---|---|---|
| Data source | Photographed panoramas, or offline renders from a model, output as equirectangular rectangles | Model and textures delivered from the modeling or rendering side, converted to glTF and split into packs on the web side | Gaussian scene reconstructed from multiple photos or video |
| Download size | Number of panoramas times the resolution of each; a single showroom panorama is 4096 x 2048 | Model packs plus textures, tiered by device; the main building loaded on phones in the showroom is about 7 MB | Raw files can reach hundreds of MB; one scan of about 800 MB holds about 15 million Gaussians |
| Interaction | Look around from a fixed point, move to the next position through hotspots | Free viewpoint, with sectioning, floor selection and unit selection | Free-viewpoint walkthrough; the scene carries no structure such as floors or units |
| Update cost | Re-shoot or re-render the affected positions when the space changes | Edit the model and repackage; the sectioning and floor-selection code is reused | Re-scan and rebuild |
| Mobile behavior | GPU memory for one decoded panorama grows with the square of the resolution | Controlled through texture tiers and model packs | Controlled through a budget on the number of resident Gaussians |
360 panoramas: capture or render, hotspot data and resolution
The standard output for a panorama is an equirectangular projection covering 360 degrees horizontally and 180 degrees vertically, so the aspect ratio is 2 to 1. Google's Photo Sphere XMP specification (page undated) supports only this projection, with a full-sphere example of 4000 x 2000 pixels; when an image is scaled, the metadata must be scaled with it, and an image that does not keep the aspect ratio should not be displayed as a panorama. The source can be a panoramic camera or an offline render of a model in rendering software. Both produce the same format, and the web side handles both the same way.
The showroom's interior panoramas are rendered at 4096 x 2048, with a separate 512 x 256 low-resolution base image. Entering a panorama loads the base image first; the high-resolution version is uploaded to the GPU in tiles and fades in over it in 350 milliseconds once complete. Panoramas are rendered per unit type, and every unit of the same type shares one set of positions, so the panorama count depends on the number of unit types and amenity spaces, not on the number of units: over 1,000 units and 10 unit types come to more than 70 panoramas. Units with a mirrored layout reuse the same panoramas, flipped horizontally in code.
The hotspots that lead to the next room are data. The rendering side delivers a camera position file for each position, and a build script aligns the camera coordinates to the interior floor plan and records, for each position, the bearing to each neighboring position, the floor-plan distance, and whether a wall lies between the two. For a position in line of sight, the hotspot is placed on that position directly; for a position behind a wall, the hotspot is placed at the center of the exit door. All of this is computed at build time and only read while browsing.
The limit of a panorama is the fixed point. Visitors can only rotate the view from a position; visitors cannot see the space between positions, change height, or cut through a wall. Viewing the same room from another angle takes another position, which means another panorama.
- 8 MB2048 x 1024about 11 MB
- 32 MB4096 x 2048Showroomabout 43 MB
- 128 MB8192 x 4096about 171 MB
Scope: equirectangular panoramas computed at 4 bytes per pixel in RGBA, uncompressed; the mipmap column applies MDN's figure of about 33% extra for square textures as an estimate. This is arithmetic, not a measurement on a device.
MDN's WebGL best practices (2026) note that WebGL has no portable way to query maximum video memory, so the budget should be conservative and the webglcontextlost event, fired when GPU memory runs out, should be handled. Panoramas are therefore tiered by device: the showroom caches 6 panoramas on desktop and 1 on phones, plus 8 low-resolution base images so that a transition always has something to show.
Prerendered environments: baking lighting ahead of time
Panoramas can also reduce the load of real-time rendering. One outdoor scene was rebuilt with an offline-rendered environment: an 8192 x 4096 panorama at 1.5 meters of eye height, rendered in Cycles at 128 samples per pixel with denoising, plus a near-ground top-down texture and a reflection map. The web page keeps only the main subject and a small amount of real geometry, and the panorama supplies the rest of the environment.
The measurement setup was a 390 x 844 phone viewport at a pixel ratio of 2 with touch dragging, run in a headless browser on a desktop GPU in two interleaved rounds. The machine carries about 1.5 to 2 milliseconds of noise, and the run is not an iPhone measurement. The GPU time of the two passes, main draw plus specular, fell from 8.13 milliseconds to 4.66 milliseconds, which cuts the per-frame work on phones by roughly a third; the phone download fell from 7.35 MB to 5.35 MB. The trade-off is in reflections: the subject's reflection in puddles on the ground is off by default, and turning it on costs about 1 millisecond more. Reflections on dry ground differ somewhat from the offline render, so the gain is fitted pixel by pixel against two renders.
3D tours: model source, sectioning and performance budget
The data of a 3D tour is a model. The glTF 2.0 specification (Khronos, 2021) positions glTF as an efficient, interoperable format for transmitting and loading 3D content, with small files and runtime processing efficiency among its design goals, which is why models on the web are mostly delivered as glTF. The production master is usually far larger than what a web page can load. The showroom's render master is about 1.4 GB and nearly 50 million faces; the exterior model used on the web is a reduced version of about 2.06 million faces, split into 4 model packs. The main building and its surroundings come from different delivery files, and the packs follow the source files directly.
- Render masternearly 50 million facesabout 1.4 GB
- Reduced exterior modelabout 2.06 million facesProcessed with Blender batch scripts
- Packs4 model packsTower shell, podium and platform, site ground, surrounding city
Scope: the showroom's exterior model; face counts are approximate, and the surroundings of about 870,000 faces are included in the 2.06 million.
Sectioning is an interaction that only a 3D tour offers. When a floor is selected in the showroom, a set of clipping planes cuts away the model above that floor slab. There are 3 clipping planes, attached once at load time to every material that needs cutting, so changing floors only changes the plane position values and needs no shader recompile. The cut building is shown with a semi-transparent shell that shares the same geometry, which avoids editing the hundreds of original materials one by one. A panorama cannot do this, because a panorama is one fixed image.
The performance budget of a 3D tour goes to textures and model packs. The showroom splits textures into 4 tiers of 512, 1024, 2048 and the original. Phones load the 512 tier, which comes to about 7 MB for the main building and about 8 MB for the surroundings; on desktop the 2048 tier is about 18 MB for the main building and about 13 MB for the surroundings. The surroundings, at about 25 MB in total, are not loaded by default and start downloading only when a button is pressed. MDN lists GPU compressed formats as a way to cut bandwidth and memory on mobile devices, and texture tiers and compression belong to the same family of techniques. Details are in how the 3D property showroom website is built.
The limit of a 3D tour is the model itself: someone has to build it or scan it. A model produced elsewhere must pass intake and acceptance before it goes on a web page, and pack names and component naming must be agreed in advance between the production side and the web side. The intake and acceptance process is covered in intake and acceptance for outsourced 3D models.
Scan-based: the third data source
Scan-based capture skips modeling and reconstructs from photos or video of a real location. The 3D Gaussian Splatting paper by Kerbl et al., 2023 represents a scene captured from multiple photos or video as 3D Gaussians and reaches real-time novel-view synthesis at 30 fps or more at 1080p. The bottleneck for web deployment is data size and overlap depth: in one scan of about 800 MB and about 15 million Gaussians, the average overlap is about 180 layers and the median Gaussian is only about 5 mm. The cost is mostly the number of layers blended per pixel, so the answer is offline merging and thinning, plus chunked streaming. The full measurements and tiering are in 3D Gaussian Splatting on the web: bottlenecks and tiering.
Scan-based capture suits places where the real location must be preserved as it is. The scene holds no structural information such as floors or units, so interactions like floor selection, sectioning and unit details need separate data.
Layering: 3D for floor selection, panoramas for interiors
- Confirm what visitors need to do in the spaceThe order of overview, narrowing the range, then inspecting detail
- Walk, section or pick floors3D tourThe data source is a model
- A close look from one spot360 panoramaThe data source is photographs or renders
- Keep a real place as it isScan-basedThe data source is photos or video
- From the whole into one room3D floor selection, then 360 panoramas insideThe showroom's path
Scope: selection criteria for the three approaches in this article.
The showroom joins both approaches on one path. The exterior is a 3D model, and buyers narrow the range from the whole building to a floor to a unit. Once a unit is chosen, the camera pushes in, the last frame is saved as a snapshot that scales up and fades out, and the panorama field of view widens from narrow to its default over about 1.3 seconds. Interiors are panoramas, with a floor-plan minimap in the lower left marking the current position and view direction. The exterior needs sectioning and floor selection, so it uses a model; the interior needs clear materials and furniture, so it uses panoramas. Each approach sits where it fits best.
The 3D world map of the online expo divides the work differently. 289 exchange schools stand as small houses on 10 zoned islands, with a single WebGL renderer for the whole page. The ground uses textures with lighting baked in, drawn with unlit materials, and the scene has no lights, shadow maps or post-processing. The cloud texture is 4096 x 2048 on desktop and scaled to 2048 x 1024 on phones. Precomputed lighting follows the same idea as the prerendered environment in the earlier section. How this scene is built is covered in front-end techniques for online expos.
Mobile: resource settings and tiering
Panoramas and models alike need a separate set of resource settings on phones. The showroom applies a different configuration at page load on touch-first devices: the pixel ratio cap drops from 2 to 1.5, the shadow map from 4096 to 1024, the environment reflection map from 512 to 256, and model textures are fixed at the 512 tier. The Draco decoding workers number 2 for the main building and 1 for the surroundings on phones, against 4 on desktop.
Panorama memory can be calculated ahead of time. One 4096 x 2048 panorama takes about 32 MB uncompressed and about 43 MB with mipmaps, multiplied by the number of panoramas cached at once. Which features stay on phones and which are desktop-only should be worked out item by item before development starts; the method is in memory budgets for mobile 3D.
Search and sharing
Panoramas and models are both drawn inside a canvas, and crawlers cannot read text inside a canvas. Information such as unit types, floors and position names should be provided separately in an HTML text layer, with each level at its own URL, so that it can be searched and shared. Prerendering, crawler routing and structured data are covered in search and AI crawler readability for 3D websites.
Situations where each approach is used less often
- Only three to five fixed viewpoints, and the space will not change soon: the data and maintenance for 360 panoramas are simpler, and a full model is unnecessary.
- No model exists and none is planned: a 3D tour needs a model source first, so 360 panoramas or a scan reconstructed from the real location are the alternatives.
- Sectioning or filtering by floor, unit or component is required: panoramas cannot be sectioned, so a 3D tour is needed.
- The space is browsed mainly on phones and has many positions: the GPU memory of each panorama must be calculated first, and the number cached at once limited.
What to prepare before the first discussion
- Space data: floor plans, an existing model or render files, or existing panoramas with a list of positions.
- Range to be viewed: how much detail the whole, a floor and a single space each need, and whether sectioning is needed.
- Device mix: the ratio of phone to desktop visitors, which sets the trade-offs in resource tiering.
- Update frequency: how often the space data is expected to change, which sets the pack and naming conventions.
Once these are ready, contact the development team and start from the scope of the 3D web design service to confirm item by item whether a 3D tour, 360 panoramas or a layered combination fits.
