Gaussian Splatting for Ecommerce: A Founder's Guide
A founder-voice guide to Gaussian splatting for premium DTC product pages, what it solves on mobile, where it beats a 3D model, and where it does not.
If you sell a premium product on a Shopify or headless storefront and your buyer is on a phone, the question stops being "should we show a 3D version of the product" and starts being "which capture method actually loads, looks honest, and lifts the conversion." Gaussian splatting is the one that quietly changes the answer for most premium DTC categories. Smartphones already drive about 45% of US web traffic, and a 12 MB GLB model does not survive a coffee-shop LTE connection.
What is Gaussian splatting and how does it apply to ecommerce?
Gaussian splatting is a 3D capture method that turns a small set of product photos into a scene a buyer can rotate on a product page. It uses a cloud of small, colored, semi-transparent shapes to recreate how light bounces off the real product. The result looks closer to a photo than a hand-built 3D model does, and it can be compressed small enough to load on mobile.
The technique was published as 3D Gaussian Splatting for Real-Time Radiance Field Rendering at SIGGRAPH 2023 by Kerbl, Kopanas, Leimkühler, and Drettakis at INRIA. For ecommerce, the part that matters is the practical math: a usable product splat now ships in 1 to 4 MB after compression, where a comparable photoreal GLB with PBR materials runs 8 to 15 MB.
Why this matters for premium DTC right now
Most premium DTC stores have the same conversion gap. Desktop converts at 3 to 5%. Mobile sits at 1 to 2% on the same product. The buyer is the same. The page is what changes.
The page changes because the desktop visitor sees photos at 1600 pixels wide, hovers over them, and reads spec text without thumb-scrolling. The mobile buyer sees the same photos shrunk to 400 pixels, gets no hover state, and stops scrolling when the page goes silent. A flat photo cannot answer "what does the back look like" on a phone. A 12 MB 3D model cannot load before the buyer leaves.
A small Gaussian splat can. It gives the buyer rotation, real lighting, and material honesty in a file small enough to ship over LTE.
What a Gaussian splat actually is, in plain terms
Picture a cloud of tiny colored dots floating in 3D space. Each dot has a position, a color, a transparency, and a soft fuzz around its edges. Stack a few hundred thousand of those dots in the right places and they look like the real product. Rotate the camera around them and the buyer sees the product from any angle.
A 3D model works differently. An artist sculpts a mesh, paints materials, and tunes the lighting until it looks right. The output is clean and editable. It also takes hours to make per product and does not always match the real thing.
A splat skips most of that. It captures what is already there.
Where splats beat a 3D model on a product page
Materials. Splats are honest about how light hits leather, brushed metal, and gemstones. A 3D model with PBR materials can get close, but a splat captures the real reflection, not a guess.
Time to capture. A small product splat takes 2 to 8 photos and a few minutes of processing. A 3D model takes a 3D artist a half day per product, longer for complex geometry. For a catalog over 50 SKUs, that math decides whether you actually ship 3D at all.
Mobile load. A compressed splat for a single product is 1 to 4 MB. An uncompressed photoreal GLB with PBR is 8 to 15 MB. On LTE, the splat reaches the buyer; the GLB does not.
Editability for non-3D teams. A merchandiser can recapture a product if the lighting changes. A 3D artist has to re-sculpt the mesh.
Where splats lose to a 3D model
Geometry-driven categories. Furniture, lighting, sport equipment, and bicycles sell on shape. A splat is photoreal but soft at sharp edges. A clean GLB with hard surfaces still wins for chairs, lamps, and frames.
Configurators. If a buyer needs to swap fabric on a sofa or change the dial color on a watch, you need a 3D model with separable parts. A splat is one frozen capture.
Animation. A splat does not animate parts. A 3D model can show a watch crown winding, a kettle boiling, or a car door opening. For categories where motion sells, splats are not the right tool yet.
AR placement at full size. AR Quick Look and Scene Viewer ship USDZ and GLB. Splats are getting native AR support in iOS and Android, but the path is still uneven. For furniture and lighting, ship a small GLB for AR and a splat for the in-page view.
What a splat-based product page needs to ship right
The capture is the easy part. The page is where most splat experiments fail.
Compress the file. A raw splat from training is 100 to 300 MB. A buyer-ready splat after standard compression is 1 to 4 MB. If your splat is over 5 MB on a single product, the conversion lift is gone.
Lazy-load the splat. The buyer should see the hero photo first, then the splat fades in when they scroll near it or tap to rotate. A splat that loads on page open eats the buyer's first second on the page.
Match the studio lighting to the rest of the PDP. A splat captured in soft daylight and rendered next to a hero photo shot under tungsten reads inconsistent. Set the capture lighting to match the brand's photography brief.
Give the splat a job in the story. A rotation by itself is a toy. A rotation that triggers the right proof point ("here is the stitch detail Allbirds reviewers ask about") is a sale. The page has to know what the buyer just looked at.
Ship an AR fallback. For furniture, lighting, and any large product, the buyer wants to see it in their room. iOS Quick Look and Android Scene Viewer still want USDZ and GLB. Ship a small GLB alongside the splat for the AR button.
What this looks like with two to eight product photos
The Oeave capture pipeline takes 2 to 8 photos of a product, processes them into a Gaussian splat, and compresses it to a 1 to 4 MB file the renderer can load over CDN.
The shoot rules that matter:
- Even lighting. Soft, diffuse light from two sides. No hard shadows, no direct sun.
- Clean background. White, gray, or a single color the splat training can ignore.
- Spaced angles. Walk around the product taking photos every 45 degrees. Add one from above. That is 9 photos for full coverage; 5 to 6 work for symmetrical products.
- Sharp focus. Phone cameras are fine. Hand shake at the edges is not.
A merchandiser can do all of that in 10 minutes per product. That is the part that decides whether 3D is realistic for a 200-SKU catalog.
The buyer flow that closes the loop
A splat alone does not lift conversion. The buyer flow does.
The pattern that works on premium DTC:
- Hero photo loads first. Static, fast, brand-correct.
- Buyer scrolls. The splat fades in and shows the product rotating slowly.
- Buyer rotates the splat by hand. The page tracks which side they linger on.
- The proof block under the splat updates to match what they looked at. Stitch detail, weight, materials, sound profile, whatever the category sells on.
- AR button is visible the whole time. Tapping it opens iOS Quick Look or Android Scene Viewer with a small GLB.
- Buyer adds to cart on a checkout that does not unmount the splat.
Most splat experiments stop at step 3. The conversion lift comes from steps 4 and 6.
Where to start
If you have not shipped 3D at all on your product pages, start with one product. Pick the one where buyers ask the most pre-purchase questions. Capture 6 photos. Process them into a splat. Put it on the PDP under the hero, with a single proof block that updates when the buyer rotates.
If you already have a 3D viewer app on Shopify and the conversion lift has not shown up, the issue is usually the file size, the studio lighting, or the lack of a proof block under the viewer. The honest 3D product viewer comparison covers which app fits which constraint.
If your category sells on geometry and you are tempted to skip splats entirely, do not. Even furniture and lighting brands ship a splat for the in-page rotation and a small GLB for the AR placement. The two are not in competition; they fix different problems.
The buyer is the same on every device. The page is what has to change.