Two to eight photos work for most products. Here are the capture rules that matter more than the count, and the shoot checklist for premium DTC merchandisers.

If you sell a premium product on Shopify or a headless storefront and you have started looking at Gaussian splats, the first question is always the same: how many photos do I need to take? The honest answer is fewer than most founders expect, and the photo count matters less than the rules around how you take them.
Two to eight photos shot from spaced angles around the product, plus one from above, are enough for a usable Gaussian splat of most small to medium products. More photos help complex geometry like leather goods with deep stitching or audio equipment with grille detail. Fewer photos work for symmetrical products like ceramics, glassware, and cosmetics packaging. The rules that matter more than the count are even lighting, a clean background, sharp focus, and visible overlap between adjacent shots.
The technique that powers this was published as 3D Gaussian Splatting for Real-Time Radiance Field Rendering at SIGGRAPH 2023 by Kerbl, Kopanas, Leimkühler, and Drettakis at INRIA. The pipeline learns where the camera was for each photo, then places semi-transparent dots to recreate the scene. With overlapping coverage, even a small set of photos gives the system enough signal.
A merchandiser with a phone and 10 minutes can shoot a buyer-ready splat. The same merchandiser with a DSLR, a tripod, and 30 minutes can shoot a worse splat by breaking one of these rules.
Even lighting. Soft, diffuse light from two sides. No hard shadows. No direct sun through a window. The splat learns the lighting along with the geometry, so harsh shadows bake in and make the buyer-side rotation look wrong from angles you did not photograph.
Clean background. White, gray, or a single color the splat training can ignore. Cluttered backgrounds confuse the camera-pose estimator and add fuzzy artifacts to the final splat. A roll of studio backdrop paper costs $30 and saves you a week of artifacts.
Sharp focus. Phone cameras are fine. Hand shake at the edges of the photo is not. If you cannot hold the camera steady, use a tripod or rest the phone on a stack of books. A splat trained on a blurred photo produces a blurred splat.
Visible overlap. Each photo should share about 30% of its frame with the next photo. The splat trainer needs that overlap to figure out where the camera was when each photo was taken. A photo set with no overlap produces a splat with seams where the trainer guessed wrong.
These four rules account for most of the quality difference between a useful splat and a throwaway one. Photo count comes after.
Two to eight is the working range. Within that, the right count depends on the product.
Cosmetics packaging, candles, glassware. 2 to 4 photos. These products are symmetrical or near-symmetrical. The splat trainer can fill in the back from the front and sides. Add one shot from above if the cap or lid matters.
Ceramics, kitchenware, small accessories. 4 to 6 photos. One every 60 to 90 degrees around the product, plus one from above. Most premium kitchenware buyers want to see the rim and the base, so include both.
Jewelry. 6 to 8 photos. Surface detail matters at premium price points. Include extra macro shots of any cut, stone, or finish detail the buyer will ask about. Do not skip the underside on rings and pendants.
Leather goods, watches, audio. 8 to 12 photos. Stitching, grain, and material transitions need overlap from multiple angles. Watches in particular need the case, the dial, the crown, and the back, plus a side profile.
Furniture and lighting. 12 to 20 photos. Larger products with crisp edges need more coverage. Walk around at three heights: low, eye-level, and slightly elevated. For a sofa, that is roughly 15 to 18 photos including angle shots of the legs and the seat cushion seams.
For a 200-SKU catalog with a mix, plan a shoot that averages 6 to 8 photos per product. A merchandiser working through that list can ship a buyer-ready splat for the whole catalog in a week.
Run this checklist on your first product before you scale.
The whole shoot for a small product runs 8 to 12 minutes once the lighting is set. Resetting the lighting between products is the slow part, which is why batch shoots work better than one-product-at-a-time.
The splat training itself is GPU-bound. On a typical pipeline, a small product trains in 10 to 30 minutes and a large product in 30 to 60. The output before compression is 100 to 300 MB of raw Gaussian data.
The buyer never sees that. The splat goes through a compression step that drops it to 1 to 4 MB while keeping the rotation usable on a phone. PlayCanvas SuperSplat is one open-source tool for that compression and clean-up step. Most ecommerce splat services run the same kind of pipeline server-side and hand back the compressed file ready to embed.
If you have not shot a splat before, start with a single product. Pick one where buyers ask the most pre-purchase questions. Shoot 6 to 8 photos following the checklist above. Process them through your splat pipeline. Embed the result on the product page under the hero photo with a single proof block that updates when the buyer rotates.
If the splat looks blurry or noisy, the cause is almost always lighting or background, not photo count. The founder's guide to Gaussian splatting for ecommerce walks through the splat capture pipeline end to end. The splat vs 3D model decision covers when to ship a splat instead of a sculpted model in the first place.
The buyer is paying for the surface. A clean shoot is what shows it.