How to prepare product images and 3D files for AI tools: formats and naming
File formats, consistent images and identifiers that make visual assets usable by generative and 3D tools.
KEY TAKEAWAYS Summary by the editors
- Choose 3D formats by destination: GLB for web viewers and commerce pages, USDZ for Apple AR, FBX or OBJ for exchange between design tools.
- glTF, the specification behind GLB, supports physically based materials and textures and became an ISO/IEC standard in 2022.
- Consistent image rules (aspect ratio, background, angles) matter more to downstream tools than individual image quality.
- Each sellable size and colour variation needs its own GTIN, and file names should carry the style and colour identifiers used in your product master.
- Keep an original, uncompressed master of every asset and convert copies, because repeated conversion can degrade files.
To prepare assets for AI tools, standardise the images, choose 3D formats by destination, and name every file with identifiers that match your product master. AI and 3D tools rely on consistency: the same angles, backgrounds and identifiers across a collection. Heavy post-processing is less important than a clear, repeatable convention.
Which 3D file formats should you use?
VNTANA's overview recommends choosing the format by destination. GLB, an open Khronos format, is suited to web viewers and commerce pages and stores geometry, materials, textures and animation in one file. USDZ, from Apple and Pixar, fits iOS augmented reality. FBX is widely used for animation and game pipelines and as an exchange format, and OBJ for universal interchange, though OBJ has no animation support and keeps textures in a separate material file. VNTANA advises choosing correctly from the start, since converting files repeatedly can cause problems. Khronos notes that glTF supports physically based material properties such as base colour, roughness and metallic, and became ISO/IEC 12113:2022.
| Format | Best for | Note |
|---|---|---|
| GLB (glTF) | Web viewers, 3D commerce | Open standard; geometry, materials and textures in one file |
| USDZ / USD | Apple AR, try-on on iPhone | Tied mainly to the Apple ecosystem |
| FBX | Animation, game engines, exchange | Proprietary (Autodesk) |
| OBJ | Universal interchange | No animation; textures in a separate file |
Garment design tools have their own export lists. Marvelous Designer, for instance, documents import and export of OBJ, FBX, USD and Alembic, so check which formats your design tool actually supports before planning the pipeline.
How should product images be prepared?
Marketplace and platform guidelines show the pattern. The Folklore's visual guidelines specify a 2:3 vertical aspect ratio, a lead image on a white backdrop (ghost mannequin for fashion), and further shots covering front, back, sides and close-up details. Whatever your own standard, document it and apply it to every style: same ratio, same background, same angle order. Consistent images make it easier for AI tools to classify products, generate variations or train on your own catalogue.
How should files be named?
There is no universal naming convention, so define one and keep it. Build names from identifiers in the product master: style number, colour code, view and version. A style number identifies a design, while each size and colour variation of a sellable product needs its own GTIN, according to GS1. SPS Commerce adds that internal SKUs are unique only inside a company, so do not substitute them for GTINs in data shared with partners. Avoid spaces and special characters, and keep names stable when a file is updated, using a version field instead.
- Style number, for example a short alphanumeric code held in the product master.
- Colour code, not a free-text colour name.
- View or angle, in a fixed list (front, back, side, detail).
- Version or date of the file.
- File extension in lowercase.
What checks should come before using AI tools?
- Keep an unmodified master file and convert copies.
- Confirm you hold the rights to use images and 3D assets with third-party tools, and check each tool's terms on data use and storage.
- Test one style end to end before processing a collection.
- Record metadata: season, category, colour, material and the origin of the asset.
- Review output against the source to catch colour shifts, since screen colour differs from fabric colour.
How can quality be checked at scale?
Once a convention exists, quality can be checked automatically: file names that do not match the pattern, images outside the agreed aspect ratio, missing views and 3D files without textures can all be flagged by a short script or a digital asset management system. Manual review then focuses on what automation cannot see, such as colour accuracy and garment fit in a render.
Keep a simple log for each batch: source, date, tool used and settings. If an AI tool alters an asset, store the output as a new version and keep the original. This is good practice for audit and for any later question about where an image came from, which is increasingly relevant as transparency rules for AI-generated content develop.
- Run automated checks on names, ratios and completeness.
- Review a sample visually for colour and fit.
- Version every output and keep the original.
- Record the tool and settings used.
Which data should travel with each asset?
An asset without metadata is hard to reuse. At a minimum, each image or 3D file should be linked to the style number, colour, season, category, material and the GTIN of the variants it represents. For 3D garments, add the size the model represents, the fabric simulation settings used and the date. AI tools that classify, generate or search assets work better with structured attributes, and the same attributes help your own teams find what they need.
- Style number, colour code and season.
- Category, fit and material.
- Size represented (for 3D and fit models).
- Rights information: who created it, where it may be used.
- Version and date.
What should you avoid?
Avoid heavy edits that bake in colour changes, because they make the asset a poor record of the product. Avoid exporting the only copy in a proprietary format that your future tools may not read: keep an interchange format such as OBJ or FBX alongside the working file. Do not rely on a single conversion for everything, since VNTANA warns that converting files many times can cause problems. Finally, avoid putting unreleased collections into external AI tools without checking terms, because assets may be stored or used by the provider according to its own policy.
Start small. Prepare one category end to end, measure how long it takes and where errors arise, and only then extend the convention to the full catalogue. The first pass nearly always reveals naming or metadata gaps that are cheaper to fix on a hundred files than on ten thousand.
Frequently asked questions
Which 3D format is best for fashion e-commerce?
GLB is widely recommended for web viewers and commerce pages, with USDZ for Apple AR. The right choice depends on where the asset will be shown.
What is glTF?
glTF is a royalty-free specification for transmitting 3D models, maintained by the Khronos Group, with support for physically based materials. It became ISO/IEC 12113:2022.
Do I need a different GTIN for each size and colour?
Yes. GS1 states that each product variation needs its own GTIN, for example one per colour for a one-size item.
Should I use the same naming for images and 3D files?
Yes where possible. Using the same style and colour identifiers across all asset types makes it easier to link assets to the product master.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
