Rights in Product Images: Photographers, Brands and Retailers

This is not legal advice. Treat the points below as diligence questions and clause topics for your counsel and the licensor.

Why ownership is rarely where buyers expect

A retailer's product pages look like the retailer's content. Behind them, images arrive from many directions: brands upload packshots through supplier portals, content-syndication services distribute brand assets to many retailers at once, the retailer's own studio or a contract studio shoots some items, marketplace sellers upload their own, and customers add photos to reviews. Each route has its own owner and its own terms, usually written for displaying and selling the product, not for training models.

For a humanoid program this matters because catalog images are the most accessible object data, as product catalog data for robot perception explains. Licensing them from a party that cannot grant the rights leaves the risk with you.

Where product images come from

Two neighboring pages go deeper on specific holders: photo studio archives, where studios' client contracts govern reuse, and trademarks in training images, which treats brand marks as a separate question from copyright.

Layers of rights in a single packshot

One photo of a cereal box can involve several holders at once:

1. The photograph, typically the photographer's or studio's unless assigned.
2. The packaging artwork it depicts, typically the brand's or its design agency's.
3. Trademarks and logos on the pack.
4. People and places, in lifestyle or in-store images.
5. Contract terms under which the retailer received the image, which may restrict use regardless of who owns copyright.

A license from the retailer can only carry the layers the retailer controls. Ask which layers each image set clears.

The training question and why express permission helps

The US Copyright Office's report on generative AI training, released in pre-publication form, analyzes fair use and examines licensing options for training on copyrighted works (U.S. Copyright Office, 2025). Robotics uses go beyond training, and some reproduce the images themselves, so a written grant naming each use is the practical route.

Open licenses and labels

Openly licensed images and assets still carry conditions. Creative Commons describes BY-NC-ND as its most restrictive license, and notes that CC licenses do not affect exceptions and limitations such as fair use (Creative Commons). Whether masks, crops, textures or renders made from an image count as adaptations is a question for counsel. Even permissive assets carry attribution duties: Google Scanned Objects assets are released under a Creative Commons license, CC-BY 4.0 according to a redistribution page (Downs et al., 2022).

Labels are unreliable. Dataset cards carry a license field in their metadata that is displayed on the dataset page (Hugging Face), but that field is only as good as whoever filled it in. A large audit of dataset licensing cut the share of "Unspecified" licenses from 72%+ to 30% by re-annotating them (Longpre et al., 2023), which shows how often a license label is missing even when the terms can be traced. Checks for public datasets are on public dataset commercial license checks.

Uses and derivatives the license should name

• Training and fine-tuning recognition, retrieval and attribute models.
• On-robot galleries. Reference images stored on deployed robots for retrieval are copies distributed with your product; name this use explicitly.
• Synthetic data. Compositing product images into generated scenes, or extracting textures for 3D assets.
• Derived labels. Masks, boxes, embeddings and attribute labels: who owns them, and do they survive termination?
• Evaluation and publication. Whether sample images may appear in papers, benchmarks or sales material.
• Retention. What happens to images of discontinued products, and to models trained on images later withdrawn.

General robotics clause topics are covered in licensing terms for robotics training data; SourceX's licensing terms glossary defines terms such as sublicense.

Illustrative image-rights manifest

# Illustrative only — synthetic parties and counts; not a legal template
delivery: catalog_images_retailer_R31_v1
image_sets:
  - set: R31_studio_2023_2026
    count: 41,200
    creator: retailer in-house studio (employees)
    rights_basis: retailer owns photographs
    packaging_artwork: brand-owned; depicted incidentally
    status: included
  - set: R31_contract_studio_CS4
    count: 12,900
    creator: contract studio CS4
    rights_basis: assignment clause in 2022 contract (copy on file)
    status: included
  - set: brand_syndicated
    count: 63,000
    rights_basis: supplier-portal terms limited to display and sale
    status: excluded pending brand permissions
  - set: customer_review_photos
    status: excluded (platform terms; people and homes visible)
per_image_fields: [image_id, set, creator, rights_basis, sha256, gtin]
permitted_uses: [training, on_robot_gallery, synthetic_compositing, internal_evaluation]
takedown: replace or delete flagged images within an agreed period; notify licensee

Delivery mechanics

• Per-image rights flag and creator field, so exclusions can be enforced in training pipelines.
• Set-level evidence on file: studio contracts, assignment clauses, brand permissions.
• Hashes and manifest for acceptance and for proving deletion later.
• Exclusion list and confirmation that excluded sets are absent.
• Embedded metadata such as photographer credits preserved for provenance rather than stripped.
• Takedown process for images later found to be outside the licensor's rights.

If ownership is too tangled, commissioning new photography under work-for-hire terms can be simpler; see the multi-view product photography specification.

Limitations

A rights manifest is only as reliable as the supplier's records. Older images often lack creator fields, and brand-supplied images can be filed as in-house shoots, so some third-party content may survive review. Deleting a flagged image after delivery does not undo what a model has already learned from it, which is why takedown terms belong in the license. Clear rights also do not change what the images can teach: packshots support detection and identification, not how a product behaves when handled.

Questions to ask suppliers

1. Which share of your product images did your own staff shoot, which a contract studio, and which the brands supplied?
2. Do your studio contracts assign rights to you or grant a limited license?
3. Under what terms do brands supply images, and do those terms permit uses beyond selling?
4. Can you exclude brand-supplied, marketplace and customer images from the delivery?
5. Were model and property releases obtained for lifestyle images?
6. Can you flag each image with its creator and rights basis?
7. Will you permit reference images to be stored on deployed robots?
8. How will you handle a takedown request from a photographer or brand after delivery?

Discuss your requirements

SourceX checks the data and the supplier's licensing permissions, agrees pricing and allowed uses in a license, and coordinates delivery; nothing is contracted until a supplier agrees. Describe the product images and uses you need licensed to SourceX, including on-robot galleries and synthetic data. It reviews the request and follows up about potential sources; a request does not guarantee a match. See also the product and object data hub, the humanoid training data guide and the page on licensing product catalogs and descriptions.

Explore sourcing and licensing options with SourceX

Sources

• Copyright and Artificial Intelligence, Part 3: Generative AI Training (pre-publication version) — fair use analysis and licensing options for AI training.
• About CC Licenses — restrictive CC terms and their relation to fair use.
• The Data Provenance Initiative: A Large Scale Audit of Dataset Licensing & Attribution in AI — how often license labels are missing or wrong.
• Dataset Cards (Hugging Face Hub documentation) — license metadata on dataset cards.
• Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items — openly licensed object assets with attribution terms.