When producing an AI video, someone may start by finding a reference photo: not a celebrity, just a blogger found online. That choice can still bring rights risk into the project. Whether the person is famous is not the deciding factor. If the individual is recognizable and the image is used without permission for the intended project, it should not be used to generate commercial content.
The less obvious version is blending live-action footage into a generated image, or extracting only a face, pose, clothing pattern or distinctive feature and assuming that a new background or style makes the source irrelevant. Recognizable elements may remain. If the person objects or brings a claim, the consequences can extend beyond removing one image to suspending distribution, advertising and an entire video rework.
First line of defense: reject unauthorized likenesses at the source
The client and the production team should identify where the person material came from, who can grant permission and which uses the permission covers. Commercial advertising, corporate communications, social posting, derivative edits and distribution in different territories may require different scope. Labels such as public, downloadable or free to use do not by themselves grant portrait or commercial-use rights.
When a fictional person is needed, use an original character, a clearly licensed real-person reference or a newly designed person that does not depend on a specific individual's recognizable traits. For face swaps, likeness compositing or close imitation, document the authorized party, permitted uses, term and channels before production. One extra approval step early is cheaper than replacing a character and rebuilding storyboards midway through the project.
Second line of defense: check the model and LoRA separately
Open source describes how something is made available; it is not a blanket commercial pass. Models, LoRAs, workflows and asset packs can carry different licenses. Some restrict use to research or personal work, while others add conditions for derivatives, redistribution or advertising. Even when a model provider permits use of its outputs, that does not mean it guarantees that every output is free of portrait, copyright, trademark or other third-party claims.
Before production, record the exact model and version, the source of every LoRA or plugin, the account, generation date and the license page or service terms effective at that time. Do not rely on a chat message saying commercial use is supported. Before publication, review the reference photos, character designs, music, fonts and other assets separately against the actual intended use.
Third line of defense: turn the workflow into a rights file
A final video alone may not show how it was created if a dispute arises. Save the brief, character design, script, storyboard, source and permissions for references, positive and negative prompts, parameters, seed values, model version, generation times and raw outputs by version. For selected shots, keep the reason for selection, manual edits and before-and-after files.
Keep project files, export versions, approvals and delivery files for compositing, repainting, color work, editing, voiceover, music, captions and effects. Use consistent names and read-only archives, then fix time and file integrity with hashes, trusted timestamps or a compliant electronic evidence service. Evidence preservation can help establish the workflow and show that files were not altered later; it cannot make an unauthorized likeness lawful.
In the Chinese mainland context, a natural person's voice is also treated under rules that refer to portrait-right protection. Face replacement and identifiable voice cloning should not bypass permission. The value of a rights file is practical: it shows who supplied each asset, which permission supported its use, and what a person selected or modified, instead of leaving the team to reconstruct the story from memory.
Avoiding portrait infringement in AI-generated video therefore starts before final publication. Reject the idea of downloading any person's photo from the internet, verify the licenses for models, LoRAs and third-party assets, and preserve the creation record from start to finish. Companies without this production experience can explore AI video production services and ask the team to build permissions and rights records into the script, asset, generation and delivery stages.
