OpenAI’s new image model is designed to keep the person or product intact while changing everything around it
OpenAI launched ChatGPT Images 2.5 on September 8. The model is designed to preserve subjects in reference photographs more faithfully while transforming settings, styles and compositions, and to make targeted edits without unnecessarily changing surrounding details. OpenAI says it also maintains better consistency through multiple editing turns. Two API versions are available: GPT-Image-2.5 Flare for faster high-volume generation and GPT-Image-2.5 Sunburst for more precise production-oriented editing. OpenAI continues to attach C2PA metadata and invisible watermarking to images produced with its tools.
Why visual creators should care
Better identity and product preservation makes AI much more useful for ordinary commercial photography workflows, but it also makes the original reference photograph more consequential. The model is increasingly capable of manufacturing a new scene while retaining the recognisable person, product or visual identity supplied by the creator.
Practical takeaway
Treat the reference photograph as a rights-bearing input. Retain the original, record who authorised its use and distinguish permission to photograph or license the source from permission to manufacture new AI derivatives from it.
What to watch next
Watch how OpenAI, Adobe Firefly, Runway and other platforms integrating Images 2.5 expose reference-image provenance, consent information and model identity when the finished asset leaves their systems.
Sources
Sony and Reuters are carrying camera provenance through the near-live newsroom
Sony and Reuters announced a near-live professional video workflow on September 8 that preserves C2PA provenance from capture through newsroom production. Sony’s PXW-Z300 camera records authenticity information when footage is shot. That information survives near-live file transfer into a TAMS-compatible store and is retrieved by Reuters alongside the video for editing and packaging. The workflow also uses watermarking so the original provenance information can potentially be recovered if ordinary metadata is removed.
Why visual creators should care
Content Credentials become much more useful when they survive real production. A provenance record that exists only in the camera original is easily lost once footage is transferred, clipped, edited and redistributed. This workflow is explicitly designed around that newsroom problem.
Practical takeaway
When assessing provenance technology, test the entire production chain rather than only capture. Confirm what happens during live transfer, transcoding, editing, clipping and distribution and whether authenticity information remains recoverable after ordinary metadata loss.
What to watch next
Sony and Reuters will demonstrate the workflow at IBC in Amsterdam from September 11 to 14. Watch for details on how provenance survives common newsroom edits and how clients can verify the final distributed footage.
Sources
Fraunhofer is putting provenance inside compressed live video rather than leaving it beside the file
Fraunhofer HHI will demonstrate at IBC 2026 an end-to-end live video-authentication workflow using Digitally Signed Content in H.265/HEVC. The system carries digital-signature information inside standardized supplemental enhancement messages in the compressed stream. A JPEG Trust manifest containing C2PA provenance and authenticity information can be embedded into that signed content, allowing verification data to remain linked with live video from encoding through later handling.
Why visual creators should care
Live video creates a difficult provenance problem because footage is continually encoded, segmented, transmitted and repackaged. Embedding authenticity information into the compressed-video workflow itself could make provenance substantially more resilient than relying on conventional file metadata alone.
Practical takeaway
Broadcasters and filmmakers evaluating provenance should ask whether authenticity survives the codec and live-production workflow. Sidecar metadata that disappears during normal video processing is much less useful than a verifiable record that travels with the stream.
What to watch next
Watch the September 11–14 IBC demonstration and Fraunhofer’s September 12 presentation on open-source C2PA live-streaming ingestion for evidence of interoperability with ordinary broadcast tools.
Sources
Creative Fabrica requires designers to tell buyers when the product itself was made with AI
Creative Fabrica’s current designer terms require any product generated wholly or partly with AI tools to be accurately categorised as AI-generated when it is uploaded. Designers remain responsible for holding the necessary rights to submitted material, and identifiable people, brands or private property require appropriate releases or permissions. Its separate Studio terms say users retain ownership of AI-generated assets while Creative Fabrica and underlying AI providers receive the limited licence needed to process and operate the generation service.
Why visual creators should care
This separates disclosure from ownership. A designer can retain rights in what they upload or create while still having marketplace obligations to label AI involvement and prove that recognisable people or protected material were properly cleared.
Practical takeaway
Keep the generator record and releases with the marketplace asset. An AI label does not replace model, property, trademark or copyright clearance, and retaining ownership does not mean the platform never passes material to an underlying AI provider.
What to watch next
Watch whether marketplaces begin requiring model names or provenance metadata in addition to a general AI-generated category.
Sources
Dreamstime accepts AI stock — but generated human faces create a release problem
Dreamstime’s current contributor guidance accepts AI-generated work, but contributors must have sufficient rights in the output and confirm that the AI tool they used gives them the necessary rights for stock licensing. Submissions must be placed in the Generative AI category and identified as AI-generated in the title or description. Dreamstime generally rejects generated human faces because contributors cannot provide a conventional model release for a fictional generated person, with limited exceptions for obviously stylised or non-identifiable figures.
Why visual creators should care
A convincing synthetic person creates a peculiar stock-licensing problem. There may be no actual model to sign a release, yet the generated face can look enough like photography that buyers may treat it as an image of a real person.
Practical takeaway
Before licensing generated people as stock, check whether the library accepts synthetic faces at all. Do not assume that a generator’s commercial-use permission satisfies a library’s separate model-release requirements.
What to watch next
Watch whether stock agencies develop a standard synthetic-person declaration or provenance record to replace the conventional model-release mechanism for fictional AI people.
Sources
Depositphotos treats the generated image as licensed stock — but the reference photograph still needs its own permission
Depositphotos’ current AI Image Generator allows generated output to be downloaded under its Standard licence for commercial use. The company says its generator is trained on commercially licensed material and that generated outputs remain private rather than being added to its public stock library. The tool also accepts reference images, allowing a user to upload a photograph and control how strongly the result follows it.
Why visual creators should care
The generated output may arrive with a platform licence, but that does not erase the rights attached to a reference photograph supplied by the user. Professional creators can therefore have a licensed final AI image built from a source file they were never authorised to transform.
Practical takeaway
When using reference generation, record the rights to the reference independently from the licence attached to the output. Client delivery permission, stock licensing and permission to create AI derivatives are not automatically the same thing.
What to watch next
Watch whether commercially licensed generators begin preserving the identity and licence status of reference images in provenance metadata or project records.
Sources
