The iOS 27 developer beta, released by Apple ahead of the expected iPhone 18 launch in September, introduces two Apple Intelligence image-altering tools in Photos: Spatial Reframing and Extend. According to TechRadar reviewer Lance Ulanoff, the tools rely on "a powerful private-compute cloud-based diffusion model built, in part, with Google" — a departure from previous on-device-only AI features. Ulanoff described the new capability as "unlike any AI Apple has ever presented before."
Private Cloud Diffusion Model: Architecture and Performance
Spatial Reframing lets users turn photo subjects to see elements the camera never captured. Extend fills in cropped-out areas, guessing at missing content. Apple uses a private-compute cloud-based diffusion model to generatively alter images, a technology distinct from earlier on-device approaches. Ulanoff noted that Apple "is taking the AI image-altering strategy" further with these tools.
In testing, Ulanoff took tight and wide shots of the same scenes to compare Extend’s output against reality. He found that "Photo Extend did a pretty good job" at reconstructing landscapes and orderly scenes. However, the tool will not extend body parts — Ulanoff tried cropping out fingers and the app refused, which he praised as preventing six-finger artifacts. Extend also "likes order" and "seems almost allergic to clutter," generating clean, minimal additions.
| Feature | Description | Performance Notes |
|---|---|---|
| Landscape extension | Fills in missing background and scenery | Generally accurate; closely matches original wide shots |
| Body part extension | Attempts to regenerate missing limbs or fingers | Will not extend; avoids generating unrealistic anatomy |
| Clutter handling | Adds content only in clean, orderly patterns | Avoids introducing complex elements; prefers minimal additions |
Limitations and Quality Control
Ulanoff reported that the tools are still in developer beta and "will likely change quite a bit before they arrive fully baked." He emphasized Apple’s current approach: "let guesswork be your guide." The AI does not know what was actually cropped out but uses intelligence to read the image and fill blanks. Some results were "unintentionally comical," though if viewers had not seen original photos, they "might never know that significant portions of the images were generated by AI."
These limitations highlight the importance of validation in generative AI — a lesson applicable to enterprise deployments where accuracy and trust are critical.
Broader Implications for Enterprise AI Deployments
While the immediate use case is consumer photography, Apple’s choice of a private-compute cloud model — rather than sending data to third-party servers — addresses enterprise concerns around data sovereignty and privacy. The partnership with Google for the diffusion model also shows how large language models can be composed with infrastructure providers. For technology leaders evaluating AI for logistics, trade documentation, or customs systems, the iOS 27 approach suggests that private cloud inference can deliver advanced capabilities without exposing sensitive corporate data. As Ulanoff noted, Apple "opens itself up to a lot of questions about whether it still prizes image truth over aesthetics," a tension that will resonate in any industry relying on AI-generated outputs.
Sources:
