According to this weekend’s Power On newsletter, Apple is preparing a major upgrade to its image generation models used in Genmoji and Image Playground. These models are set to receive a ‘big boost’ in visual quality when iOS 27 launches later this year. Currently, the output from these tools has been widely criticized for lacking the fidelity and detail seen in competing image generators from companies like OpenAI and Google.

Genmoji, which allows users to create custom emoji-style images based on text prompts, and Image Playground, a more general image creation app, debuted in iOS 18.2. While the concept was innovative, the execution left many users underwhelmed. Image Playground, in particular, was noted for generating blurry, low-resolution results that often failed to match the described subject matter. Genmoji fared slightly better but still lagged behind the realism of standard emoji or third-party alternatives.

Background on Apple’s image generation journey

Apple first introduced on-device AI image generation with iOS 18.2, positioning it as a privacy-focused alternative to cloud-based services. The company emphasized that all processing happens locally on the device, using Apple’s own neural engine and small language models. This approach ensures that user data never leaves the iPhone, a key selling point for privacy-conscious consumers. However, running complex image generation algorithms on-device imposes strict limits on model size and computational resources. As a result, the quality of generated images has been noticeably inferior to server-side solutions like OpenAI’s DALL-E or Google’s Imagen.

Over the past year, Apple has quietly been refining its image generation pipeline. The upcoming iOS 27 update appears to be the culmination of those efforts. According to sources familiar with the matter, the new models leverage advanced training techniques and possibly more efficient architectures that allow for higher resolution and more coherent outputs without increasing the processing burden. This could mean that even older devices will see noticeable improvements, as the algorithms become more efficient at extracting meaningful features from prompts.

What the upgrade entails

The newsletter specifically mentions a “big boost” in visual quality for both Genmoji and Image Playground. While exact technical details remain under wraps, it is believed that Apple has incorporated a larger dataset of high-quality images and improved the conditioning mechanisms that map text descriptions to visual elements. Additionally, the new models may adopt a diffusion-based approach similar to state-of-the-art systems, but optimized for Apple’s Neural Engine.

Another significant change is the expansion of third-party model support. Currently, Image Playground only integrates OpenAI’s ChatGPT for image generation. With iOS 27, Apple plans to open the platform to other providers. Industry speculation suggests that Google’s Nano Banana models — a lightweight variant of their larger image generation system — could be among the first to be supported. This would give users the flexibility to choose different models for different tasks, much like how they can select different search engines in Safari.

The move to support third-party models also hints at a broader strategy: Apple may be creating an ecosystem where developers can offer their own image generation models through the App Store, similar to how keyboard extensions work today. Such a marketplace would allow niche creators to provide specialized styles or capabilities, from photorealistic rendering to cartoon-like illustrations.

Comparative analysis with competitors

To understand the scale of this improvement, it helps to compare Apple’s current offerings with what is available elsewhere. OpenAI’s DALL-E 3, for example, can generate highly detailed images with accurate compositions and subtle lighting effects. Google’s Imagen Video even supports short video clips. By contrast, Image Playground often produces images that appear soft, with inconsistent anatomy and unnatural textures. Genmoji struggles with complex concepts involving multiple characters or actions.

In the enterprise space, tools like Adobe Firefly leverage generative AI to assist designers, offering professional-grade output that can be directly used in marketing materials. Apple’s tools, while fun for casual users, have not reached that level of utility. The iOS 27 upgrade aims to close this gap, potentially making Genmoji and Image Playground viable for tasks like social media content creation, educational materials, and even light graphic design work.

Privacy remains a differentiating factor. Even with improved quality, Apple’s on-device processing ensures that no images or prompts are uploaded to remote servers. For users concerned about their creative data being used to train public models, this is a crucial advantage. The trade-off, however, is that on-device models may never match the absolute peak quality of cloud-based counterparts due to hardware constraints. But with the efficiency gains hinted at in the report, the gap may become negligible for many practical applications.

Technical challenges and future possibilities

Running large image generation models on a mobile device is no small feat. The iPhone’s Neural Engine is powerful, but memory and power consumption are limiting factors. Apple has likely employed model quantization, pruning, and distillation techniques to reduce the model size while preserving quality. The “big boost” could also come from leveraging the A18 or A19 chips’ improved neural cores, which might include specialized hardware for attention mechanisms used in diffusion models.

Another possibility is that Apple is moving toward a hybrid approach, where some processing is done on-device and some in a secure enclave or Trusted Execution Environment. However, given Apple’s strong stance on privacy, any offloading would have to be encrypted and ephemeral. The report doesn’t clarify whether the new models remain purely on-device, but it is a critical question for users who prioritize data security.

Beyond image quality, Apple may also introduce new features such as style transfer, text-to-background generation, and intelligent composition suggestions based on user’s photo library. Earlier leaks have indicated that Genmoji will gain the ability to suggest characters based on faces in the user’s contacts or photos. This personalization layer would make the tool more useful and engaging, differentiating it from generic third-party generators.

The integration with third-party models also raises intriguing possibilities. For example, a photography enthusiast could use a model specialized in simulating vintage film stock, while a graphic designer might prefer a model trained on corporate branding guidelines. Apple’s platform approach could turn Image Playground into a versatile creative tool rather than just a novelty app.

In the broader context, Apple’s investment in on-device AI is part of a larger push to make the iPhone an intelligent personal assistant. Image generation is just one piece; we can expect similar improvements in text creation, music composition, and even video editing in future iOS updates. The company is clearly committed to enabling users to express their creativity without compromising privacy or relying on external services.

As iOS 27 development progresses, developers and early adopters will have the chance to test these enhancements in beta builds. If the reported quality boost materializes, Apple’s image generation tools could become a standout feature of the operating system, rivaling dedicated apps and services. For now, the wait continues, but the promise of significantly better Genmoji and Image Playground outputs is encouraging news for anyone who has been disappointed by the initial implementations.

Source: 9to5Mac News