The article exposes the tension between AI features that simply generate content and those that genuinely educate users. Adobe‘s experimental iOS camera app Project Indigo deliberately moves away from prompt-based editing, which Marc Levoy argues can be tricky, toward button-driven, deterministic outputs. The app now uses LLMs to deliver detailed critiques covering framing, lighting, colors, and emotional impact, plus tailored reshoot or editing suggestions. This contrasts with generic AI helpers like Google’s Pixel camera coach, which offered only broad framing advice, and aims to make users aware of photography nuances rather than just creating slop.

The concrete technical path includes two main critique features: a professional opinion on composition and mood, and capture/edit suggestions that guide users on reshoot adjustments or improve existing photos using Lightroom controls. Object removal shifts from manual circling to toggles for categories like people, trash, wires, poles, fences, vehicles, and custom objects, yielding clean results without artifacts. The app also adds AI-generated depth of field and style transfers (watercolor, pen and ink, monochromatic, backlit subject). Under the hood, these features run on Google’s Gemini Nano Banana, with Adobe open to swapping in its own Firefly model. They are bundled in an AI playground tab, still in testing and limited to select users.

For a serious builder, this signals a practical shift: AI can be made more educational and deterministic, reducing the chance of slop. The button-driven design over prompt-based editing suggests that constrained, reliable outputs often outperform open-ended generation for real-world photography help. The use of on-device models hints at performance benefits, but the experimental nature means these features may never ship widely. The takeaway is that thoughtful interface design, even with powerful LLMs, matters more than blind generation, and that teaching users about photography nuances can be a viable product direction.