Adobe's Camera Lab Tests AI That Teaches Photography, Not Just Edits

A Different Approach to AI-Assisted Capture

Adobe has begun testing a suite of AI features inside Project Indigo, the company's experimental iOS camera application that first appeared in 2025. The new capabilities center on critique rather than creation: the app now evaluates framing, lighting, color balance, and emotional resonance, then suggests concrete changes a photographer might make before or after the shutter clicks. At DailyTechWire, we've tracked the steady migration of large language models into mobile imaging workflows, and this implementation stands out because it prioritizes education over automation.

Marc Levoy, who previously architected computational photography systems for Google's Pixel line, leads the Indigo effort. His rationale is straightforward: prompt-based generative tools demand trial and error that mobile shooters rarely have patience for. By replacing open-ended text fields with labeled buttons and targeted feedback, Adobe aims to deliver predictable results while teaching users why a given adjustment improves an image.

The critique module evaluates four dimensions. It comments on compositional balance, the quality and direction of light, the palette's harmony or tension, and the photograph's ability to convey mood. These assessments arrive as natural-language paragraphs rather than numeric scores, a design choice that mirrors the way a mentor might review a portfolio. The second feature, capture and edit suggestions, activates inside the viewfinder. It may recommend repositioning a subject to exploit leading lines, adjusting exposure compensation before capture, or removing distracting elements that fragment attention. In one test scenario, the system flagged a hexagonal white object in the lower third of the frame and advised its removal before the shot was finalized.

A parallel pane surfaces post-capture edits mapped to Adobe Lightroom's control schema: curves, selective color, clarity, and similar sliders. This dual-mode feedback loop, pre-capture plus post-process, reflects an understanding that mobile photography is rarely a single-moment discipline; decisions cascade from composition through to final export.

Object Removal Without Manual Masks

Generative fill and content-aware tools have existed in desktop and mobile photo editors for years, but they typically require the user to paint a mask or trace a boundary. Project Indigo introduces category toggles: background figures, trash receptacles, utility poles and wires, fencing, vehicles, and miscellaneous clutter. A custom-object field accepts short text descriptions for edge cases. The system then isolates and removes the specified elements without manual selection.

In practice, the feature handled complex occlusions cleanly. A test image containing a person holding an object in the mid-ground was processed successfully; both the figure and the held item disappeared, and the algorithm reconstructed plausible texture behind them. No visible seams or repetitive patterns emerged in the fill region, a common failure mode for earlier inpainting models. This level of reliability matters in mobile contexts, where users expect single-tap solutions and lack the time to refine masks across multiple iterations.

Apple Photos, Google Photos, and Adobe's own Photoshop have offered similar capabilities, but the toggle-based interface lowers the cognitive load. Instead of deciding what to select, the user decides which category of distraction to suppress. It is a subtle but meaningful shift in interaction design, one that aligns with the broader industry move toward intent-based controls rather than parameter-heavy panels.

Depth Simulation and Style Transfer

Project Indigo also includes a depth-of-field generator that synthesizes background blur from a single frame. Computational bokeh is not novel; Apple and Google have shipped portrait modes for years, and Samsung's camera software performs similar transformations. What differentiates this implementation is its integration with the critique and suggestion workflow. The app may recommend enabling depth simulation to isolate a subject, then evaluate whether the resulting separation enhances emotional impact.

Style transfer options round out the feature set: watercolor, pen-and-ink, ink line with color wash, monochromatic rendering, and backlit-subject effects. These filters echo the aesthetic experiments popularized by Prisma nearly a decade ago. While the current generation of diffusion models produces more refined output, style transfer remains a niche use case. Most professional and enthusiast photographers prioritize fidelity over stylization, and the feature feels more like a legacy checkbox than a core pillar of the product.

Despite Levoy's stated skepticism toward prompt-driven interfaces, Project Indigo does include a free-form edit field. Users can describe transformations that fall outside the preset button array, invoking editing sequences not otherwise exposed in the UI. This creates a tension: the app's design philosophy champions determinism, yet it still offers an escape hatch into the unpredictability it seeks to avoid. The risk is that users who rely on the prompt field will encounter the same frustrations that motivated the button-based approach in the first place.

Model Strategy and Rollout Timeline

Adobe is using a variant of Google's on-device language model, referred to internally as Nano Banana, to power the critique and suggestion features. The choice of an edge model reflects both latency requirements and privacy considerations; processing image metadata and generating feedback locally avoids round-trip server calls and keeps user photos off external infrastructure. However, Adobe has signaled flexibility: if performance or capability gaps emerge, the company may substitute its own Firefly foundation model or another third-party stack.

The features currently reside under an "AI Playground" tab and remain gated to a limited test cohort. Adobe has not committed to a public release timeline, and some capabilities may never graduate beyond the experimental phase. This cautious rollout mirrors the company's broader approach to generative AI, which has emphasized opt-in workflows and clear labeling to mitigate backlash over training data and output authenticity.

Implications for Mobile Imaging

Project Indigo's design choices illuminate a fork in the road for computational photography. One path leads toward full automation: the camera decides composition, exposure, and post-processing with minimal user input, optimizing for shareability and algorithmic engagement. The other path treats the camera as a teaching instrument, surfacing the reasoning behind each adjustment and encouraging users to internalize principles they can apply manually.

The critique and suggestion features lean toward the latter. By explaining why a reframing improves balance or why removing a background element strengthens focus, the app functions as a portable workshop rather than a black-box filter. This aligns with Adobe's historical positioning as a toolmaker for creative professionals, even as consumer expectations push toward one-tap solutions.

Yet the educational value depends entirely on the quality of the critique. If the feedback is generic or misaligned with the photographer's intent, users will ignore it. If it is overly prescriptive, it risks flattening artistic diversity into a narrow aesthetic. The challenge for Adobe is calibrating the model to offer useful guidance without imposing a singular photographic doctrine.

The object-removal toggles represent a different trade-off. They prioritize convenience and speed, abstracting away the manual labor of mask refinement. In doing so, they also abstract away control. A toggle for "people in the background" cannot distinguish between a stranger who disrupts the composition and a friend who adds context. The user must trust the algorithm's judgment about what constitutes clutter, a trust that will vary widely depending on shooting context and personal style.

What Comes Next

At DailyTechWire, we've observed that mobile camera software is increasingly a battleground for on-device AI differentiation. Apple's computational photography pipeline, Google's real-time HDR and portrait light adjustments, and Samsung's scene optimization all rely on specialized silicon and tightly integrated models. Adobe enters this arena without its own hardware, which means Project Indigo must deliver value through software alone, and it must do so in a way that justifies launching a separate camera app rather than enhancing Lightroom Mobile.

The critique and suggestion features offer a plausible answer: they provide a layer of instruction that platform-native camera apps do not. Whether that instruction resonates with users, and whether it translates into sustained engagement, will determine if these experiments migrate into Adobe's mainline products. For now, Project Indigo remains a signal of intent, a demonstration that AI in photography can serve purposes beyond generating slop or automating away creative decisions.

The industry will be watching to see if deterministic, button-driven AI can coexist with the open-ended, prompt-heavy paradigm that dominates generative tools today. If Adobe succeeds in making the camera a better teacher, it may chart a path that other mobile imaging platforms choose to follow.