When Adobe shipped Photoshop 24.6 in May 2023, it embedded a generative AI model directly into the toolbar millions of photographers already had open. The feature — Generative Fill — let users select any region of an image, type a text prompt, and have the surrounding pixels seamlessly replaced or extended within seconds. Within 72 hours of the public beta launch, Adobe reported users had generated more than 100 million images using the feature. It was the fastest adoption of any Photoshop feature in the software's 33-year history.
The underlying model, Adobe Firefly, had been trained exclusively on Adobe Stock images, openly licensed content, and public-domain works — a deliberate choice that Adobe made to sidestep the copyright controversies that had engulfed competitors like Stability AI and Midjourney. That training decision shaped the tool's outputs in measurable ways: Firefly's photorealistic results tend to be clean, commercially safe, and stylistically consistent with professional stock photography.
Generative Fill operates through a three-stage pipeline. First, Photoshop builds a context mask — a pixel-level map of what surrounds your selection, capturing color temperature, luminosity gradients, lens characteristics, and perspective cues. Second, that context map is passed to the Firefly diffusion model along with your text prompt (or an empty prompt for automatic in-painting). Third, the model outputs three candidate fills, each generated at the native resolution of your document, which Photoshop places on a non-destructive generative layer — meaning your original pixels are never touched.
The non-destructive layer architecture is critical for professional workflows. Every generative fill lives on its own layer with a mask, so photographers can blend opacity, swap between the three variants, or delete the fill entirely at any time. Adobe also stamps each generated file with Content Credentials metadata (based on the C2PA standard), creating a traceable chain of custody that records when and how AI was used in the image.
By late 2023, Adobe had expanded Firefly integration to include Generative Expand (extending the canvas outward), Remove Tool (one-click object removal), and Background Generation in Adobe Express. Each feature uses the same Firefly model but with different context-injection strategies tuned to the specific task.
Adobe's decision to train Firefly on licensed and public-domain content was not purely ethical — it was also strategic. When Getty Images filed a lawsuit against Stability AI in January 2023 for allegedly scraping 12 million images without consent, it signaled that enterprise clients would increasingly require provenance guarantees before deploying AI-edited images in commercial contexts. Adobe positioned Firefly as the commercially safe choice, and enterprise clients responded: by Q4 2023, Adobe reported Firefly had been used to generate over 3 billion assets.
The trade-off is that Firefly's stylistic range is narrower than models trained on the full breadth of the internet. It excels at photorealistic fill — replacing a sky, removing a distraction, extending a background — but it is less adept at highly stylized or abstract prompts compared to Midjourney or DALL-E 3. For photographers working in commercial contexts, this constraint is often irrelevant; for fine-art practitioners, it may be a meaningful limitation.
Generative Fill does not replace retouching skill — it accelerates it. The tool still requires photographers to make precise selections, write clear prompts, and evaluate whether the AI output matches the lighting and perspective of the original scene. Poorly made selections produce seams; vague prompts produce inconsistent results. The photographer's eye remains the quality filter.
Adobe's 3 billion Firefly-generated assets by end of 2023 represents a speed of adoption unprecedented in creative software history. For context, it took Instagram roughly 14 months to reach its first billion photos shared.
You are a commercial photographer who has just shot a product photo for a client. The client wants three variations: one with a neutral studio background, one with an outdoor lifestyle setting, and one with the original background removed entirely. You'll use this chat to think through how to write effective Generative Fill prompts, which Photoshop selection tools to use, and how to evaluate whether AI output is commercially usable.
Have at least three exchanges with the AI assistant — ask about prompt writing, selection strategies, or how to assess output quality. The more specific your questions, the more useful the responses.
In October 2021, Adobe shipped Lightroom 11 with a feature called AI Masking. For the first time, photographers could click "Select Sky" or "Select Subject" and watch Lightroom build a pixel-perfect mask of the sky or the human figure in their photograph — in under two seconds. Before this, the same task required manually painting masks in Photoshop for anywhere from five minutes to over an hour depending on complexity. In a single update, Adobe collapsed that time gap to near-zero for a huge proportion of real-world editing scenarios.
The underlying model was trained on tens of millions of labeled images to recognize semantic regions — sky, foreground, people, backgrounds — rather than just detecting edges or color differences. This distinction matters: the AI can correctly mask a person standing in front of a wall that shares their skin tone, something that had previously defeated even experienced retouchers using traditional selection tools.
Lightroom's subject and sky masks use a semantic segmentation model — a neural network that assigns a class label (sky, person, animal, foreground object) to every pixel in the image. This is fundamentally different from older selection methods like Photoshop's Magic Wand (which selects contiguous similar colors) or Select Color Range (which selects all pixels of a specific hue). Semantic segmentation understands context: it knows that the pixels at the top of an image inside an arch are probably "sky" even if they share a similar blue with the subject's shirt.
The practical result for photographers is dramatic. Portrait retouchers can now apply targeted adjustments — a +0.5 EV exposure boost, a skin-tone dehaze, a background blur — that affect only the subject, with a mask generated in two seconds rather than fifteen minutes. Landscape photographers can independently adjust sky exposure, saturation, and haze without affecting the foreground, enabling looks that previously required Photoshop compositing.
Adobe extended the AI masking system further in 2022 with People Masks, which can separately select the face skin, body skin, hair, clothing, and eyebrows of each person in a group photo. This granularity enables per-person retouching within Lightroom's non-destructive raw workflow — a capability that had previously required exporting to Photoshop and applying layer-based adjustments manually.
In March 2023, Adobe added Denoise to Lightroom — an AI-powered noise reduction tool that represented a step-change from the luminance and color noise sliders that had existed since Lightroom 1.0 in 2007. The new Denoise feature uses a deep neural network trained specifically on the noise patterns produced by different camera sensors at different ISO settings, and it produces a new DNG file with noise removed while preserving fine detail at a level that traditional luminance reduction cannot achieve.
The processing is computationally intensive — on a 45-megapixel file, Denoise can take 30 to 90 seconds to generate its output DNG, depending on GPU availability. But the results in high-ISO photography are striking: images shot at ISO 12,800 or ISO 25,600 that would previously have been considered unusable emerge with fine texture, retained color fidelity, and none of the smearing artifacts that traditional noise reduction produces. Wedding photographers, sports photographers, and photojournalists — all of whom routinely shoot in available light at high ISO — have reported it as one of the most practically significant Lightroom updates in years.
The technical reason for the quality gain is that the AI model has been trained to distinguish between noise patterns (random, sensor-generated, statistically predictable) and actual image detail (edges, textures, fine structure). Traditional sliders apply uniform blurring that cannot make this distinction, which is why they sacrifice detail. The AI can suppress one while preserving the other.
Lightroom's AI tools operate within its non-destructive raw processing pipeline — ideal for batch adjustments across hundreds of images. Photoshop's Generative Fill is better for complex compositing, content replacement, and tasks requiring precise pixel-level control. Many professional workflows now use both: Lightroom for culling, AI masking, and Denoise, then Photoshop for any generative or compositing work.
Before AI Denoise, concert photographers shooting at ISO 12,800 routinely discarded 40–60% of images as "too noisy to deliver." Multiple photographers on professional forums (including Luminous Landscape and Photography Life) reported that AI Denoise in Lightroom 2023 allowed them to recover and deliver images they would previously have deleted — effectively increasing usable yield from difficult shoots.
You are editing a set of 80 wedding reception photos — mix of dance floor shots at ISO 6400–12800, portrait group shots, and outdoor ceremony images. Your client wants consistent edits delivered in 48 hours. Work through with the AI assistant how to use Lightroom's AI Masking and Denoise to handle this job efficiently — including which mask types to use for which shots, when to apply Denoise, and how to manage the processing time on your specific machine.
Have at least three exchanges. Try asking about specific scenarios (e.g., a dance floor shot with motion blur AND noise, or a group portrait where two people have similar skin tones to the background wall).
While Adobe integrated AI into its established products, a parallel ecosystem of specialized AI photography tools emerged from independent developers. Skylum's Luminar Neo, Topaz Labs' suite (DeNoise AI, Sharpen AI, Gigapixel AI), and DxO PhotoLab's DeepPRIME each carved out specific niches — offering capabilities that either preceded Adobe's equivalent features by years or went deeper into their specific domain than Adobe chose to.
Topaz Labs released DeNoise AI in 2019 — four years before Adobe's AI Denoise — and it developed a devoted following among wildlife and sports photographers who regularly pushed cameras to ISO 25,600 or beyond. When Adobe eventually released its competing tool in 2023, professional reviews consistently found the two broadly comparable, with advantages depending on camera model and shooting conditions. The competitive pressure from Topaz had clearly accelerated Adobe's timeline.
Topaz DeNoise AI uses a model trained specifically on high-ISO noise patterns from a library of camera bodies, with separate noise models for different sensor types. Its flagship mode, "Low Light," is tuned for extremely high ISOs (12,800–102,400) and uses what Topaz calls a "subject isolation" pass to ensure faces and fine textures are denoised with extra care. As a standalone application and Photoshop/Lightroom plug-in, it processes one image at a time with full GPU acceleration.
Topaz Sharpen AI addresses a different problem: images that are slightly out of focus, motion-blurred, or camera-shake-blurred. It distinguishes between three types of unsharpness — out-of-focus blur, motion blur, and softness — and applies different correction algorithms to each. This is a capability that no in-camera or traditional software sharpening can replicate; Topaz's model reconstructs detail by learning what "sharp" looks like for thousands of subject types.
Topaz Gigapixel AI performs AI-driven upscaling — enlarging images to 2×, 4×, or 6× their original resolution by inferring detail that was never captured. It was used commercially to upscale archival images for large-format print, and Associated Press archivists have reported using it to prepare historical photographs for modern high-resolution reproduction. The tool outputs images that pass visual inspection at print sizes where the original would show obvious pixelation.
Skylum's Luminar Neo (released 2022) took a different strategic direction — positioning AI as a creative enhancement platform rather than a technical correction tool. Its Sky AI feature detects the sky in any landscape or exterior photo and replaces it with one from a library of dramatic alternatives, automatically adjusting the foreground lighting to match the new sky's color temperature and direction. This is technically a compositing task that Photoshop users had performed manually for decades; Luminar Neo automated the entire process including lighting relighting.
Its Relight AI feature uses a depth-estimation model to analyze the three-dimensional structure of a scene and then applies simulated lighting that respects depth — brightening foreground subjects while leaving distant backgrounds darker, or vice versa. This simulates the behavior of physical lighting in a way that flat dodging-and-burning cannot. Portrait Bokeh AI uses subject segmentation to add or increase background blur with accurate depth falloff, effectively adding the shallow depth-of-field look of a fast prime lens to images shot with kit zooms.
Critics of Luminar Neo note that many of its AI effects lean heavily toward dramatic or "Instagram-optimized" aesthetics that can look artificial. Photojournalism and documentary applications are largely incompatible with most of its AI tools — Sky AI replacement, for instance, would be an ethical violation in editorial contexts. Luminar Neo is primarily a tool for commercial, portrait, and social media photography.
DxO PhotoLab's DeepPRIME and DeepPRIME XD algorithms represent a different architectural choice: rather than working as a post-processing plug-in, DeepPRIME is integrated directly into DxO's raw demosaicing pipeline. This means the AI noise reduction is applied at the earliest possible stage of processing — before the raw data is converted to RGB — allowing it to work on luminance patterns in the native sensor data rather than the already-processed image. In blind tests conducted by DPReview and Imaging Resource in 2022, DeepPRIME XD was rated the best noise reduction for high-ISO raw files among all tested tools, narrowly ahead of Topaz DeNoise AI.
For photographers already in the Adobe ecosystem, Lightroom's AI Denoise and masking cover 80–90% of use cases without leaving the application. Topaz tools add value when pushing cameras to their ISO limits or rescuing technically imperfect images. Luminar Neo serves photographers who prioritize creative transformation over technical accuracy. DxO PhotoLab is the specialist choice for raw file quality at any ISO.
Topaz DeNoise AI launched in 2019 — four years before Adobe's equivalent. The competitive pressure from independent AI tool developers has consistently accelerated Adobe's product roadmap, benefiting all photographers regardless of which ecosystem they use.
You are a freelance photographer with access to Adobe Lightroom, Topaz DeNoise AI, Topaz Gigapixel AI, Luminar Neo, and DxO PhotoLab. A client has sent you four jobs: (1) a wildlife series shot at ISO 25,600 on a Nikon Z9, (2) a landscape portfolio for a gallery print run at 40×60 inches from 24MP files, (3) an interior architecture shoot where skies through windows look blown out and dull, and (4) a batch of 300 portrait headshots for a corporate LinkedIn campaign.
Work through with the AI assistant which tools to use for each job, why, and in what order. Have at least three exchanges — push for specific reasoning, not just tool names.
In April 2024, the World Press Photo Foundation disqualified three winning photographs — including the Photo of the Year — after post-submission technical analysis found that the images had been digitally manipulated beyond the organization's permissible standards. The winning photograph, by Palestinian photographer Mahmoud Ajjour, was disqualified after forensic examination suggested that elements of the image had been substantially altered. The disqualification triggered a global discussion about what constitutes "acceptable" editing in an era when AI tools can seamlessly reconstruct, extend, and replace image elements.
The World Press Photo controversy was not an isolated incident. In 2023, Reuters photographer Yannis Behrakis had already warned that AI editing tools represented "the most serious threat to photojournalism's credibility since the invention of Photoshop." The Associated Press, Reuters, and AFP each updated their photo editing guidelines in 2023–2024 specifically to address AI-generated and AI-assisted modifications, distinguishing between permissible technical corrections and prohibited content alterations.
Major wire services and news organizations converged in 2023–2024 on a common framework distinguishing technical corrections (permissible) from content alterations (prohibited in editorial contexts). The Associated Press guidelines explicitly prohibit: removing or adding any element to the frame, altering colors to change meaning, using generative AI to modify any part of an image. They permit: tonal adjustments (exposure, contrast, color balance), cropping, and standard noise reduction — as long as these do not obscure or alter the factual content of the scene.
The distinction matters enormously in practice. Using Lightroom's AI Denoise on a high-ISO news photograph is permissible — it makes existing information more visible. Using Generative Fill to remove a distracting element from a news photograph is prohibited — it alters the factual record of what was present. The technology operates on a continuum, and the ethical line runs through the question of whether information is being enhanced or fabricated.
The National Press Photographers Association (NPPA) Code of Ethics states photographers must "not manipulate images or add or alter sound in any way that can mislead viewers or misrepresent subjects." The NPPA's guidance on AI tools (updated 2024) recommends full disclosure when any AI-assisted editing has been applied to editorial images, even for permissible technical corrections.
The strict standards of photojournalism do not apply uniformly across all photography contexts. In commercial advertising photography, extensive retouching — including AI background replacement, generative object removal, and composite building — has been standard practice for decades, and AI tools simply make these processes faster. The ethical considerations shift to transparency in advertising (the UK's Advertising Standards Authority, for example, has ruled that digitally altered body images in advertising must be labeled) and to copyright in generative outputs.
In portrait photography, AI skin retouching tools like PortraitPro and Lightroom's AI skin masking have raised questions about unrealistic beauty standards and the psychological impact of routinely delivering images of people that remove their natural skin texture, asymmetries, and age marks. A 2022 survey by the British Journal of Photography found that 67% of portrait photographers used some form of automated AI skin retouching, and 41% said clients now explicitly requested "AI-smoothed" results — a shift in expectation driven by social media AI filters.
In fine art photography, the question is not authenticity in the journalistic sense but authorship and intent. Photographers like Andreas Gursky have been manually compositing and retouching prints for decades — his 1999 photograph "Rhine II" (which sold for $4.3 million in 2011) had entire elements removed in post-processing. The fine art context treats the final image as an expression of vision rather than a document of reality, making AI editing tools a legitimate extension of the photographer's toolkit.
The C2PA Content Credentials standard (discussed in Lesson 1) represents the technology sector's proposed solution to the AI editing transparency problem. By embedding cryptographically signed records of editing history into image files, Content Credentials allow any viewer — or any publication — to verify what was done to an image and by which tools. Adobe, Canon, Nikon, Sony, Leica, and Microsoft are all members of the C2PA coalition, and Leica's M11-P (released 2023) became the first camera to capture Content Credentials at the point of shooting, signing the original capture before any editing occurs.
The practical adoption of Content Credentials remains uneven. Platforms must choose to read and display the metadata, which most currently do not. But the infrastructure is being built, and the expectation within the industry is that within three to five years, editorial image submissions will routinely be checked against Content Credentials before publication — making AI-assisted manipulation detectable even when it is visually indistinguishable from the original.
Before applying any AI editing tool to an image you intend to publish or submit professionally, ask: Does this edit make existing information more visible — or does it add, remove, or change information about what was present in the scene? The first is enhancement. The second is fabrication. The distinction defines the ethical boundary across every professional photography context.
Released in October 2023, the Leica M11-P is the first camera to incorporate C2PA Content Credentials signing into its firmware. Every image captured by the camera is automatically signed with a cryptographic certificate, establishing an immutable record of the original capture before any editing. At $9,195 USD, it is not a mainstream tool — but it established a precedent that other manufacturers are expected to follow.
You are a photo editor at a mid-size publication that covers news, commercial clients, and maintains a portrait studio. In the past week, three situations have landed on your desk: (1) A staff photographer used Lightroom's AI Denoise on breaking news images from a protest — a colleague objects that this constitutes prohibited manipulation. (2) A commercial client is asking your studio to use Generative Fill to remove an unflattering shadow from their CEO's headshot before the annual report. (3) A freelancer submitted a stunning landscape that appears to have had its sky replaced using Luminar Neo — you can't be certain, there are no Content Credentials, and the photographer denies it.
Work through each scenario with the AI assistant. What decisions do you make, what standards apply, and what processes would you put in place? Have at least three exchanges.