Photography
Last reviewed
May 13, 2026
Sources
40 citations
Review status
Source-backed
Revision
v2 · 5,351 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 13, 2026
Sources
40 citations
Review status
Source-backed
Revision
v2 · 5,351 words
Add missing citations, update stale details, or suggest a clearer explanation.
See also: Photography ChatGPT Plugins
Artificial intelligence in photography covers a wide span of techniques, from the computational pipelines baked into modern smartphones to the generative editing tools now built into Photoshop and Lightroom, the AI culling apps used by wedding photographers, and the cryptographic provenance systems being added to flagship cameras by Leica, Sony, Nikon, and Canon. The field has changed faster between 2022 and 2025 than at any point since the move from film to digital, and the changes have triggered a parallel argument among photographers, news organisations, and contest juries about what still counts as a photograph.
AI now sits inside almost every consumer image, often without the user knowing. A handheld photo of the moon from a 2023 Samsung phone may have been silently re-textured by a neural network trained on moon images. A portrait taken on a Pixel 8 may have swapped a different frame's smile into the final shot through Best Take. A landscape exported from Lightroom may have had a tourist erased by Adobe Firefly. Stock libraries from Getty, Adobe, and Shutterstock now sell purely generated images alongside real photographs. At the same time, news agencies are scrambling to keep the chain of trust intact: the Associated Press, Reuters, and AFP have all started testing in-camera signing systems based on the C2PA standard.
The phrase "computational photography" predates the current wave of AI by more than a decade. It was first used by Steve Mann in 1995 and given its modern definition by Marc Levoy at Stanford in 2004 and 2005. Levoy ran a Stanford course on the subject in even-numbered years from 2004 onward, and the field grew up around the idea that a digital photograph is no longer a single exposure but a software-mediated reconstruction from multiple captures.
The current wave is different in two ways. First, the underlying models are deep neural networks rather than hand-tuned algorithms, so the results depend on training data and not just optics. Second, generative diffusion models can now invent image content from scratch, so the line between editing a photograph and synthesising a new image has become thin. Tools like Photoshop's Generative Fill (May 2023), Lightroom's Generative Remove (May 2024), Pixel's Magic Editor (October 2023), and Apple's Clean Up (October 2024) all sit on that boundary.
The authenticity response, the Content Authenticity Initiative (CAI) and its technical arm the Coalition for Content Provenance and Authenticity (C2PA), has been led mostly by Adobe with support from Microsoft, the BBC, Intel, Arm, Truepic, and a growing list of camera makers. The Leica M11-P, announced in October 2023, was the first production camera to ship with C2PA Content Credentials enabled at the point of capture.
The first wave of computational photography arrived inside Google's Nexus and Pixel phones. Google's HDR+ launched in 2014 and shipped widely on the Nexus 6 and later Pixel devices. It worked by capturing a burst of underexposed frames, aligning them in software, and merging them to reduce noise and recover highlight detail. The pipeline was hand engineered by Marc Levoy's Gcam team at Google Research, and it produced JPEGs that often outperformed cameras with much larger sensors.
Night Sight followed in November 2018, built on top of HDR+, and let Pixel phones take handheld photos in light so dim that traditional camera autofocus and metering would fail. The technique extended the exposure time per frame and increased the number of frames in the burst, then used machine learning to set the white balance correctly when the colour of the light was unknown. Levoy retired from Stanford in 2014 to join Google full time, led the imaging team through Pixel 4, and moved to Adobe in 2020 to lead a new universal camera project there.
Apple's Deep Fusion arrived with the iPhone 11 in 2019. It selected the sharpest frame from a short burst, then merged it pixel-by-pixel with multiple supporting frames using a neural network. The Photonic Engine, introduced with the iPhone 14 in September 2022, moved Deep Fusion earlier in the pipeline so that it ran on uncompressed image data, and Apple said the change improved low-light performance by up to 2.5 times on the main camera and 2 times on the ultra wide and TrueDepth cameras.
By 2019, machine learning had moved from in-camera burst merging to standalone desktop tools. Pixelmator Pro shipped ML Super Resolution in December 2019, using a convolutional neural network running on Core ML to upscale images by 3x while preserving sharpness. Topaz Labs released Gigapixel AI, DeNoise AI, and Sharpen AI as separate desktop applications across 2018 to 2021, then consolidated them into Topaz Photo AI in 2023.
Adobe added neural raw denoising to Lightroom in the April 2023 release. The feature, simply called Denoise, was trained on millions of pairs of high noise and low noise image crops to demosaic and denoise raw files in a single step. Adobe positioned it as a way to produce clean images at ISO 51200 from a 20 megapixel camera, and it supported both Bayer and X-Trans sensor patterns. DxO had been doing similar work in parallel with its DeepPRIME and DeepPRIME XD models inside PureRAW 3 and PhotoLab 7; PureRAW 4 and PhotoLab 8 followed in 2024 with DeepPRIME XD2.
The generative phase started in mid-2022 with Stable Diffusion and DALL-E 2 and reached photography editing tools in 2023. Adobe Firefly launched on March 21, 2023 as a separate web app, then Photoshop's Generative Fill arrived in the public beta on May 23, 2023, followed by Generative Expand in July 2023 and general availability in late 2023. Within the first year, Adobe reported that more than 13 billion images had been generated through Firefly across all products.
Google responded with Magic Editor, announced at I/O 2023 and shipped with the Pixel 8 on October 12, 2023. Apple followed about a year later with Apple Intelligence Clean Up, released in iOS 18.1 on October 28, 2024, and Image Playground, which shipped separately in iOS 18.2 on December 11, 2024. Samsung jumped in earlier, debuting Galaxy AI with Generative Edit on the Galaxy S24 on January 17, 2024. Generative Edit on the S24 ran in the cloud using Google's Imagen 2 model and stamped a visible AI watermark on the saved output.
The table below lists the main computational and AI features added to flagship phones and cameras since 2018. Some features run entirely on the device, some require a cloud round trip, and some are hybrid.
| Feature | Maker | Launched | Hardware | What it does |
|---|---|---|---|---|
| HDR+ | 2014 | Nexus 6, Pixel | Burst align and merge for dynamic range and noise | |
| Night Sight | November 2018 | Pixel and later | Handheld low light, machine learning white balance | |
| Deep Fusion | Apple | September 2019 | iPhone 11 | Frame selection plus pixel-by-pixel neural merge |
| ML Super Resolution | Pixelmator | December 2019 | Mac, Core ML | 3x upscale with edge preservation |
| Magic Eraser | October 2021 | Pixel 6 | One tap removal of people and objects | |
| Photonic Engine | Apple | September 2022 | iPhone 14 | Deep Fusion on uncompressed data, low light gain |
| Best Take | October 2023 | Pixel 8 | Swap heads between group photo frames | |
| Magic Editor | October 2023 | Pixel 8, Google Photos | Move, resize, and regenerate parts of an image | |
| Audio Magic Eraser | October 2023 | Pixel 8 | Remove background sound from video | |
| Generative Edit | Samsung | January 2024 | Galaxy S24, Imagen 2 | Erase, reposition, expand with cloud AI |
| Clean Up | Apple | October 2024 | iOS 18.1, Apple Intelligence | Brush over objects to remove them |
| Image Playground | Apple | December 2024 | iOS 18.2 | Generate images from prompts and Photos library |
| Real-time Recognition AF | Sony | 2022 onward | Alpha 7R V and later | Pose estimation, animal and bird tracking |
Sony's Alpha line moved subject recognition out of generic contrast and phase detection in 2022 with the Alpha 7R V, which added a dedicated AI processing unit using human pose estimation. The Alpha 1 II later widened the recognition set to animals, vehicles, and insects, and Sony has continued to ship recognition improvements through firmware. The Real-time Recognition AF+ in firmware 4.0 added more stable tracking of partially obstructed subjects, with about a 30 percent improvement in animal recognition and 50 percent in birds compared to the original Alpha 7R V models.
Nikon's Z9, Z8, and Z6 III similarly include AI based subject detection, while Canon's R3, R5 Mark II, and R1 use deep learning models trained on faces, eyes, and animals. None of these high end cameras use generative AI inside the image pipeline; the AI is used only to drive autofocus and metering.
Desktop editing tools have rapidly become AI assemblies. The table below covers the main applications used by photographers as of 2025.
| Tool | Maker | Notable AI features | Released |
|---|---|---|---|
| Adobe Lightroom | Adobe | Denoise, Lens Blur, Generative Remove, Distraction Removal | Denoise April 2023; Generative Remove May 2024 |
| Adobe Photoshop | Adobe | Generative Fill, Generative Expand, Remove Tool, Neural Filters | Generative Fill beta May 2023 |
| Photo AI 4 | Topaz Labs | Denoise, Sharpen, Recover Faces, Gigapixel upscaling, Super Focus, Dust and Scratch | Photo AI consolidated 2023; v4 May 2025 |
| Luminar Neo | Skylum | GenErase, GenSwap, GenExpand using Stable Diffusion | October 2023 onward |
| Pixelmator Pro | Pixelmator | ML Super Resolution, Magic Eraser, ML Match Colors | ML Super Resolution December 2019 |
| ON1 Photo RAW | ON1 | NoNoise AI, Resize AI, TackSharp AI, AI Masking | NoNoise AI 2021; Resize AI 2026 |
| Affinity Photo 2 | Serif (Canva) | Inpainting Brush, AI selection helpers | Affinity Photo 2 November 2022 |
| PureRAW 4 | DxO | DeepPRIME XD2, DeepPRIME XD3 | PureRAW 4 March 2024; XD3 2025 |
| PhotoLab 8 | DxO | DeepPRIME XD2s, ClearView Plus | October 2024 |
| Capture One | Capture One | AI Mask, Auto Adjustments | Smart Adjustments and AI Mask 2024 |
Lightroom's Denoise generally produces results competitive with Topaz and DxO, but it is GPU intensive and can take 20 to 60 seconds per image on a modest machine. Generative Remove, added in May 2024, can erase a tourist or a stray power line without burning a Firefly credit and works on raw files non-destructively. Distraction Removal, an extension of Remove introduced in October 2024, identifies and offers to remove unwanted background figures automatically.
Luminar Neo started shipping its generative tools in October 2023, with GenErase first, GenExpand and GenSwap following within two months. The tools are explicitly described as using a Stable Diffusion base model with Skylum's own fine tuning, and the licence is bundled with Luminar Neo subscriptions for a 12 month window.
Topaz Photo AI 4, released on May 8, 2025, added a Dust and Scratch model for old prints, a Super Focus model that combines multiple in-focus regions, refined Face Recovery with separate neck and hair controls, and roughly 30 percent faster processing on typical files.
Generative editing is where the technical story crosses into the cultural debate. The distinction matters: a 2023 Lightroom Denoise pass is a cleaner version of the same scene the camera saw, while a 2024 Generative Fill insertion of a sky or a person is by definition not in the original frame.
Adobe's Photoshop Generative Fill, announced on May 23, 2023, takes a selection and a text prompt and uses Firefly to fill the selection with photorealistic content. Generative Expand, added in July 2023, lets users widen the canvas and have Firefly invent the new edges. Both tools are non-destructive, save their output on a new layer, and stamp the result with Content Credentials.
Magic Editor on the Pixel 8 lets the user select a subject, drag it to a new position, and have Google's diffusion model regenerate the background. It can also enlarge or shrink the subject, change the sky, and remove unwanted objects. Magic Editor was originally exclusive to Pixel 8 and 8 Pro, then rolled out free to all Pixel 6 and later devices and to Google One subscribers across iPhone and Android in mid-2024.
Best Take, another Pixel 8 feature, addresses the classic group photo problem of someone blinking. It uses bursts of frames captured before and after the shutter, asks the user to pick a preferred expression for each face, then swaps the chosen face into the chosen frame. The output is one composite image that may never have existed as a single moment in time.
Apple Clean Up, part of Apple Intelligence in iOS 18.1, lets users brush over objects in Photos to remove them. The model runs partly on device for newer iPhones with the A17 Pro or A18 chip and partly on Apple's Private Cloud Compute servers. Image Playground, released in iOS 18.2, is a separate experience for generating images from prompts in three styles called Animation, Illustration, and Sketch; Apple does not currently offer a fully photorealistic style in Image Playground.
Outside the major suites, Black Forest Labs released FLUX.1 Kontext in 2024 and 2025 as a context-aware image editing model that takes both an input image and a text prompt and produces a new image with the requested change. The Pro and Max variants run as a hosted API; the Dev variant is open weight under a non commercial licence, with 12 billion parameters, designed to fit on consumer hardware. Kontext is widely used by photographers experimenting with style transfer, character preservation across multiple scenes, and selective edits without classic masking.
Magnific AI, founded in November 2023 by Javi Lopez and Emilio Nicolas in Murcia, Spain, became a popular generative upscaler. It uses latent diffusion to enlarge images up to 16x and invent plausible new detail, which makes it a strong tool for cleaning up old or low resolution shots but a problematic one for anyone trying to claim documentary accuracy. Freepik acquired Magnific in May 2024, and renamed itself Magnific in April 2026. Krea AI offers a similar enhancement pipeline alongside real-time generation and is widely used by concept artists.
Wedding, event, and portrait photographers routinely shoot thousands of frames per session and then spend hours selecting the keepers. Culling AI tries to compress that work into minutes.
Aftershoot, founded in 2020 by wedding photographer Justin Benson and engineer Harshit Dwivedi, became the dominant tool in this space. The app analyses each frame for sharpness, eyes-closed and eyes-open detection, facial expression, and visual duplicates, then ranks the photos. By 2025 the company had raised roughly 16 million dollars, employed about 100 people, and offered editing and retouching modules in addition to culling.
Narrative Select, from New Zealand based Narrative, takes a slightly different approach focused on speed and visual grouping. It clusters similar shots, highlights focus and closed-eye issues, and integrates with Lightroom for export. Some photographers prefer it for the lower automation level; others prefer Aftershoot's fully automatic mode.
FilterPixel and Picsellia handle smaller niches, and both Capture One and Lightroom have started experimenting with subject-aware selection inside their own catalogues.
The other major thread in AI photography is the attempt to prove what is real. The argument took shape in 2019, when Adobe formed the Content Authenticity Initiative (CAI) together with the New York Times Company and Twitter. Two years later, in February 2021, Adobe co-founded the Coalition for Content Provenance and Authenticity (C2PA) with Arm, the BBC, Intel, Microsoft, and Truepic. C2PA defines an open technical standard for tamper-evident metadata that travels with an image from capture to publication.
The metadata, branded Content Credentials, captures the camera make and model, the time and date, the location if the user allows it, and a record of any edits made in compliant software. Each entry is cryptographically signed, so a viewer with a C2PA reader can see if the chain was broken.
| Camera | Maker | Content Credentials availability |
|---|---|---|
| Leica M11-P | Leica | First production camera with C2PA, October 2023 |
| Sony Alpha 1 | Sony | Camera Authenticity Solution firmware March and April 2024 |
| Sony Alpha 9 III | Sony | Camera Authenticity Solution firmware April 2024 |
| Sony Alpha 7S III, 7 IV | Sony | C2PA support in 2024 firmware |
| Nikon Z6 III | Nikon | C2PA firmware demonstrated October 2024, released mid 2025 |
| Canon EOS R1, R5 Mark II | Canon | C2PA firmware July 2025 |
| Leica SL3-S | Leica | Content Credentials shipped 2025 |
Leica's M11-P, announced on October 26, 2023, was the first camera in the world to embed Content Credentials at the moment of capture. The camera contains a dedicated secure chipset that stores a certificate from the Leica certificate authority and signs each JPEG and DNG file as it is written. The authenticity information stays attached through any subsequent edit in a C2PA compliant application like Photoshop or Lightroom.
Sony rolled out a similar Camera Authenticity Solution through firmware updates in early 2024 to the Alpha 1, Alpha 9 III, Alpha 7S III, and Alpha 7 IV, and worked with the Associated Press on field testing through 2024. Nikon demonstrated its Z6 III C2PA implementation at Adobe MAX 2024 on October 14 to 16 and shipped firmware version 2.00 with full C2PA support in mid 2025, alongside bird detection autofocus and webcam mode. Canon joined the CAI in January 2023, added C2PA to the EOS R1 and R5 Mark II via firmware in July 2025, and continues to expand support across its mirrorless lineup.
Truepic, an early CAI member founded in 2016, provides a software development kit called Lens that allows mobile apps to capture images with C2PA signatures from inside iOS and Android. Lens runs 35 authenticity tests on each capture, including device attestation, screen recapture detection, and online provenance checking. Truepic was recognised in Time magazine's Best Inventions list in 2022 and ran a joint provenance pilot with Microsoft in 2023.
The past few years have produced a string of cases that forced editors, jurors, and the public to look at AI in photography more carefully.
German photographer Boris Eldagsen submitted a black and white portrait titled Pseudomnesia: The Electrician to the Open category Creative section of the 2023 Sony World Photography Awards. The image was made with DALL-E 2 and a number of post-generation edits. It won the Creative category in April 2023.
At the award ceremony in London on April 13, 2023, Eldagsen revealed that the image was AI generated and refused the prize. In a statement on his website he said he had entered "as a cheeky monkey" to test whether competitions were prepared for AI images, concluded they were not, and argued that AI images and photographs are different things and should not compete against each other. The World Photography Organisation said it had been aware of the AI co-creation; Eldagsen disagreed that this had been made clear. The incident is widely cited as the moment when major photography contests started writing explicit AI policies.
In June 2024, photographer Miles Astray entered an actual photograph of a flamingo, titled F L A M I N G O N E, into the AI category of the 1839 Color Photography Awards. The bird had folded its head behind its body to scratch itself, so the image looked like a pink ball on legs and read convincingly as a generative composition.
Astray won both the People's Vote and a Jury Award in the AI category. He then revealed that the photo was real. The 1839 Awards disqualified the entry, which Astray supported, and Lily Fierman, co-founder of the awards, said the experiment made an important point. The case ran as a mirror image of the Eldagsen story: a real photo passing as AI rather than AI passing as a real photo.
In March 2023, a Reddit user posting as ibreakphotos documented an experiment with the Samsung Galaxy S23 Ultra. They photographed a small, deliberately blurred image of the moon displayed on a monitor in a dark room and found that the resulting picture contained sharp lunar craters that were not present in the source on screen. The detail had been invented by Samsung's Scene Optimizer, which recognised the moon and applied a neural detail-enhancement model trained on moon images.
Samsung confirmed in a community post that since the Galaxy S21 series, Scene Optimizer had used "a detail enhancement engine" for the moon, but argued that the base photo was still real and the AI was only adding clarity. Critics countered that the enhancement was inserting detail that was not in the captured light, which the average user could not tell from a normal moon shot. The episode is a standard example in discussions of where computational photography ends and synthesis begins.
The 2024 1839 AI category also gave its first place to a fully AI generated image, a separate result from the Astray submission. The Australian News Photographer of the Year contest in 2024 disqualified a winning entry after suspicion that AI tools were used outside the rules. World Press Photo, after considering allowing AI submissions in 2023, formally banned generative AI in all categories for the 2024 contest, requiring entrants to submit camera files for any image that reached the final round.
Outside consumer photography, AI is doing some of its most useful work in scientific image collection.
Google launched Wildlife Insights on December 17, 2019, in partnership with Conservation International, the Smithsonian, the Wildlife Conservation Society, World Wildlife Fund, Map of Life, and the Zoological Society of London. The platform takes camera trap photographs uploaded by researchers and uses a custom AI model running on Google Cloud to classify the species. The system reportedly processes images about 3,000 times faster than humans, at around 3.6 million photos per hour, and launched with more than 4.5 million records contributed by partners. Google released a derivative open source model called SpeciesNet in 2025.
Wildbook, a project of the Wild Me non-profit, provides photo identification for terrestrial and marine species, and powers domain-specific instances such as Flukebook (whales and dolphins) and ACW (whale sharks). Flukebook combines seven identification algorithms across 15 species and 37 species-specific pipelines, and held more than 2 million photographs of 52,000 individual animals as of 2024. Microsoft AI for Earth funded both Wildlife Insights and Wildbook work as part of its conservation programme in the late 2010s.
In astrophotography, AI has changed the post-processing workflow as much as it has changed the camera. PixInsight is the dominant raw processing tool for serious astrophotographers, and a separate community of plugin developers has built deep learning tools that integrate with it. Russell Croman of RC-Astro shipped StarXTerminator, BlurXTerminator, and NoiseXTerminator for PixInsight and Photoshop; the tools are trained in a non-generative way to remove stars cleanly, deconvolve atmospheric blur, and reduce noise. Croman received the Photographic Society of America's Progress in Photography award in 2024 for that work. Topaz also markets a Star Tracker tool for astro use, though most serious users prefer the RC-Astro suite.
Medical and scientific imaging communities use AI heavily for tasks like microscopy denoising, histology stain normalisation, and automatic counting of cells or galaxies, but those use cases sit closer to image processing than to the consumer photography pipeline.
As generative tools have improved, a parallel industry has grown up around detecting fakes. The flagship academic figure is Hany Farid, a professor at the UC Berkeley School of Information who has been called the father of digital image forensics. Farid was part of the Microsoft team that developed PhotoDNA, used globally to fingerprint child sexual abuse imagery, and won a Guggenheim Fellowship in 2006 for his digital forensics work. He co-founded GetReal Security in 2024 to commercialise deepfake detection for enterprises, focusing on inconsistencies in shadows, geometry, and pixel statistics.
Reality Defender, founded in 2021 and based in New York, runs an enterprise API that scans images, audio, video, and text for AI manipulation without relying on watermarks. The company was named a Technology Pioneer by the World Economic Forum in 2025. Sensity AI focuses on visual deepfake detection, with engines that look at metadata, behavioural cues, and cross-modal inconsistencies. Truepic and the C2PA standard sit on the other side of the same problem, trying to make authenticity provable at capture rather than detectable after the fact.
For a working photojournalist, the practical advice as of 2025 is roughly: shoot to a camera with C2PA enabled, keep the original raw files, use Lightroom or Photoshop with Content Credentials turned on so the edit history travels with the file, and avoid generative tools entirely on news work. For investigative work the inverse is true: assume incoming images may be synthetic and run them through both metadata checks and a detection service like Reality Defender or Sensity before publishing.
The broader argument is whether a photograph still means what it used to. The question is not new. Pictorialists argued with straight photographers in the early 20th century about whether painterly manipulation was acceptable. Photoshop set off a similar argument in the 1990s. What is different now is the volume and the autonomy of the changes: a phone can rewrite a photo without the photographer asking.
The major photography press has tracked the shift closely. PetaPixel, DPReview, Fstoppers, and PhotographyLife have all published explainers on Generative Fill, Magic Editor, the Samsung moon case, Best Take, and the C2PA rollout. Wired, The Verge, the New York Times, and Bloomberg have run feature pieces on the philosophical question of whether a Pixel 8 portrait counts as a photograph or as a composite. World Press Photo's policy update, banning generative AI for the 2024 contest, was widely read as a marker of where the journalism community had landed.
Professional guilds have responded unevenly. The American Society of Media Photographers (ASMP) and the National Press Photographers Association (NPPA) have both pushed for clearer disclosure rules and supported C2PA. Stock agencies have done the opposite by accepting AI submissions. Adobe Stock began accepting AI generated content with mandatory labels in late 2022. Shutterstock partnered with OpenAI in October 2022 and launched its own AI generator in early 2023. Getty's Generative AI by Getty Images, trained on its own library and built on NVIDIA Picasso's Edify architecture, launched on September 25, 2023, and offers full indemnification for commercial use.
Is it still a photograph? In a strict sense, a Pixel 8 image with Best Take applied is a composite of frames the camera did capture, so a defender can argue it is still photographic. A Lightroom Generative Remove edit removes content, which is arguably similar to dodging or burning. A Photoshop Generative Fill that inserts a brand new person crosses a line that the journalism community has now formally drawn. Most photographers I talk to land somewhere in the middle: comfortable with neural denoising and small distraction removal, uneasy with anything that fabricates content, and quietly frustrated that the camera makers and phone makers do not always make the difference clear.
The verification side may be the most consequential thread of all. If Leica, Sony, Nikon, and Canon ship C2PA across their lineups, and if news organisations adopt it as a baseline, then within a few years the question will not be "is this real?" but "does this image carry a verified Content Credential?" That is not the same thing, but it is a workable proxy.