This tracker collects Google and Nvidia filings on photo and video editing, covering per-region edits, text-guided recoloring, compression, and smoother zoom previews. The batch shows camera software moving from single-pass fixes toward layered, context-aware editing built into the pipeline.
108 filings · tracking since Apr 2026 · latest Aug 2026 · updates automatically as new filings publish
US 2026/0229000 A1
A generative model trained to infer absent objects from scene context rather than user prompts pushes the editing race toward fully automated completion, moving past text-guided tools toward systems that recognize what a scene should contain.
US 2026/0228858 A1
The tracker has focused on post-capture adjustments and guided edits. Qualcomm's filing shifts upstream by segmenting objects during the capture stage itself, letting the camera meter each element independently before any blending occurs.
US 2026/0230685 A1
Motion transfer between unrelated subjects extends the editing race beyond appearance changes into movement synthesis, letting users impose choreography or gesture patterns without frame-by-frame animation.
US 2026/0228949 A1
The tracker so far maps which editing operations get automated; this filing shows the selection process itself becoming algorithmic, letting software pick the right reconstruction method before committing compute resources.
US 2026/0228944 A1
The tracker's focus on regional edits finds confirmation here: Google is building explicit face-preservation zones into its editing pipeline, treating facial regions as protected areas while allowing free modification elsewhere in the frame.
US 2026/0228932 A1
The race so far has focused on editing existing photo content; this filing moves upstream to object insertion, automating where new elements land rather than requiring manual placement work.
US 2026/0228856 A1
Reconstructing detail from compressed feature maps during inference sidesteps the blurriness that standard upscaling introduces when vision systems need to work with shrunk representations.
US 2026/0228508 A1
Nvidia trains generators to sharpen output by applying refinement in a secondary pass, steering models away from the statistical averaging that produces soft features in early generation stages.
US 2026/0220735 A1
Dual-level region analysis sharpens object boundary detection, a core bottleneck when users need clean selections in camera editing workflows where background separation determines final image quality.
US 2026/0220901 A1
The race so far has focused on editing existing pixels. This filing moves into synthetic insertion, automating the geometry and lighting work that currently demands manual adjustment to avoid visible compositing.
US 2026/0220746 A1
Splitting noise removal from detail recovery lets the pipeline restore texture in low-light shots without amplifying grain, shifting the phone's post-processing toward selective refinement rather than blanket smoothing.
US 2026/0222669 A1
The race includes live compositing: Sony's method isolates screen content during recording so broadcast graphics and talent stay on separate layers, skipping the rotoscoping work that usually follows.
US 2026/0220741 A1
The per-region edits strand gains a key mechanism: semantic segmentation that sorts pixels by content type before applying distinct filter chains, solving the washed-out face problem.
US 2026/0220739 A1
Reconstructing detail from low-resolution input by first mapping the blur structure itself rather than guessing at missing pixels. Extends the tracker's compression and zoom preview coverage into satellite and remote imaging where pixel recovery matters most.
US 2026/0220834 A1
Auto-extracted keywords let users skip typing prompts, reducing friction in the text-guided editing workflow that other players require. Confirms the race is moving toward lower-barrier interfaces for AI recoloring and scene swaps.
US 2026/0220881 A1
Reconstructing 3D geometry from multi-angle shots lets editors reposition light sources in post rather than reshooting or manually correcting each frame, a shift from 2D per-image fixes toward object-space relighting.
US 2026/0220803 A1
The race so far has focused on what edits change, color, compression, region masks. This filing adds depth sensing to make those edits physically plausible across foreground and background layers.
US 2026/0222693 A1
Storing a reference face model lets the system correct white balance drift by comparing live skin tones to known baseline, sidestepping the camera's general-purpose light guessing that fails under mixed lighting.
US 2026/0220898 A1
The timeline has tracked text-guided edits and per-object changes; Samsung's filing adds temporal state shifts, aging or freshening a single object backward or forward in time without reshaping the photo around it.
US 2026/0220823 A1
Dual-model inference for decompression pushes beyond single-path restoration, letting competing approaches fight for artifact removal where one trained method alone leaves traces behind.
US 2026/0222692 A1
The race's white balance problem gets regional: Sony's filing shows how to handle mixed lighting in a single frame instead of picking one source globally.
US 2026/0212661 A1
The race so far has focused on what AI does after seeing a full image. This filing reveals how Samsung trains the model itself, using masked tiles to force deeper visual reasoning rather than surface pattern matching.
US 2026/0214134 A1
Adaptive bitrate assignment during encoding lets compression parameters shift frame-by-frame based on motion and scene content rather than applying uniform settings across an entire video, improving efficiency for mixed-pacing clips.
US 2026/0212458 A1
The tracker has focused on guided edits and regional control; this filing extends into automated multi-defect repair, showing Google moving from user-directed fixes toward systems that identify and correct several problems without manual problem specification.
US 2026/0212468 A1
The race so far has focused on what to change in a photo or video. Samsung's filing introduces a prior question: whether a visible flaw is actually a flaw. The system distinguishes intentional tilts from accidental ones before attempting correction.
US 2026/0212473 A1
The race to move beyond one-size-fits-all edits now includes region-specific enhancements. Samsung's system identifies sky and face separately to apply tuned contrast and color to each, solving the sunset-photo tradeoff.
US 2026/0212110 A1
Syncing transcript edits back to raw media files lets users reshape audio and video from the text layer instead of the timeline, shifting the editing surface from waveform manipulation to document-style word deletion.
US 2026/0212454 A1
Separating brightness correction from color grading during multi-shot merging lets Samsung preserve natural tones where exposures overlap, a key step toward seamless HDR and panorama output that won't need manual fixes.
US 2026/0212460 A1
Selective noise filtering based on human visual perception lets the system prioritize smooth faces and subjects over peripheral areas, reducing computational load while maintaining perceived image quality in camera processing.
US 2026/0205558 A1
Users could watch stretched or gap-filled video without the blurring and stuttering that plague current frame-insertion tools, shifting the burden from post-production timing choices to predictive motion modeling.
US 2026/0203960 A1
Users won't need to describe what they want changed; the system scans the image itself and offers preset alternatives for each detected element, speeding up the edit-select workflow that text-guided approaches require.
US 2026/0203881 A1
Merging multiple exposures into one frame requires separate AI models for highlights and shadows, a two-stage approach that avoids forcing a single model to balance conflicting exposure demands.
US 2026/0203867 A1
Poor facial visibility in dim environments blocks smooth video calls. Microsoft's filing describes real-time relighting that reconstructs underlit faces without hardware fixes or manual repositioning.
US 2026/0203996 A1
Reconstructing stable 3D scenes from video requires filtering transient objects like pedestrians and cars during model training, a step Google's method automates rather than leaving to manual cleanup afterward.
US 2026/0203977 A1
Users won't have to manually dial in edit intensity for each step, the system learns when to apply instructions heavily versus lightly across different image regions, reducing over-processing and artifact creep that text-guided edits typically suffer from.
US 2026/0203972 A1
Semantic photo reconstruction from natural language lets editors skip manual masking and regeneration workflows by describing desired changes in plain text.
US 2026/0203956 A1
Compression artifacts need selective recovery. Google's filing lets a single compressed file generate multiple detail levels on decode, pushing the recreation work to the viewing device rather than baking in one quality choice at compression time.
US 2026/0195847 A1
Adaptive training on source footage itself rather than fixed datasets lets the upscaler tune to each video's specific noise patterns and content, sharpening results beyond generic models that must work across all input types.
US 2026/0195946 A1
Predicting future lighting and color shifts within a single frame moves beyond static filters into temporal reasoning, extending the tracker's coverage beyond per-region edits into physics-aware scene dynamics.
US 2026/0192191 A1
The timeline so far has focused on single-frame edits and static images. Sony's filing shifts the problem to temporal consistency, where frame-to-frame coherence matters more than individual frame quality in video and games.
US 2026/0197561 A1
The confidence-based white balance system extends the race beyond binary correction into graduated adjustments, letting the camera pull back when lighting conditions resist a clean answer.
US 2026/0195948 A1
Within the race's focus on per-region edits, Adobe's filing confirms the race is moving into spatial reasoning: automatically inferring light direction and shadow geometry to make composited objects blend into their new backgrounds.
US 2026/0197538 A1
The race has focused on guided edits and compression; this filing shows IBM automating the selection and sequencing layer itself, pulling finished videos from raw libraries on command.
US 2026/0196248 A1
Swapping video backgrounds per viewer without re-rendering the foreground object adds a personalization layer to the editing race, moving beyond one-size-fits-all recoloring into dynamic content delivery that preserves the hero product across versions.
US 2026/0187769 A1
Sampling color values from adjacent frames to correct isolated color shifts in animation sequences fills a gap between Samsung's earlier text-guided recoloring work and the broader video consistency problem that Google and Nvidia have been pursuing.
US 2026/0187860 A1
Preserving 3D geometry during generation lets users reposition elements and adjust lighting after the fact, moving AI images closer to the malleable asset status of traditional 3D renders rather than locked photographs.
US 2026/0187971 A1
Separating compression noise from real image detail during streaming, Nvidia's layered detection approach identifies miscolored pixels without blurring edges, a direct fix for the banding and grid patterns that plague low-bandwidth video playback.
US 2026/0187815 A1
Frame-to-frame motion tracking lets the upscaler maintain pixel continuity across fast cuts and pans, addressing the compression artifacts that plague quick camera movements in low-res video.
US 2026/0189799 A1
Exposure bracketing during capture lets the camera recover detail across extreme lighting ratios instead of forcing a single-shot compromise, extending the editing race beyond software reconstruction to hardware capture strategy.
US 2026/0187755 A1
The compression and preview category now includes selective sharpening: Intel's approach separates intentional blur from quality loss before applying filters, letting devices sharpen only what needs it rather than degrading soft backgrounds.
US 2026/0189723 A1
Selective AI upscaling based on codec entropy signals reduces compute waste by skipping frames where compression artifacts are minimal, addressing the efficiency gap between per-frame processing and practical deployment.
US 2026/0187808 A1
The restoration race has focused on uniform processing across images. Samsung's filing suggests a priority-based training method, restoring background areas before subjects, which could give AI models a steadier foundation for reconstructing faces and details.
US 2026/0187764 A1
The noise-reduction race now extends to hardware artifacts. Microsoft's approach learns sensor-specific dark current patterns rather than relying on post-shot filtering, which could let cameras deliver cleaner raw files upstream of editing software.
US 2026/0187774 A1
Retrieving similar photos from a database to justify quality comparisons lets the system ground its verdicts in real examples rather than abstract metrics, moving the feedback loop closer to how humans actually evaluate camera output.
US 2026/0170616 A1
Photos with mismatched details across regions, a dog's leg disconnecting from its body, a couch edge that vanishes, would become rare if AI regions could share information during generation instead of working in isolation.
US 2026/0170722 A1
The race so far has focused on rewriting photos after capture, but Adobe's approach lets users preset exact attribute levels before generation happens, shifting control from post-shot prompts to pre-generation dials.
US 2026/0170654 A1
Batch object removal speeds up the core editing workflow by identifying and extracting multiple subjects simultaneously rather than forcing users through repetitive individual selections.
US 2026/0170264 A1
Automating prompt refinement shifts the bottleneck from user skill to machine interpretation, letting Adobe's software bridge the gap between casual description and generator-ready specification without manual editing cycles.
US 2026/0170323 A1
Breaking images into numbered chunks like text tokens lets Google run photo editing through the same AI models that already power its language systems, sidestepping the need for separate image-specific architectures.
US 2026/0162341 A1
Users typing casual commands like "zoom out" could get photos recomposed as if the camera had physically moved, shifting the race from editing pixels to repositioning the virtual camera itself.
US 2026/0162328 A1
The race expands beyond fixing what's already visible: Google is now patenting the ability to reconstruct objects that got cut off or blocked entirely, then reposition them anywhere in the frame.
US 2026/0162320 A1
Pinning down which words map to which image regions during training. Google's method uses contrastive learning to force its generator to match text descriptions with sharper spatial precision, moving beyond approximate semantic alignment.
US 2026/0154897 A1
Photo edits need to know which parts they're inventing versus observing. Google's approach runs multiple reconstructions to flag where different 3D interpretations fit the same 2D image, letting the system mark its uncertainty rather than guess confidently.
US 2026/0154857 A1
The race has focused on style transfer and scene generation separately. Google's filing shows how to preserve object identity while applying artistic styles, solving the problem of detail loss that happens when you blend a photo into a painting.
US 2026/0154872 A1
Keeping bad outputs invisible: Google embeds a quality check inside generation itself rather than showing you failures and asking you to retry.
US 2026/0156368 A1
Splitting exposure control by object instead of by frame region lets the system preserve face detail without blown-out skies, a different approach from traditional HDR's zone-based compromise.
US 2026/0154795 A1
Photos lose detail when bright areas max out the sensor. Google's method reconstructs those clipped pixels by training a generative model to invent plausible detail, rather than trying to recover what was never recorded.
US 2026/0154861 A1
The race has focused on editing existing photos, but Google's filing expands the battlefield to generation itself, showing how to make AI consistently reproduce a user's specific subject rather than generic versions.
US 2026/0154903 A1
Reconstructing individual objects in 3D space from limited reference images solves a bottleneck in photo-to-model workflows where generic generation fails to preserve real-world specifics like wear patterns and exact colorways.
US 2026/0154796 A1
Running the removal process twice, first to fill the gap and then to refine the boundary, Adobe targets the artifact that haunts content-aware fill: the visible seam where synthetic pixels meet the original photo.
US 2026/0148426 A1
The race so far has focused on fixing images after generation. Adobe is working the problem backward: building exclusion logic into the generation itself so unwanted elements never appear in the first place.
US 2026/0148473 A1
Scaling 3D human generation from controlled studio data to unstructured internet photos, eliminating the need for specialized capture hardware.
US 2026/0148474 A1
Training a generative model on messy, uncontrolled 2D photos sidesteps the need for expensive 3D capture rigs, letting companies build poseable digital humans from internet-scraped images instead of controlled studio data.
US 2026/0148429 A1
Making realistic pose changes with minimal user input. Adobe's system lets editors drag sparse control points to reshape objects, cutting the manual masking and warping that currently makes physical adjustments tedious.
US 2026/0148449 A1
Splitting image generation across multiple models in sequence lets each one focus on a specific quality problem, roughness, then color, then sharpness, rather than one model juggling everything at once.
US 2026/0148427 A1
Embedding character traits into the model itself rather than regenerating them per-prompt, so the same person stays recognizable across a multi-image sequence without manual intervention between frames.
US 2026/0148475 A1
Building realistic 3D human models from random internet photos instead of studio footage would let photo editors generate new poses and angles without reshoots, a major cost cut for the AI retouching pipeline.
US 2026/0148448 A1
The race so far has centered on which tools can eliminate manual selection work. UniTune pushes further: it removes the need for masks, sketches, or crops entirely, letting plain text commands drive edits without any preparatory steps from the user.
US 2026/0148337 A1
Fitting photos to different aspect ratios without losing subject matter requires generating plausible pixels where none existed. Google's approach expands the image first, then crops, rather than choosing between distortion and loss.
US 2026/0148478 A1
Real-time radiance field rendering would let editors apply light and view changes instantly rather than waiting for re-renders. Nvidia's Morton-code sorting accelerates the voxel lookup bottleneck that currently makes these edits feel sluggish.
US 2026/0148479 A1
The race so far has focused on editing existing photos. Nvidia's filing pivots to building editable 3D scenes from photos in real time, a foundation that could let users rewrite images from multiple angles before flattening them back to 2D.
US 2026/0141619 A1
Reconstructing intermediate frames from multi-camera video in 3D lets editors pull usable shots from moments never actually recorded, filling gaps between the camera's capture rate and creative intent.
US 2026/0141590 A1
Users could reshape their photos by pointing to reference images instead of describing what they want, speeding up the editing workflow and reducing reliance on written prompts to guide the AI.
US 2026/0141631 A1
Gaussian splatting with learned motion tokens lets the system infer dynamic 3D geometry and movement directly from ordinary photo sequences, avoiding the depth sensors or manual annotation that currently bog down video-to-3D conversion.
US 2026/0141594 A1
Companies racing to edit photos after capture need AI that follows spatial rules. Adobe's condition-map approach shifts the control earlier, letting users blueprint where objects land before generation even starts.
US 2026/0141502 A1
Photos with flawed exposure or focus could be auto-corrected before the user ever sees them, shifting the race from editing after capture to preventing bad shots in the first place.
US 2026/0141572 A1
Keeping spatial relationships stable in generated images requires enforcing layout constraints before rendering, which Adobe's method does by mapping object positions during the generation process itself rather than correcting them after.
US 2026/0141498 A1
Generating inpainting in two stages, global layout first, then local detail refinement, lets the AI match surrounding textures and perspective without the blurry compromise of single-pass methods.
US 2026/0141593 A1
Photos edited by AI would stay visually consistent with your existing brand instead of drifting toward generic templates. Adobe's approach extracts stylistic rules from your reference images to constrain what the generator can produce.
US 2026/0141615 A1
The race so far has centered on reconstructing 3D scenes from photos in real time. Nvidia's patent removes a practical barrier: making Gaussian-based rendering work with the wide-angle and specialty lenses that phones and cameras actually use.
US 2026/0134545 A1
Photos could be edited faster if one AI model handles cutting out subjects, isolating skies, and detecting people in a single pass instead of running three separate networks.
US 2026/0134546 A1
The race so far has focused on rewriting pixels after capture. Adobe's approach targets the selection problem itself, training a network to automatically refine messy masks rather than fix the final image.
US 2026/0134260 A1
Generating clothing that physics engines can actually use cuts out the manual rigging step that currently requires technical artists to make garments move realistically in simulations.
US 2026/0134524 A1
Picking the right AI model for inpainting depends on scene complexity. Adobe's system analyzes visual difficulty first, then routes the job to whichever model handles that specific type of scene best.
US 2026/0134547 A1
A neural network that refines masks separately for each disconnected region stops Adobe's selection tool from averaging detail across a person's hand, fork, and hair as one blob.
US 2026/0134591 A1
Semantic segmentation of body parts before color assignment lets the network apply different color logic to faces versus hair versus fabric, solving the greenish-skin problem that plagues simpler colorization tools.
US 2026/0134348 A1
Photos edited after capture expand from still images into video: Nvidia's approach converts footage into riggable 3D models, letting creators animate people rather than just retouch them.
US 2026/0134602 A1
The race so far has focused on editing existing photos. Nvidia is pushing sideways into generating characters from scratch, automating the entire rigging and skeleton setup that normally requires separate specialist work.
US 2026/0134627 A1
Generating photorealistic edits requires 3D geometry that won't break under physics simulation, Nvidia's approach builds that constraint directly into the generation step rather than fixing problems afterward.
US 2026/0134516 A1
Within the race to move photo AI onto devices rather than cloud servers, Google is focusing on the training problem: how to pack restoration quality into models small enough for smartphones to run without lag.
US 2026/0134603 A1
Automating the rigging and skeleton step, making generated characters actually movable, pushes the race beyond static image editing into animated content creation.
US 2026/0129292 A1
Running two depth sensors in parallel lets the system reject bad focus locks when one sensor gets fooled by close subjects, improving the raw material that AI photo editing later works with.
US 2026/0127723 A1
Photos edited with this system would rely on AI trained through a structured pipeline that separates the learning process itself, potentially making the fill-in results more consistent and predictable than current erase tools.
US 2026/0129298 A1
The race so far has focused on what happens after you shoot. Google is now patching the moment before, smoothing the preview itself so zooming feels like one continuous motion instead of a jarring skip.
US 2026/0127726 A1
Filtering out photos too sharp or well-composed to need work saves processing power by skipping unnecessary enhancement cycles on already-good shots.
US 2026/0120288 A1
Applying different edits to sky, subject, and background in a single pass cuts out the manual masking step that slows down even quick phone edits.
US 2026/0120346 A1
Generative models that iterate toward a text description reveal how companies are automating color grading entirely, moving the workflow from slider adjustments to natural language prompts.
US 2026/0120334 A1
Storing a personalized AI model alongside compressed photos means the reconstruction happens on your device, not in the cloud, reducing the server load that photo editing at scale would otherwise demand.
Most filings in this batch come from Google and cluster around making photo edits smarter rather than flashier. The company is patenting systems that treat different regions of an image separately, decide which photos deserve heavier processing, recolor scenes from text prompts, compress and rebuild images without losing detail, and smooth out jumpy zoom previews. Several filings also target autofocus on small close-up subjects and better methods for removing unwanted objects from photos.
Nvidia's contribution to this storyline points in a different direction: turning descriptions and video into animatable 3D characters. One filing builds 3D characters from text prompts, another reconstructs them from video footage, and a third generalizes the pipeline for simulation-ready output. Compared with Google's per-pixel refinements, Nvidia's filings suggest interest in generating whole synthetic figures rather than editing existing photos.
Watch for filings that blend these two threads: automated decisions about which images or scenes need heavy AI intervention, and pipelines that move beyond flat photos into layered or synthetic representations. The steady flow of restoration, inpainting, and pipeline filings suggests both companies are building toward AI editing that runs with less user input, deciding on its own what a photo needs before touching it.
These are patent filings, which show research direction, not confirmed features. Google has filed systems for per-region editing, text-based recoloring, compression, and zoom smoothing, but a patent does not guarantee any of these ship in a camera app. It signals where engineering time is going, not a release date.
Nvidia's filings in this batch focus on turning text descriptions or video footage into animatable 3D characters, which overlaps with photo and image work through shared AI techniques like generative modeling. It is not photo editing in the traditional sense, but the underlying methods for building and refining visual content connect the two companies' patents.
Several filings return to the same challenges: deciding which photos need heavy AI processing, editing different regions of an image with different rules, and shrinking images without losing quality on rebuild. Restoration and inpainting also appear more than once, suggesting both companies see removing flaws and filling gaps as an ongoing area worth patenting repeatedly.
Not necessarily, and not on any predictable timeline. Patents describe technical approaches companies want to protect, sometimes years before a feature reaches a device, if it ever does. This storyline is a useful way to see what problems Google and Nvidia are actively working on, not a preview of your next software update.
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →
See plain-English patent summaries from every company we track.