This watchlist tracks Nvidia's filings on turning text prompts, photos, and video into rigged, animatable 3D characters and full digital humans, replacing manual modeling and studio capture. Together, the filings point to steady research into generating simulation-ready characters from whatever input is easiest to get.
28 filings · tracking since May 2026 · latest Nov 2026 · updates weekly
based on all tracked filings in this watchlist · refreshes every week
Nvidia's 25 filings all point toward one goal: letting computers build, move, and talk through realistic digital characters using AI, starting from simple inputs like photos, text, or audio.
The filings concentrate most on two problem areas: stopping AI-animated faces and bodies from producing visual glitches, and teaching characters to move naturally from basic instructions like scripts or spoken words.
a dated entry each week this watchlist moves · older entries stay archived
This week's filing covers turning a single photo of a person into a moving 3D head. Nvidia is exploring ways to build animated face models from very little input.
This week's filing focuses on giving animated characters the ability to shift mood while they are still speaking. The work moves toward characters that feel more alive and less robotic in how they express themselves.
This week's filing covers turning a written description into realistic surfaces on a plain 3D shape. The focus is on using text to tell software how a 3D object should look and feel visually.
This week's filings all focus on fixing glitches. Two patents tackle AI faces and bodies breaking apart during animation, while a third looks at building 3D objects from smaller files.
This week's filing covers a system that builds 3D characters able to move into any body position. The focus is on turning text into figures that can be posed freely.
This week's filing focuses on making digital faces move and speak in real time by reading audio. The push is toward characters that react to sound as it happens, not after the fact.
This week's filing covers a tool that reads a script and automatically draws out a full set of scenes. The focus shifts from building 3D characters to planning and picturing whole stories visually.
This week's new filing focuses on cleaning up raw 3D scans so they can actually be used as digital models. Nvidia is working on ways to take messy scan data and turn it into something a computer can work with properly.
Nvidia’s tracked filings by month · newest first
15 or more filings that month · 5 to 14 · under 5
the problems Nvidia keeps filing on · each with its three newest filings · new filings join every week
Building a usable 3D character from words or a simple photo is slow and requires expert tools. These filings cover automated pipelines that take text descriptions or everyday photos and produce 3D characters ready to move and be placed in simulations.
Making a virtual character move in a way that looks natural is hard to do without hand-crafting every action. These filings cover AI approaches that generate and predict character motion from simple instructions or learned rules about how bodies move.
Creating realistic digital humans normally requires expensive studio capture sessions. These filings cover systems that build full-body digital human models from ordinary 2D photos or video footage of real people.
A description like a flowing medieval tunic becomes an object built to work inside physics simulators.
Either a written description or a single photo can be the starting point for a finished 3D model.
The face reacts to how you speak, picking up emotion in your tone as well as the words.
One model makes a complete 3D person, face, hands and body, trained mostly on ordinary 2D photos.
every tracked filing, month by month · counts are USPTO pre-grant publications, one per publication number (an application, not a granted patent)
US 2026/0350914 A1
Building volumetric head geometry from a single photo sidesteps the distortion problem that plagues 2D warping methods and preserves accessories like hair and glasses as separate 3D objects rather than losing them in facial deformation.
US 2026/0341272 A1
Rigged characters could now shift expressions within a single line of dialogue instead of freezing one emotion for the full take, making dialogue feel less wooden and closer to how real actors modulate their performance.
US 2026/0311422 A1
After generating 3D geometry from text, the system now learns material properties by reverse-engineering rendered video of a plain model, solving the lighting problem that makes generated objects look wrong when environments change.
US 2026/0301322 A1
After text-to-3D generation, the bottleneck shifts to inference speed and storage. This patent cuts data footprint by encoding 3D geometry as layered 2D grids, letting the model generate characters faster and deploy them more efficiently.
US 2026/0301289 A1
After establishing how to generate characters from text, this filing addresses the real-time responsiveness problem: keeping facial animation smooth when the system switches from looped idle states to generating a live reply.
US 2026/0301288 A1
After mouth geometry, controlling which facial regions animate during speech. This patent constrains motion to the lips and jaw, preventing the ripple artifacts that currently leak into adjacent areas like cheeks and forehead when characters speak.
US 2026/0289888 A1
Surface-element visibility tied to depth and viewing angle solves the clipping and blur problems that arise when rigged characters move through extreme poses.
US 2026/0278899 A1
After establishing how to generate 3D characters from text and video, this filing targets the animation layer: getting facial movements, mouth, eyes, teeth, to sync with speech in real time rather than requiring manual keyframing or post-processing.
US 2026/0253280 A1
Automated storyboarding from scripts removes a manual bottleneck before 3D character generation even starts, letting creators move directly from story to asset production.
US 2026/0245306 A1
Turning a raw 3D scan of a real object into a clean, usable digital model is one of the messiest problems in computer graphics. Nvidia's new patent describes a machine-learning approach that may make that process significantly more automatic.
US 2026/0237323 A1
Real-time sign language output would let deaf participants follow conversations at natural speaking pace rather than waiting for captions. This filing shows Nvidia pursuing accessibility as a direct application of its character animation tech.
US 2026/0212578 A1
Generating animations on demand from high-level commands sidesteps the need to pre-record or manually craft every possible movement a 3D character might perform, letting game engines request poses and transitions dynamically instead.
US 2026/0212165 A1
Training a single model on both motion reconstruction and generation lets the network learn shared patterns that improve performance on each task, rather than building separate specialized tools for partial data versus text-driven invention.
US 2026/0212573 A1
The pipeline shifts from static character generation to autonomous movement: after learning from motion-capture data, the AI must produce convincing motion without direct video reference, relying only on reward signals for realism.
US 2026/0212543 A1
Constraint-based motion generation lets artists specify only key behavioral rules, what must happen and what must not, while the system synthesizes natural motion between keyframes, cutting manual pose adjustment work.
US 2026/0212622 A1
The pipeline now includes a correction layer that catches anatomical implausibility before output, using learned body proportions rather than manual fixes.
US 2026/0204038 A1
A 3D character generated this way would actually look like the reference photo instead of drifting into the AI's generic defaults. The patent reveals how Nvidia fuses text and image inputs so neither one overwhelms the other.
US 2026/0195932 A1
After establishing how to generate 3D characters from text and reference images, this filing shows how to choreograph multiple characters together using motion from video references, controlling both identity and movement in a single pass.
US 2026/0188341 A1
Extracting emotion intensity from speech audio lets the system sync facial expressions to vocal tone rather than words alone, closing the gap between what a character says and how it actually sounds.
US 2026/0187860 A1
The text-to-3D pipeline needs generated images to stay editable after creation. This filing confirms Nvidia is embedding 3D structure directly into the generation process rather than trying to reconstruct it afterward.
US 2026/0148473 A1
A single foundation model trained on 2D internet photos eliminates the need for separate specialized networks for face, hands, and body, compressing what normally requires studio equipment into one unified system.
US 2026/0148474 A1
Ready-to-animate characters need consistent body geometry across poses. This filing shows how to extract that consistency from messy, uncontrolled photos instead of requiring studio capture rigs.
US 2026/0148475 A1
The 3D character pipeline so far relies on controlled capture data; this filing shows how to work backward from random internet photos with conflicting angles and lighting to reconstruct consistent human geometry.
US 2026/0134260 A1
Generating clothing that behaves correctly under physics simulation requires predicting how fabric deforms and moves. This filing shows how Nvidia chains text-to-3D generation with physics-aware training to produce garments that don't need manual rigging.
US 2026/0134348 A1
Multi-angle video capture lets the system build skeletal structure and surface geometry in a single pass, collapsing what normally requires separate motion-capture and modeling stages into one automated workflow.
US 2026/0134602 A1
Animators would skip the manual rigging step entirely, with the diffusion model generating skeletons and joint controls automatically alongside geometry. This shrinks the pipeline from separate modeling, rigging, and animation phases into one model output.
US 2026/0134627 A1
Within the character pipeline, this filing shows how Nvidia routes text descriptions directly to physics-compatible geometry, cutting the traditional bottleneck of manual asset creation that slows down simulation environments.
US 2026/0134603 A1
Rigging animated characters requires skeleton structures and joint definitions that most AI models generate poorly. This filing shows how to extract usable skeletal data directly from the diffusion model's output, bypassing manual rigging work.
Across the filings, Nvidia is trying to replace manual 3D modeling and rigging with an AI pipeline that starts from a text prompt, a photo, or video footage and outputs a character ready to animate and simulate. The filings describe research direction, not a shipped product, so the exact workflow may change before anything reaches users.
The filings describe simulation-ready 3D characters and full digital humans generally, without naming a specific game or product. The same pipeline could apply to games, film, or other 3D content, but the patents themselves stay focused on the underlying generation method rather than a named use case.
Existing tools generally need a trained artist to build and rig a character by hand, or a controlled capture studio to scan a real person. Nvidia's filings describe skipping both steps, generating a rigged, animatable character directly from a text description, ordinary video, or uncontrolled 2D photos instead.
Patents describe what a company is exploring, not what it has committed to ship. This watchlist shows Nvidia filing multiple related pipelines around text-to-3D and photo-to-3D characters, which signals sustained research interest, but there is no guarantee any of it becomes an announced product.
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →