Toggle search
Toggle menu
5.5K
6.1K
231
53.7K
VR & AR Wiki
Toggle preferences menu
Toggle personal menu
Not logged in
Your IP address will be publicly visible if you make any edits.

Unity Sentis

vrarwiki.com · VR & AR Wiki
More actions
Sentis
Information
Type Neural network inference library (Unity package)
Industry Video games
Developer Unity Technologies
Operating System Cross-platform (all Unity runtime platforms)
License Unity Terms of Service
Supported Devices Unity runtime platforms; Meta Quest 3 and Meta Quest 3S in Meta's samples
Release Date June 2023 (closed beta); 14 August 2024 (version 2.0.0)
Website https://docs.unity3d.com/Packages/com.unity.ai.inference@latest


Unity Sentis, listed in the Unity Package Manager as Sentis, is a neural network inference library for the Unity engine, made by Unity Technologies. It imports trained models in the ONNX, LiteRT and PyTorch formats and runs them inside a Unity application on the end user's device, using its central processing unit (CPU) or graphics processing unit (GPU).[1] Unity has named object identification, speech recognition and "smart NPCs" as example uses, and says that the majority of AI models can run this way "without the need for cloud infrastructure".[2]

Sentis replaced Barracuda, an inference library developed by Unity Labs that Unity introduced in 2019 as the Unity Inference Engine, bundled with the Unity ML-Agents Toolkit.[3][4] Sentis entered closed beta in June 2023[5] and reached a release version, 2.0.0, in August 2024.[6][7] In May 2025 Unity renamed the package Inference Engine and gave it a new package ID, com.unity.ai.inference; in October 2025 the display name returned to Sentis, while the new package ID stayed.[8][1] As of November 2026 the latest version is 2.6.1, dated 2 April 2026, which requires Unity 6 or later.[9]

In XR development, Meta's samples for the Passthrough Camera API and its AI Building Blocks for Unity use the library to detect real-world objects on Meta Quest 3 and Meta Quest 3S headsets, and researchers have measured Barracuda's inference times on HoloLens 2.[10][11][12]

Reviewed 1 November 2026. Checked all dates, versions, names, quotes and figures against Unity's package changelogs, registry, manuals and announcements, Meta's developer documentation and GitHub samples, and the full texts of the two HoloLens 2 studies. About review dates.

History

Barracuda

Unity introduced the library on 1 March 2019 by bundling it with version 0.7 of the ML-Agents Toolkit, which it said "now uses the Unity Inference Engine (codenamed Barracuda)". The new engine replaced TensorFlowSharp, a third-party plugin that had limited ML-Agents to five platforms: Windows, Linux, Mac, iOS and Android. Unity described the engine as a product of its Unity Labs research team, based on IL2CPP, Burst and compute shaders, with a size of 600 KB at the time. According to Unity, building the 3D Balance Ball example for iOS produced a 135 MB build with TensorFlowSharp and an 83.5 MB build with the new engine. An earlier version had run agents trained with ML-Agents in Jelly Bowl, an Xbox game by What Up Games that Unity showed at Unite Los Angeles in October 2018.[3]

Barracuda then became a package of its own. Version 0.2.7 (27 September 2019) was the first available in the Unity Package Manager, 0.3.0 (29 October 2019) added an importer for .onnx files, and 1.0.0 (1 June 2020) was the first verified release. The last major version, 3.0.0 (24 January 2022), combined the 2.1 to 2.4 improvements into a verified release.[13] Its documentation calls Barracuda production-ready for ML-Agents and several other network architectures and in a preview stage for other uses, and states that the Sentis package has replaced it.[4]

Sentis beta and release

Unity announced Unity Sentis on 27 June 2023, together with the Unity Muse creation tools. The press release called Sentis "a cross-platform runtime inference engine" for deploying AI models in Unity projects and said both products were in closed beta, with general availability expected later that year.[5] The first beta build, 1.0.0-exp.6, is dated 26 June 2023. Version 1.2.0-exp.2 (20 September 2023) removed a watermark on scenes that referenced Sentis and the requirement for a license key.[6] By late November 2023 the package was in a free open beta for all Unity developers,[2] and the documentation for versions 1.3 to 1.6 described it as a "pre-release" package to be verified for release near the end of 2024.[14][15] ML-Agents moved to Sentis in version 3.0.0-exp.1 (9 October 2023), which also deprecated Barracuda.[16]

In January 2024 Unity announced an integration with Hugging Face: models that its team had tested with Sentis, tagged "Unity Sentis" and preconverted to the .sentis format. The initial 11 models covered tasks such as sentence similarity, speech to text and object detection. The same announcement said Sentis would officially launch later that year as part of the Unity 6 release.[17] Sentis 2.0.0, dated 14 August 2024, was the first version whose documentation described the package as being in a "release" state.[7] It introduced a generic Tensor class, renamed Execute to Schedule throughout the API and removed the separate GPUCommandBuffer backend in favour of a unified GPUCompute backend.[6]

Inference Engine and the return to Sentis

On 15 May 2025 Unity announced Unity AI with the Unity 6.2 beta. The announcement said that once Unity 6.2 reached general availability, Unity AI would replace both Muse and Sentis. It listed Inference Engine as one of the three components of Unity AI, alongside Assistant and Generators, and stated: "Inference Engine replaces Sentis."[18] Inference Engine 2.2.0, which its changelog calls the "First version of Inference Engine", is dated the same day. Sentis 2.1.3, in the old com.unity.sentis package and also dated 15 May 2025, added a documentation notice about the rename.[8][6] Moving a project from Sentis 2.1 to Inference Engine 2.2 meant replacing the package with com.unity.ai.inference and changing references from the Unity.Sentis namespace to Unity.InferenceEngine, either by hand or with Unity's automatic API updater.[19]

The changelog entry for version 2.4.0 (22 October 2025) reads "Renamed Inference Engine to Sentis in package name and documentation".[8] The user manual describes the change as affecting only the package display name: the package can be found under either name in the Package Manager, and projects need no code changes.[1] The package ID is still com.unity.ai.inference, and code in the 2.6 documentation still uses the Unity.InferenceEngine namespace.[9][20] ML-Agents followed the package through both names, upgrading to Inference Engine 2.2.1 in version 4.0.0 (28 August 2025) and to Sentis 2.6.1 in version 4.0.3 (17 April 2026).[16] As of November 2026, Unity's AI product page answers the question "What is happening to Sentis?" with "Sentis remains active and serves as the engine for running neural network models natively within the Unity Runtime."[21]

Version history

Dates are those given in the package changelogs. The package registry shows later publication dates for some versions.

Name Version Changelog date Notes
Barracuda 0.2.7 27 September 2019 First version in the Unity Package Manager[13]
Barracuda 1.0.0 1 June 2020 First verified release[13]
Barracuda 3.0.0 24 January 2022 Verified release combining the 2.1 to 2.4 improvements[13]
Sentis 1.0.0-exp.6 26 June 2023 First beta build[6]
Sentis 1.2.0-exp.2 20 September 2023 Watermark and license-key requirement removed[6]
Sentis 1.4.0-pre.1 5 April 2024 Quantization API, functional API, serialization reworked to use FlatBuffers[6]
Sentis 2.0.0 14 August 2024 First version documented as a release; generic Tensor class; unified GPUCompute backend[6][7]
Sentis 2.1.2 18 December 2024 Support for DirectML acceleration "in future Unity versions"[6]
Sentis 2.1.3 15 May 2025 Documentation notice of the rename to Inference Engine[6]
Inference Engine 2.2.0 15 May 2025 First version under the new name and package ID[8]
Inference Engine 2.3.0 15 July 2025 Model Visualizer[8]
Sentis 2.4.0 22 October 2025 Display name changed back to Sentis; LiteRT import; tokenization API; STFT and DFT operators[8]
Sentis 2.5.0 23 January 2026 PyTorch import; Hugging Face tokenizer parser; CoreCLR-compatible APIs[8]
Sentis 2.6.0 20 March 2026 ONNX opset support up to version 25[8]
Sentis 2.6.1 2 April 2026 Documentation fix; latest version as of November 2026[8][9]

Features

Sentis imports ONNX models with opset versions 7 to 25, LiteRT models (the format formerly called TensorFlow Lite) and PyTorch exported programs decomposed to Core ATen IR operators.[1] Models with ONNX opset versions outside that range might still import, but the documentation warns that results can be unpredictable. Unity also points developers to the validated models on its Hugging Face page.[22] An imported model can be serialized as a .sentis file, which uses FlatBuffers and can be saved to the project's StreamingAssets folder.[23] Sentis is made available under the Unity Terms of Service.[24]

To run a model, an application creates a worker for a chosen backend. The worker breaks the model into tasks and processes its layers in sequence.[20][25]

Backend Runs on Documented use
CPU CPU, as Burst jobs Faster than the GPU for small models or when inputs and outputs are on the CPU[20]
GPUCompute GPU, with compute shaders and command buffers "Generally the fastest backend for most models"; uses DirectML on platforms that support DirectX 12[20]
GPUPixel GPU, with pixel shaders Only for platforms that lack compute shader support[20]

Output tensors can still be pending when they are returned. Reading one back to the CPU with ReadbackAndClone blocks until the backend finishes, while methods such as ReadbackAndCloneAsync avoid blocking the main thread. When an operator is supported on the CPU but not on the GPU, Sentis can fall back to the CPU, which requires reading tensors back from the GPU and uploading results again.[25] For models that are too slow for one frame, the ScheduleIterable method runs a model a layer at a time. The documentation's example spreads a 50-millisecond model over 10 frames at 5 milliseconds each, because running it in one frame "might cause stuttering or low frame rates in gameplay".[26]

Feature Added in Description
Functional API 1.4.0-pre.1 Building models in code with "torch-like syntax"[6]
Weight quantization 1.4.0-pre.1 Stores float weights used by operations such as Dense, MatMul and Conv as Float16 or Uint8 to reduce disk and memory use; Sentis converts them back to higher precision at runtime, and Uint8 may significantly affect accuracy, depending on the model[6][27]
Model encryption 1.3.0-pre.1 (sample project) Saving and loading a model through an encrypted stream; the documentation gives an AES example[6][28]
Model Visualizer 2.3.0 Read-only node graph of an imported model inside the Unity Editor[8][29]
Tokenizer 2.4.0 Converts text into numerical tokens for models that process text; compatible with the tokenizer.json files of the Hugging Face tokenizers library, with a Hugging Face parser added in 2.5.0[8][30]
Spectral operators 2.4.0 STFT, DFT and window-function operators[8]

Supported platforms

Unity states that Sentis supports all Unity runtime platforms, and that performance depends on the model's operators and complexity, the device's hardware and software constraints, and the type of engine used.[1] The CPU backend relies on Burst, which compiles to WebAssembly on WebGL and may be slower there. GPUCompute needs compute shader support, and GPUPixel is meant for platforms without it.[20] Barracuda's GPU inference did not support OpenGL ES on Android and iOS, where Vulkan or Metal was needed; OpenGL Core on Mac, where Metal was needed; or WebGL, where only CPU inference was available.[31]

Package versions Minimum Unity Editor version
Barracuda 3.0 2019.4.29f1[4]
Sentis 1.0 to 1.2 2021.3[32]
Sentis 1.3 to 1.6 2023.2[32]
Sentis 2.0 to 2.1 2022.3[32]
Inference Engine 2.2 to 2.3, Sentis 2.4 to 2.6 Unity 6 (6000.0)[9]

Use in XR development

Unity samples

Unity's November 2023 introduction to Sentis described an augmented reality sample in which a depth-estimation network processes camera video so that real-world objects can occlude virtual ones. Unity said the sample works on mobile devices with a camera and does not need a lidar sensor.[2] The sample's documentation states that its inference is split over two frames and runs at a lower frequency than the refresh rate, to keep the phone from overheating.[33] The same post suggested the Ultralytics YOLO model for detecting objects in a VR game scene, or real-world objects in an AR device's camera feed.[2]

Meta Quest

The GitHub repository of Meta's Unity samples for the Passthrough Camera API begins with an initial commit dated 12 March 2025. That first version included a MultiObjectDetection sample that fed headset camera images to Sentis 2.1.1 and a YOLO model to recognize real-world objects. The README of that first version described the camera API as experimental and warned that the model, optimized for Quest 3 and 3S, "may not achieve 100% accuracy".[34] The current README lists Unity Inference Engine 2.2.1 for that sample, Unity 6000.0.38f1 or newer, and Quest 3 or Quest 3S with Horizon OS v74 or higher.[35]

Meta's guide to the integration, last updated 18 May 2026, lists these requirements: a Quest 3, Quest 3S or Quest 3S Xbox Edition with Horizon OS v74 or later, Unity Editor 6000.0.66f2 or later, version 2.2.1 of the com.unity.ai.inference package, Mixed Reality Utility Kit 85.0.0 or later, and camera and spatial data permissions for the app. The sample uses the Yolov9t model from the MultimediaTechLab/YOLO repository, converted to a .sentis file with an added non-maximum suppression layer and quantized to Uint8, for a size of about 2.3 MB. The model was trained on 80 classes, so some objects are grouped into broader classes. Meta notes that some classes are hard to identify, giving cell phones as an example, and that Quest 3 controllers may be missed or labelled as remote controls.[10]

Meta's AI Building Blocks for Unity, first released in version 83, pair an agent component with a provider that sets where inference runs: in the cloud, on a local machine or on the headset. In the overview's table of typical use cases, the recommended provider for real-time object detection is "Unity Inference Engine / HuggingFace", and Meta describes the Unity Inference Engine provider as running "optimized models directly on Meta Quest hardware". Meta's recommended setup is Unity 6 or newer, a Meta Quest 3 or 3S, and Meta XR Core SDK and MR Utility Kit v83 or later. The page's release notes for version 85 and later state that this provider supports object detection, image segmentation and on-device small language model chat.[11] Meta's v85 API reference describes UnityInferenceEngineProvider as supporting object detection and text-only LLM chat with SmolLM, Qwen, Phi and GPT-2 models, and as supporting on-device inference only.[36]

Research

Two 2023 studies measured the on-device inference times of Unity Barracuda on HoloLens 2, whose Qualcomm Snapdragon 850 chip includes an Adreno 630 GPU.[12][37]

Zaccardi and colleagues (Sensors, 2023) compared Barracuda 3.0, running on the HoloLens 2 GPU with the ComputePrecompiled worker in Unity 2021.3, with Windows Machine Learning (WinML) in a Universal Windows Platform application. Each model was timed over five repetitions of 200 images. Barracuda had the lower mean inference time for 7 of 8 synthetic test models, and the difference between the two frameworks was statistically significant for all of them (Mann-Whitney U test, p < 0.005). The authors concluded that with Barracuda, models with fewer than 107 multiply-accumulate operations could reach application frame rates above 30 fps, while models of about 1010 operations, such as ResNet50 and U-Net, would likely have inference times "of the order of seconds". A surgical model containing gated recurrent units could not be tested with Barracuda because that layer type was not supported.[12]

Łysakowski and colleagues ran YOLOv8 object detection on HoloLens 2 with Barracuda, in a paper first posted as an arXiv preprint in June 2023 and published at the 2023 IEEE International Conference on Edge Computing and Communications. They noted that their Barracuda version supported ONNX opsets only up to version 9, so models had to be exported accordingly, and they aimed for end-to-end processing under 100 ms. With the smallest model, YOLOv8n, and 160 by 160 pixel input, detection took less than 100 ms and still detected every instance of the test object, a smartphone, at distances closer than 2.5 m. The authors also noted that Barracuda did not support FP16 or INT8 quantization at the time.[37]

Study Model and setup Barracuda inference time (mean) Comparison
Zaccardi et al. 2023 YOLOv2-Tiny 630 ± 32 ms WinML 1330 ± 56 ms[12]
Zaccardi et al. 2023 EgoPose 384 ± 14 ms WinML 1164 ± 43 ms[12]
Zaccardi et al. 2023 ResNet50 701 ± 14 ms WinML 1842 ± 64 ms[12]
Łysakowski et al. 2023 YOLOv8n, 160 by 160 input, ComputePrecompiled GPU backend 89.84 ms Compute backend 103.64 ms; CSharpBurst CPU backend 216.11 ms[37]

Limitations

Unity's documentation lists model types that Sentis does not support: tensors with more than eight dimensions, sparse input tensors or constants, string tensors and complex-number tensors. Some data types, such as booleans, are converted to floats or integers, which can increase memory use.[22] Not every operator runs on every backend, CPU fallback adds readback and upload time, and some inputs, such as the shape input of an Expand operation, have to be read on the main thread.[25]

For Meta Quest, Meta's guide states that the library runs on the main thread of the Unity app and therefore affects other main processes such as the render pipeline, and that "Large generative models and LLMs do not perform well on devices like Meta Quest 3". Loading large models can cause long loading times, lag on the main thread during the first inference and a smaller memory budget for other resources. The guide also says the library "doesn't use NPU or hardware acceleration to run the inferences" and runs on Quest "as a regular Android platform with no specific accelerations". It recommends layer-by-layer inference to avoid blocking the main thread and Uint8 quantization to reduce loading times.[10] In January 2024 Unity had described plans to integrate with hardware acceleration libraries and neural processing units through platform interfaces such as Microsoft DirectML, Apple Core ML/MPS Graph and Google NN API.[17] The 2.6 documentation lists DirectML acceleration for the GPUCompute backend on DirectX 12 platforms.[20]

See also

References

  1. ↑ 1.0 1.1 1.2 1.3 1.4 "Sentis overview (Sentis 2.6.1)". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/manual/index.html. Retrieved 2026-11-01.
  2. ↑ 2.0 2.1 2.2 2.3 Bill Cullen (2023-11-29). "Create next-gen features with AI models using Unity Sentis". Unity Blog. Unity Technologies. https://unity.com/blog/games/create-next-gen-ai-models-with-unity-sentis. Retrieved 2026-11-01.
  3. ↑ 3.0 3.1 Vincent-Pierre Berges, Renaldas Zioma (2019-03-01). "Unity ML-Agents Toolkit v0.7: A leap towards cross-platform inference". Unity Blog (Internet Archive copy). Unity Technologies. https://web.archive.org/web/20210514172114/https://blogs.unity3d.com/2019/03/01/unity-ml-agents-toolkit-v0-7-a-leap-towards-cross-platform-inference/. Retrieved 2026-11-01.
  4. ↑ 4.0 4.1 4.2 "Introduction to Barracuda". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/manual/index.html. Retrieved 2026-11-01.
  5. ↑ 5.0 5.1 "Unity Unleashes the Power of AI for Creators, Supercharging Content Creation and User Engagement". Unity Investor Relations. Unity Technologies. 2023-06-27. https://investors.unity.com/news/news-details/2023/Unity-Unleashes-the-Power-of-AI-for-Creators-Supercharging-Content-Creation-and-User-Engagement/default.aspx. Retrieved 2026-11-01.
  6. ↑ 6.00 6.01 6.02 6.03 6.04 6.05 6.06 6.07 6.08 6.09 6.10 6.11 6.12 "Changelog - Sentis 2.1.3". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/changelog/CHANGELOG.html. Retrieved 2026-11-01.
  7. ↑ 7.0 7.1 7.2 "Sentis overview (Sentis 2.0.0)". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/manual/index.html. Retrieved 2026-11-01.
  8. ↑ 8.00 8.01 8.02 8.03 8.04 8.05 8.06 8.07 8.08 8.09 8.10 8.11 "Changelog - Sentis 2.6.1". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/changelog/CHANGELOG.html. Retrieved 2026-11-01.
  9. ↑ 9.0 9.1 9.2 9.3 "com.unity.ai.inference". Unity Package Registry. Unity Technologies. https://packages.unity.com/com.unity.ai.inference. Retrieved 2026-11-01.
  10. ↑ 10.0 10.1 10.2 "Unity Inference Engine For On-Device ML/CV Models". Meta Horizon OS Developers. Meta. 2026-05-18. https://developers.meta.com/horizon/documentation/unity/unity-pca-sentis. Retrieved 2026-11-01.
  11. ↑ 11.0 11.1 "AI Building Blocks - Overview". Meta Horizon OS Developers. Meta. 2026-07-08. https://developers.meta.com/horizon/documentation/unity/unity-ai-building-blocks-overview/. Retrieved 2026-11-01.
  12. ↑ 12.0 12.1 12.2 12.3 12.4 12.5 S. Zaccardi, T. Frantz, D. Beckwée, E. Swinnen, B. Jansen (2023-10-25). "On-Device Execution of Deep Learning Models on HoloLens2 for Real-Time Augmented Reality Medical Applications". Sensors, vol. 23, no. 21, article 8698. https://doi.org/10.3390/s23218698. Retrieved 2026-11-01.
  13. ↑ 13.0 13.1 13.2 13.3 "Changelog - Barracuda 3.0.2". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/changelog/CHANGELOG.html. Retrieved 2026-11-01.
  14. ↑ "Sentis overview (Sentis 1.3.0-pre.3)". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/manual/index.html. Retrieved 2026-11-01.
  15. ↑ "Sentis overview (Sentis 1.6.0-pre.1)". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/manual/index.html. Retrieved 2026-11-01.
  16. ↑ 16.0 16.1 "Changelog - ML Agents 4.0.3". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/changelog/CHANGELOG.html. Retrieved 2026-11-01.
  17. ↑ 17.0 17.1 Bill Cullen (2024-01-25). "Find the right AI model for your game with Unity and Hugging Face". Unity Blog. Unity Technologies. https://unity.com/blog/games/hugging-face-ai-models-and-more-sentis-updates. Retrieved 2026-11-01.
  18. ↑ Bill Cullen (2025-05-15). "Unity 6.2 Beta is now available". Unity Discussions. Unity Technologies. https://discussions.unity.com/t/unity-6-2-beta-is-now-available/1639999. Retrieved 2026-11-01.
  19. ↑ "Upgrade to Sentis 2.5 or 2.6". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/manual/upgrade-guide.html. Retrieved 2026-11-01.
  20. ↑ 20.0 20.1 20.2 20.3 20.4 20.5 20.6 "Create an engine to run a model". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/manual/create-an-engine.html. Retrieved 2026-11-01.
  21. ↑ "Unity's AI Tools for Game Development". Unity. Unity Technologies. https://unity.com/products/ai. Retrieved 2026-11-01.
  22. ↑ 22.0 22.1 "Supported models". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/manual/supported-models.html. Retrieved 2026-11-01.
  23. ↑ "Serialize a Model". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/manual/serialize-a-model.html. Retrieved 2026-11-01.
  24. ↑ "Sentis license". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/license/LICENSE.html. Retrieved 2026-11-01.
  25. ↑ 25.0 25.1 25.2 "How Sentis runs a model". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/manual/how-sentis-runs-a-model.html. Retrieved 2026-11-01.
  26. ↑ "Split inference over multiple frames". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/manual/split-inference-over-multiple-frames.html. Retrieved 2026-11-01.
  27. ↑ "Quantize a Model". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/manual/quantize-a-model.html. Retrieved 2026-11-01.
  28. ↑ "Encrypt a model". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/manual/encrypt-a-model.html. Retrieved 2026-11-01.
  29. ↑ "Inspect models with the Model Visualizer". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/manual/visualizer.html. Retrieved 2026-11-01.
  30. ↑ "Tokenize text for input". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/manual/tokenizer.html. Retrieved 2026-11-01.
  31. ↑ "Supported platforms (Barracuda)". Unity Documentation. Unity Technologies. https://docs.unity3d.com/Packages/[email protected]/manual/SupportedPlatforms.html. Retrieved 2026-11-01.
  32. ↑ 32.0 32.1 32.2 "com.unity.sentis". Unity Package Registry. Unity Technologies. https://packages.unity.com/com.unity.sentis. Retrieved 2026-11-01.
  33. ↑ "Depth Estimation Sample". GitHub - Unity-Technologies/sentis-samples. Unity Technologies. https://github.com/Unity-Technologies/sentis-samples/blob/main/DepthEstimationSample/README.md. Retrieved 2026-11-01.
  34. ↑ "Unity-PassthroughCameraApiSamples README (initial commit)". GitHub. Meta. 2025-03-12. https://github.com/oculus-samples/Unity-PassthroughCameraApiSamples/blob/b48d61dd43ffafbd2e2de3ae63fa3719685391c3/README.md. Retrieved 2026-11-01.
  35. ↑ "Unity-PassthroughCameraApiSamples". GitHub. Meta. https://github.com/oculus-samples/Unity-PassthroughCameraApiSamples. Retrieved 2026-11-01.
  36. ↑ "UnityInferenceEngineProvider (v85)". Meta Horizon OS Developers API reference. Meta. https://developers.meta.com/horizon/reference/unity/v85/class_meta_x_r_building_blocks_a_i_blocks_unity_inference_engine_provider. Retrieved 2026-11-01.
  37. ↑ 37.0 37.1 37.2 M. Łysakowski, K. Żywanowski, A. Banaszczyk, M. R. Nowicki, P. Skrzypczyński, S. K. Tadeja (2023-06-06). "Real-Time Onboard Object Detection for Augmented Reality: Enhancing Head-Mounted Display with YOLOv8". arXiv preprint 2306.03537; also published in the 2023 IEEE International Conference on Edge Computing and Communications (EDGE), pp. 364-371. https://arxiv.org/abs/2306.03537. Retrieved 2026-11-01.
Retrieved from "https://vrarwiki.com/index.php?title=Unity_Sentis&oldid=53298"