Unity AI Models and Partners
Unity AI is powered by a variety of artificial intelligence (AI) models that are integrated into the Unity Editor to provide a simple interface and economy for users. Unity AI may change model providers and models over time to power, or add to these functionalities. Some models are hosted on Unity’s first-party server infrastructure, and some are hosted on third-party infrastructure and accessed via a partner API (“Partner Models”). This page identifies the AI models used by Unity and each model provider.
Assistant
The Unity AI Assistant uses large language models (LLMs) to answer user questions, generate code, and run agentic actions such as modifying a large number of files. Improving Unity AI with Developer Data sharing to train and improve these models is turned OFF by default, but you can edit that setting in the Unity AI Cloud Dashboard if you wish to help us improve Unity AI for all users. Developer Data, in this case, includes your prompts, responses, interactions, code, and other content. Note that even if Developer Data sharing is turned on, your data is not sent to Partner Model providers since these models are hosted on Unity first-party servers. In any case, you continue to own all your input and output data when using the Assistant. Here are the models we use the power the Assistant:
| Model and Provider | Purpose | Hosting |
|---|---|---|
GPT-series from Azure Open AI Services | GPT-models are used at various stages in the Assistant pipeline. | Unity |
Llama-series from Meta Platforms, Inc. | Meta Llama-models are used at various stages in the Assistant pipeline. | Unity |
Model and Provider
| Model and Provider | Purpose | Hosting |
|---|---|---|
GPT-series from Azure Open AI Services | ||
| Purpose GPT-models are used at various stages in the Assistant pipeline. | ||
| Hosting Unity |
| Model and Provider | Purpose | Hosting |
|---|---|---|
Llama-series from Meta Platforms, Inc. | ||
| Purpose Meta Llama-models are used at various stages in the Assistant pipeline. | ||
| Hosting Unity |
Generators
The Unity AI Generators use several first-party (Unity) and Partner Models to generate and refine assets. Some useful information on Generators Partner Models:
- Data transfer: We send your anonymized Developer Data, including prompts, reference assets, etc. to these Partner Model providers for the sole purpose of running the services. Your data is deleted by Partner Model providers after generation, except for custom sprite model training where the resulting custom model is saved until you delete it.
- Data ownership: You own all your input and output data when using Unity AI, including when using Partner Models.
- Model training: Partner Model providers do not train their models with your Developer Data, even if you enable “Improve Unity AI” in the Unity cloud dashboard.
- Custom models: Unity AI offers custom trained model functionalities to users. When you choose to create a custom trained model, that model is retrained exclusively on the data you upload for that purpose. Only users inside your Unity organization can use this custom trained model, and the data uploaded to retrain the Custom Model is not used to improve the partner foundation model.
Here are the models we use to power the Generators:
| Model Provider | Purpose | Hosting |
|---|---|---|
Sprite | ||
Scenario models include a variety of pre-trained LoRAs (e.g. fine-tuned styles) trained on Stable Diffusion, FLUX, Bria, and GPT-Image foundation models. The models can generate images (“sprites”) from text and can optionally use several image reference types including hand drawn doodles to guide generations. Scenario models also power some post-process refinements. | Scenario | |
Layer models include a variety of pre-trained LoRAs or “styles” trained on Stable Diffusion and FLUX foundation models. The models can generate images (“sprites”) from text and several image reference types including hand drawn doodles. Layer models also power (i) post-process refinements, and (ii) custom model training with your choice of foundation models like Stable Diffusion, FLUX, or Bria. | Layer AI | |
Texture | ||
Unity | Unity first-party model that generates textures from text, image, and pattern references. | Unity |
Animation | ||
Unity | Unity first-party model that generates humanoid animations. This is accessible via the “text-to-animation” prompt. | Unity |
Kinetix model generates humanoid animations from video prompts. This is accessible via the “video-to-animation” prompt. | Kinetix | |
Sound | ||
Unity | Unity first-party model that generates sound from text and sound prompts. The sound refinements are local operations which are not powered by AI. | Unity |
Model Provider
| Model Provider | Purpose | Hosting |
|---|---|---|
Sprite | ||
| Purpose | ||
| Hosting |
| Model Provider | Purpose | Hosting |
|---|---|---|
| Purpose Scenario models include a variety of pre-trained LoRAs (e.g. fine-tuned styles) trained on Stable Diffusion, FLUX, Bria, and GPT-Image foundation models. The models can generate images (“sprites”) from text and can optionally use several image reference types including hand drawn doodles to guide generations. Scenario models also power some post-process refinements. | ||
| Hosting Scenario |
| Model Provider | Purpose | Hosting |
|---|---|---|
| Purpose Layer models include a variety of pre-trained LoRAs or “styles” trained on Stable Diffusion and FLUX foundation models. The models can generate images (“sprites”) from text and several image reference types including hand drawn doodles. Layer models also power (i) post-process refinements, and (ii) custom model training with your choice of foundation models like Stable Diffusion, FLUX, or Bria. | ||
| Hosting Layer AI |
| Model Provider | Purpose | Hosting |
|---|---|---|
Texture | ||
| Purpose | ||
| Hosting |
| Model Provider | Purpose | Hosting |
|---|---|---|
Unity | ||
| Purpose Unity first-party model that generates textures from text, image, and pattern references. | ||
| Hosting Unity |
| Model Provider | Purpose | Hosting |
|---|---|---|
Animation | ||
| Purpose | ||
| Hosting |
| Model Provider | Purpose | Hosting |
|---|---|---|
Unity | ||
| Purpose Unity first-party model that generates humanoid animations. This is accessible via the “text-to-animation” prompt. | ||
| Hosting Unity |
| Model Provider | Purpose | Hosting |
|---|---|---|
| Purpose Kinetix model generates humanoid animations from video prompts. This is accessible via the “video-to-animation” prompt. | ||
| Hosting Kinetix |
| Model Provider | Purpose | Hosting |
|---|---|---|
Sound | ||
| Purpose | ||
| Hosting |
| Model Provider | Purpose | Hosting |
|---|---|---|
Unity | ||
| Purpose Unity first-party model that generates sound from text and sound prompts. The sound refinements are local operations which are not powered by AI. | ||
| Hosting Unity |
Inference Engine
The Unity Inference Engine enables you to run AI models on your local machine in the Unity Editor, or on the end-user device in the Unity runtime. No data from these models are stored or transferred to the cloud. Inference Engine does not include built-in models, but rather allows you to import your own custom pre-trained models, or ones acquired from model gardens like Hugging Face. See our documentation for further information and supported workflows.
If you have questions about how Unity AI uses various AI models, please contact legal@unity3d.com.