Hi Creators,
Earlier this year, we introduced Cube 3D, our foundation model for text to 3D creation. Our community’s creativity continues to amaze us with over a million 3D models generated in experiences and Studio since March.
The in-experience Mesh Generation API allows users to quickly generate unique in-game items and props using simple text prompts. We believe this will make creation even more accessible within gameplay and unlock new possibilities for dynamic and custom content within experiences. We have observed that games with creation dynamics result in significantly higher engagement and retention as users become more invested in experience they can actively shape and customize.
Our new model has been trained on over 3.5 million additional assets, increased our model’s output from 512 to 1024 tokens — essentially increasing the geometric detail it can generate for each shape — and refined our mesh simplification approach resulting in higher quality results. The new model also gives you greater creative control over the generations by using bounding boxes to inform the size of the generation. You can learn more about the technical details of the Cube 3D model here.
Quality and Rate Limit Improvements
We are improving the Cube 3D model and the developer experience for the Mesh Generation API with the following updates:
-
Improved Mesh Fidelity: Our model now produces meshes with better overall forms, silhouettes, more detail, and fewer artifacts.
-
Cleaner Textures: Generated assets now have more coherent and detailed textures with less graininess and less pronounced baked in lighting, making your creations look better than ever.
-
Better Understanding and Creativity: Our improved model more accurately interprets your text prompts, including a much better ability to creatively adhere and blend different concepts and styles. This gives you more precise control over the details of your generated assets, while ensuring outputs are safe and suitable.
-
Increased Rate Limits: We are updating our API rate limit baseline from 5 generations per minute to 30 generations per minute, and this limit will be a soft limit to ensure your players are able to generate more. Additionally, in the coming weeks we will be moving to scaling the limit with concurrent player count for broader adoption.
Check out the updated Cube 3D Play Experience to see these improvements. In this experience, all the assets were built using the latest version of Cube 3D.
Control Generations with a Bounding Box
We know generated assets are useful when they fit the scenarios you create, whether that’s filling a specific spot in a room or creating a new model that slots into an existing modular set. With this update, we’re introducing a powerful and flexible new control to the GenerateMeshAsync API.
You can now provide an optional SuggestedSize parameter to guide the model to generate an asset within a specific proportion. You can optionally pass a Vector3 value representing your desired proportions and size (X, Y, Z) to the GenerateMeshAsync API. This is more than just a rescale, the model uses your Vector3 input to generate a 3D shape that respects those proportions from the start. While the model is quite flexible, requesting proportions that don’t suit your intended object may produce less predictable results.
Above you can see some examples of the model’s generations when varying dimensions and proportions are provided.
To learn more, please check out the updated Generation Service API documentation.
We also created a SuggestedSize Template experience for creators as a reference API implementation. You can go to experience details and click on … to edit it in Studio.
Enabling the Mesh Generation API
To enable this API in your experience, you will need to turn on Editable Mesh / Editable Image APIs by going toStudio File > Game Settings > Security and selecting the following:
To learn more, please check out the updated Generation Service API documentation.
Generation Improvements Now Available in Studio
These quality improvements are also live when you generate in Studio! You can take advantage of the enhanced model quality using Assistant to accelerate your prototyping, level design, or asset creation workflows.
Go to File > Beta Features and ensure “Assistant Mesh Generation” is enabled. Then, use the /generate command in the Assistant chat.
Open Source Update
We are committed to sharing our research and contributing to the open-source community. The latest version of our Cube 3D model for mesh generation is also available on GitHub and Hugging Face, so anyone can experiment with these latest improvements.
What’s Next?
This is another step in our journey to build powerful and intuitive generative AI tools. As we shared previously, we are continuing to work toward 4D capabilities that allow avatars to immediately use an asset, like a drivable vehicle, as well as expanding into full scene generation. We’re excited about what’s to come and will share more details soon.
Share Your Thoughts!
To provide feedback, you can sign up for Mesh Generation office hours. If you would like to participate, please complete the sign-up form here.
We’re also eager to see how you utilize these new features and hear your feedback. Please continue to share your planned generative AI use cases, desired features, and any thoughts you have below. Your feedback will directly contribute to refining and enhancing our models.
Thank you!
FAQs
Click here to view the FAQ!
Is Mesh replication supported?
- No, Mesh replication is not supported at this time. Assets generated in-experience are currently visible only to the user who generated them. We are working to make this available with safety at the forefront.
Do these changes affect whether generated assets persist in a user’s inventory?
- No, this update does not change how generated assets persist. Assets generated in-experience do not persist outside of that session and are not added to a user’s inventory. Assets generated in Studio via Assistant are saved to your inventory and accessible in the Creator Hub.
Is there a cost to use the Mesh Generation API and what is its rate-limit?
- There is no cost associated with using the Mesh Generation API. With this update, we’ve increased the baseline from 5 generations per minute to 30 generations per minute for an experience. This is currently a soft limit, meaning you may be able to temporarily exceed this limit. In the coming weeks, we will be moving to scaling the limit with concurrent player count for broader adoption. In the future, we may allow creators to pay for additional usage to go beyond this baseline allowance.
How can these tools benefit artists and creators?
- These tools can significantly speed up workflows, allowing you to rapidly prototype ideas, create variations and concepts, or quickly populate an experience with background objects. This allows you to focus your time and skills on the hero assets and your creative vision.
Where does the Cube 3D model training data come from?
- The model for 3D generation is trained on a library of 3.5 million 3D assets utilizing a combination of synthetic and procedural data, licensed and publicly available datasets, as well as the free assets available in our Creator Store. It does not include experience data or assets in a user’s inventory. Creators can also continue to control their preferences with respect to how their data is used as part of our AI data sharing program.
How is Roblox ensuring that developer and IP holder data is protected from cloning / IP infringement?
- We take the intellectual property rights of all rights holders seriously. Rights Manager allows creators to report content they believe is copyright infringing and file a removal request. We are continually refining our approach and have plans to provide more proactive tools to enable IP owners to better manage their IP across Roblox later this year. Stay tuned for more updates we will announce here.
