CSM AI

What is CSM AI?

CSM AI was developed by Common Sense Machines; its core product, Cube, is a 3D creation platform designed for use in gaming, design, animation, and simulation. It starts with text, sketches, product images, or character illustrations, generates reference images from these, converts them into 3D meshes, and then proceeds with tasks such as refinement, segmentation, texturing, and animation.

The platform is designed to serve as a tool for 3D creation collaboration, rather than a replacement for all modeling software. The quick Preview feature is useful for exploring shapes, while Refine is used to improve the quality of meshes and textures. The final assets can still be imported into Blender, Maya, Unity, or Unreal for further work on topology, materials, rigging, and performance testing.

Convert image to 3D image

After the user uploads the main image, Cube first separates the objects and generates multiple Spin previews, after which it creates a Preview Mesh. Once the overall shape is confirmed, the Refine option can be used to obtain higher-resolution geometry and textures.

Images with a clear front side, a simple background, a complete subject, and minimal obstruction are easier to generate. The back side, which cannot be seen in a single image, is inferred by the model; therefore, precise products, asymmetric characters, and mechanical structures benefit from using multiple views.

Difference between Turbo Preview and Refine

The Turbo or Preview stage focuses on rapid 3D prediction within a few seconds, and is suitable for selecting compositions and shapes; the Refine stage requires more time and resources to improve the geometric details, textures, lighting effects, and overall realism.

Previewing first and then making refinements can help avoid wasting Credits on incorrect compositions. Refining processes do not guarantee that the result will be ready for production, especially when it comes to joint deformations, thin structures, text, and precise dimensions.

Convert multi-view images to 3D

Multi-View allows the upload of up to 5 images of the same object from different angles, which is suitable for products or characters for which a more accurate representation from all sides is needed. Each image should maintain the same object, proportions, materials, and lighting.

When the focal length, composition, or state of the components in different photos conflict with each other, the model may exhibit double images and distortions. It is advisable to use a fixed lens and a fixed background for shooting, and to manually separate each image as needed.

Manual Segmentation – manual division

When an image contains multiple objects, a complex background, or slender structures, Manual Segmentation can be used to specify the elements that should be retained. An accurate mask helps to prevent the desk, shadows, or surrounding objects from being incorrectly included in the grid.

Text to Image to 3D

In the absence of reference images, it is possible to first create a textual description of the individual elements that will be used for 3D reconstruction, and then feed this description into Image to 3D. The prompt should specify the individual objects, their complete outlines, a simple background, and the ideal viewing angle, in order to avoid generating complex scenes from the outset.

Parts-based component modeling

Complex characters, vehicles, and mechanical objects can be broken down into separate components such as a head, body, weapons, wheels, etc., which are then assembled in 3D software. This component-based approach allows for the retention of more detailed features and makes it easier to replace or reuse asset packs.

Before combining them, it is necessary to standardize the scale, origin, coordinate axes, connection surfaces, and material names. Parts generated by AI do not automatically form a coherent engineering assembly structure.

AI Retexturing – Remapping

AI Retexturing allows for the recreation of a asset’s surface based on text and reference directions; it is useful for altering materials, colors, levels of wear, and style, and it can also be used to repair logos or text that are present in the original image.

Generative textures may still contain spelling errors in brand names or suffer breaks at UV seams. For official assets, it is necessary to check the UV density, normals, PBR channels, and the accuracy of the branding.

Retopology and texture baking

High-quality model representations are not necessarily suitable for real-time engines. Users can redo the topology in subsequent steps, reduce the number of polygons, and bake the details from the high-detail model into the low-detail model.

The character also needs to establish appropriate boundary lines for joint and facial deformations.

Text to 4D and animation

CSM offers a method for creating animations from text; it is possible to upload FBX meshes and then provide one or more descriptions of the actions to be performed in order to generate the animation. Text to 4D focuses on transforming 3D assets into animated content that can change over time.

Motion generation is suitable for rapid prototyping; however, it is still necessary to check the bones, foot sliding, collisions, center of gravity, and cyclic seams. The input mesh must conform to the official reference directions and structure.

3D Agent and Blender MCP

CSM is exploring the use of generative agents to control 3D tools, enabling users to create and modify scenes in environments such as Blender using natural language. The MCP approach allows modeling tools to be utilized by AI agents to carry out multi-step tasks.

The agent has the ability to modify files and scenes; it is necessary to run the copy of the project and set confirmation points. After code is generated automatically, assets are downloaded, or materials are modified in bulk, manual review is required.

Comparison of packages and versions

The official FAQ does not display the current amount in dollars consistently; the price information is loaded dynamically based on the account details. When making a purchase, it is necessary to check the monthly Credits used, as well as the amounts consumed by Preview and Refine functions, the generation speed, privacy settings, API limits, and taxes. It is not advisable to rely on the historical monthly fees mentioned in older reviews.

Differences in licenses between the free version and the paid version

The outputs from Tinkerer are licensed under Creative Commons Attribution 4.0, allowing their use provided that the attribution requirement is met. The outputs from Maker and Creative Pro are proprietary and belong to the customers.

Regardless of the package chosen, users must still have the legitimate rights to input images, characters, products, and trademarks.

Comparison of export formats

Size and Pivot control

The Refine API allows you to set the dimensions of the 3D bounding box as well as the pivot point. The correct scale and origin are essential for positioning, rotating, animating objects, and displaying products in a game.

After generation, the actual unit still needs to be verified in the target software.

CSM REST API

The REST API covers Image to 3D sessions, Preview, Refine, grid addresses, Text to Image, animation, Credits, and Webhooks. Tasks are handled through asynchronous sessions; after submitting a request, one must poll for status or wait for a notification.

The API Key must be stored on the server side. The download link may become invalid over time; therefore, the application should save it promptly after completing a task, and it should handle cases of failure, timeouts, duplicate Webhooks, and insufficient Credits.

Python Library

The official Python library wraps REST APIs; it allows the conversion of images to 3D formats using image URLs, local files, or PIL images, and it handles polling for results while exporting files in OBJ, GLB, FBX, or USDZ formats. It is suitable for batch asset processing and prototype development.

The public documentation of the Python wrapper library does not mean that the model is open-source; calling it still requires the CSM cloud service, API keys, and credits.

Enter suggestions

• Use high-resolution, clear images with a complete subject.
• The characters should preferably adopt a T Pose or a posture with their limbs separated;
• Avoid having transparency, reflectivity, hair, and extremely fine structures dominate the image;
• For complex backgrounds, manual segmentation is used;
• Exact objects provide up to 5 consistent multiple views;
• Complex characters are preferably broken down into components for creation.

Privacy and data

Creative Pro offers options for privacy control, with different rights associated with the free version. Before dealing with unreleased products, customer data, and trade secrets, it is necessary to verify the level of visibility, the ways in which they can be used, the retention period, the methods for deleting them, and the permissions of the team members.

GitHub and the open-source status

CSM’s Cube generation model, web platform, and cloud API are not open-source products. A search on GitHub in real time did not reveal any complete core model repository that is explicitly mentioned on the official website.

The numerous CSM projects with the same name originate from other organizations; they cannot be considered official code from Common Sense Machines.

Even if the Python client or related proxy integrations are made available, it does not mean that the Cube model can be run offline.

CSM AI Usage Guide

Complete a basic task.

1. Clarify the purpose of the image, its dimensions, style, and requirements regarding commercial licensing.
2. Upload the original images for which you have permission to use, or enter clear instructions to CSM AI;
3. Select \"Image to 3D\" to convert the image into a 3D model and generate low-cost candidates first;
4. Use Turbo Preview and Refine to adjust the composition, subject, and details;
5. Zoom in to examine faces, hands, text, edges, logos, and artifacts;
6. Export after confirming the color, resolution, and licensing, while retaining the original image;

Create reusable professional workflows

1. Create a list of brand colors, fonts, layout elements, and elements that are prohibited.
2. Test separately the conversion of Image to 3D images into 3D format, the differences between Turbo Preview and Refine, as well as the conversion of multi-view images into 3D format.
3. Use the same set of representative samples to compare quality, speed, and cost;
4. Complex edges, text, and images of key products should be handed over for manual refinement.
5. Standardize naming, dimensions, and review status;
6. Batch processing and release are carried out after random inspections;

Which users is it suitable for?

• Game developers who can quickly convert concept diagrams into 3D models;
• 3D artists who create prototypes of characters, props, and environments;
• E-commerce teams that need 3D and AR product displays;
• Creators who produce asset packages through component workflows;
• Designers who need AI-based retexturing and text animation;
• Developers who generate content in bulk using REST APIs and Python.

Main advantages

• Supports the generation of text, single images, and up to 5 multiple-view images;
• Preview and Refine allow for controlled management of costs and quality in stages;
• It offers manual segmentation, component modeling, and AI remapping;
• Covers 3D animation and Agent control aspects;
• Supports various mainstream grid formats;
• It provides REST APIs, Webhooks, and Python libraries.

Restrictions and Precautions

• A single image cannot reveal the true structure of objects that are not visible, and errors may still occur regarding precise dimensions, text, thin layers, and complex topologies.
• Refine requires more time and Credits, and the pricing can be viewed by logging in.
• Free distribution requires attribution under CCBY4.0;
• The core model is not open-source, and cloud processing involves privacy concerns regarding the materials used.
• Ultimately, game or product assets still require validation using specialized software.

Frequently Asked Questions

Is CSM AI free?

There is a basic version available under Tinkerer, with content licensed under CC BY 4.0. High-resolution grids, fast servers, and a complete workflow require Maker or Creative Pro.

How many multi-view images can be uploaded at most?

Multi-View currently supports up to 5 images of the same object from different angles.

What is the difference between Preview and Refine?

Preview is used for a quick assessment of the design, while Refine improves the quality of the geometry and textures but requires more resources.

Which export formats are supported?

The Python library explicitly supports OBJ, GLB, FBX, and USDZ; the specific download options available on the web version depend on the current session.

Does CSM provide APIs?

It offers REST APIs, Webhooks, and Python wrapper libraries for tasks such as converting images to 3D format, turning text into images, and creating animations.

Is CSM AI open source?

The core Cube platform and the generation models are not open-source; no complete model weights or server-side source code have been made publicly available by the developers.