This is the multi-page printable view of this section. Click here to print.

Return to the regular view of this page.

CreativeML Open RAIL-M Guide

CreativeML Open RAIL-M is a license used for AI image generation models such as Stable Diffusion. It grants freedom to use the model while including restrictions intended to ensure responsible use.

SPDX Identifier: CreativeML-OpenRAIL-M

Summary of Obligations

A license dedicated to AI models (different from general software)
Use, modification, and distribution of the model: Freely permitted
Notice obligation: Retain the license information
Use-based restrictions: Use for illegal or harmful purposes is prohibited
Derivative models: The same restrictions apply (the RAIL clauses must be retained)
Generated output: Separately licensed (independent of the model license)

AI Model License Characteristics

CreativeML Open RAIL-M differs from general software licenses:

Permissive: Can be used freely, including for commercial purposes
Responsible AI: Certain uses are explicitly prohibited
Separate from output: Images generated by the AI have separate copyright

Consult the OSPO when developing AI services.

What Is CreativeML Open RAIL-M?

CreativeML Open RAIL-M (Responsible AI License - Model) is an AI model license released in 2022 together with Stable Diffusion.

Characteristics of RAIL (Responsible AI License)

Open: Anyone can use it freely
Responsible: Restrictions for responsible use
Permissive: Allows commercial use, like Apache-2.0
Use-Based Restrictions: Restrictions based on use

Differences from General Open Source Licenses

Category	Traditional Licenses (MIT, Apache)	RAIL
Subject	Software code	AI model weights
Use restrictions	None	Yes (harmful use prohibited)
Freedom of use	No restrictions	Responsible use only
Commercial use	Permitted	Permitted (when use restrictions are observed)

Major Projects Using It

Stable Diffusion

Stable Diffusion v1.x: CreativeML Open RAIL-M
Stable Diffusion v2.x: CreativeML Open RAIL++-M (improved version)
Stable Diffusion XL: CreativeML Open RAIL++-M

Other Image Generation Models

Waifu Diffusion
DreamBooth fine-tuned models

What Is Permitted

What You Can Do Freely

Using the model
- Generating images
- Generating images for commercial purposes
- Providing API services
Modifying the model
- Fine-tuning
- Additional training
- Model optimization
Distributing the model
- Redistributing the modified model
- Releasing derivative models
- Integrating into commercial services
Using the output
- Commercial use of generated images
- Selling generated images
- Creating derivative works from generated images

Prohibitions (Use-Based Restrictions)

CreativeML Open RAIL-M may not be used for the following purposes:

1. Violence and Crime

Generating content that promotes or glorifies violence
Assisting in planning or carrying out crimes
Supporting terrorist activities

2. Child Exploitation

Generating child sexual abuse material (CSAM)
Content that sexualizes minors

3. Invasion of Personal Privacy

Generating deepfakes without the consent of the individual
Generating images for the purpose of identity theft
For the purpose of spreading disinformation

4. Discrimination and Hate

Discriminatory content regarding race, gender, religion, etc.
Images that promote hate speech

5. Medical and Legal Advice

For the purpose of professional medical diagnosis
For the purpose of replacing legal advice

6. Other Harmful Uses

Promoting suicide or self-harm
Promoting illegal drugs
Promoting gambling addiction

Use Scenarios

Permitted Uses

1. Generating Marketing Images

Scenario: An image generation service for advertising
Method of use: Generating product images with Stable Diffusion
Determination: Permitted (commercial use OK, not a prohibited use)

2. Generating Game Art

Scenario: Creating background images for an indie game
Method of use: Fine-tuning to generate a consistent style
Determination: Permitted

3. Educational Content

Scenario: Generating thumbnails for online lectures
Method of use: Providing an AI image generation API
Determination: Permitted

Prohibited Uses

1. Deepfake Service

Scenario: A service that generates images using another person’s face
Method of use: Compositing celebrities’ faces
Determination: Prohibited (invasion of privacy)

2. Generating Fake News

Scenario: Generating fake photos of politicians
Method of use: For the purpose of spreading disinformation
Determination: Prohibited (invasion of personal privacy, disinformation)

Requiring Review

Scenario: Profiles of virtual characters generated by AI
Method of use: Profile photos of non-existent people
Determination: Depends on the purpose; review the potential for misuse

Licensing of Derivative Models

An important characteristic of CreativeML Open RAIL-M is the propagation of the RAIL clauses.

When distributing a derivative model:

The same use restrictions must be retained
The license may be changed, but the Use-Based Restrictions are retained
In other words, the “responsible use” clauses continue to apply

Example: When fine-tuning Stable Diffusion and distributing a custom model, the same prohibited uses apply to the custom model as well.

Copyright of AI Output

Important: Generated images are separate from the model license

Model license: CreativeML Open RAIL-M (the model itself)
Output: Separate copyright (owned by the user)

Generated images:

The user owns the copyright (not the model provider)
Can be used freely for commercial purposes
However, the generation process must not violate the prohibited uses