CC0_rebuild_attempt
Version: 0.3
Summary
CC0_rebuild_attempt is a text-to-image model based on the Stable Diffusion 1.5 architecture. It is trained exclusively on CC0 images and other permissive content, aiming to produce high-quality artistic images from given text prompts. The goal is to create a robust and versatile model while ensuring the dataset used is entirely within the public domain, allowing for unrestricted use.
A mixed technic was used to create te capitons of images, the datatset was segmented my subject and for each it was used a different method such as: GIT for realistic and photo images and JointTagger for illustrations, Gemini for regularization and finaly human correction.
Training Overview
- Input: Manual captioned images at 768x
- Output: Images
- Architecture: Stable Diffusion 1.5
Performance Limitations
CC0_rebuild_attempt may face challenges in generating highly detailed or realistic images due to the constraints of the CC0 and permissive content datasets. Additionally, the model may underperform in specific domains where high-quality, diverse CC0 images are less prevalent.
Training Dataset Limitations
The model is trained on images and content from the following sources:
- Pexels: Pexels License
- LIBRESHOT: CC0
- Unsplash: Lite Dataset License
- opengameart.org: CC0
- Authors: CC0
- Contributors: CC0
- Met Museum Open Access CC0
Dataset sample:
Although the dataset consists of CC0 and permissive content, in respect of individual site policies and the creators' preferences, only images that explicitly allow redistribution will be published. The remaining images will be indexed but not redistributed.
These datasets may not cover all possible themes or subjects comprehensively. The datasets may lack representation of certain modern or niche topics due to the limited availability of such content under these licenses. Additionally, the model was trained and developed with a focus on compliance with the Brazilian Copyright Act, which imposes stricter regulations compared to other jurisdictions due to the absence of fair use provisions.
It is important to note that while every effort has been made to ensure the model generates ethical and high-quality content, it is not possible to guarantee that the model will always avoid producing unwanted content or achieve the highest quality in every instance. This project represents an attempt to create an ethical model within the constraints of the available datasets and legal considerations.
Copyright laws differ from country to country, and this project acknowledges the necessity of establishing guidelines for considering public domain content. It is hoped that this research will inspire others to build models with greater responsibility, taking into account the complexities of copyright laws and the ethical use of training data.
Contact before use in commercial projects.
Associated Risks
- The model might struggle with generating highly detailed text within images.
- There may be limitations in creating complex scenes that require deep compositional understanding.
- The quality and diversity of generated images are dependent on the availability and variety of CC0 and permissive content.
- Potential bias towards subjects and styles that are more commonly found in CC0 and permissive content datasets and the manual capition.
- Please check OpenRAIL and make responsible use of open framework
Intended Uses
- Generative art and design projects
- Educational tools and research in generative AI
- Creative experimentation and artistic expression
- Reference for ethical development
Showcase
Authors' Note
I firmly believe that intellectual property laws have often been utilized as means of controlling and restricting access to content, effectively acting as a form of censorship. When companies use unauthorized content, nothing happens, but when individuals do, the consequences might be much more severe. Many massive models used, and continue to use, unauthorized content, and this is unlikely to change. It's a sad reality about what happened and will be happening. However, this is the first time these technologies have been open source and widely available for everyone. Neither governments nor companies will decide the fate of society, but work and actions of individual people. We might have to change how we think about what is considered original. This shift in perspective is crucial as we navigate the evolving landscape of AI and copyright. I hope this work inspires more responsible and ethical development practices in the field of generative AI.