X-Ray_Alpha: A Fully Uncensored Vision Model
X-Ray_Alpha, developed by SicariusSicariiStuff, is a pre-alpha proof-of-concept vision model built upon Gemma-3 4B instruct. It distinguishes itself as one of the very few truly uncensored and fully trained vision models available, addressing the limitations of other models that often apply censorship or only fine-tune their text components.
Key Capabilities
- Fully Uncensored Vision: Unlike many existing models, X-Ray_Alpha has been trained without content moderation, allowing users to classify and analyze a wide range of images without corporate-imposed restrictions.
- In-depth Descriptions: The model generates very detailed and long descriptions for images, providing comprehensive insights.
- Foundation for Open-Source AI: It represents a critical step towards democratizing vision capabilities, particularly for tasks like mass image tagging essential for training LORAs and pretraining image diffusion models.
- Nuanced Content Moderation: Enables users to define their own content moderation and classification rules, especially for sensitive topics like art with nudity, where stock models often refuse to inference.
- Good Roleplay & Writing: The text portion of the model, while somewhat uncensored, was trained on a massive corpus of high-quality human (~60%) and synthetic data, contributing to strong roleplay and writing abilities.
Good for
- Image Tagging: Ideal for creating well-tagged datasets for training LORAs and pretraining image diffusion models.
- Custom Content Classification: Users requiring flexible and uncensored image analysis for diverse use cases, including art classification or specific content moderation needs.
- Proof-of-Concept Exploration: Developers interested in contributing to or exploring the frontiers of uncensored vision AI.
This model is currently a proof-of-concept and requires further community assistance, particularly with well-tagged, diverse image data, to enhance its accuracy and power. Instructions for running inference are provided, requiring approximately 15.9 GB VRAM for FP16.