fairy322/Qwen3-4b-Z-Image-Turbo-AbliteratedV1
The fairy322/Qwen3-4b-Z-Image-Turbo-AbliteratedV1 is a 4 billion parameter Qwen3-based text encoder, derived from Tongyi-MAI/Z-Image-Turbo. It has been specifically modified using a 'heretic method' to significantly reduce image generation and general refusal rates, achieving a 0.0004 KL Divergence and 4/100 refusal rate in torture tests. This model is optimized for generating desired content without censorship, making it suitable for applications requiring unrestricted output.
Loading preview...
Qwen3-4b-Z-Image-Turbo-AbliteratedV1: Unrestricted Content Generation
This model is an "abliterated" version of the Z-Image-Turbo text encoder, built upon the Qwen3-4b architecture. Developed by fairy322, it focuses on minimizing content refusal, particularly for image generation prompts and general conversational restrictions. The modification process involved applying a "p-e-w heretic method" over 1000 trials to target and reduce refusal behaviors.
Key Capabilities
- Significantly Reduced Refusal Rates: Achieves a low refusal rate of 4/100 in rigorous testing, indicating a high propensity to generate requested content.
- Minimal Model Alteration: Maintains a very low KL Divergence of 0.0004, suggesting that the core capabilities of the base model are largely preserved despite the refusal-reduction modifications.
- Versatile Quantization Options: Available in various GGUF formats, ranging from full precision (F16) to highly compressed (Q2_K), allowing for flexible deployment based on performance and resource requirements.
Good For
- Unrestricted Content Creation: Ideal for developers and applications that require a language model to generate diverse content without encountering built-in censorship or refusal mechanisms.
- Creative and Experimental Use Cases: Suitable for scenarios where the primary goal is to explore the full range of generative possibilities without artificial constraints.
- Integration with Image Generation Pipelines: Specifically designed to address refusals in image generation contexts, making it a strong candidate for text-to-image systems where prompt adherence is critical.