theworstgam3r/Qwen3-4b-Z-Image-Turbo-AbliteratedV1
The theworstgam3r/Qwen3-4b-Z-Image-Turbo-AbliteratedV1 is a 4 billion parameter text encoder, based on the Tongyi-MAI/Z-Image-Turbo model, specifically modified to reduce refusal rates in image generation and general responses. Utilizing a p-e-w heretic method applied over 1000 trials, this model achieves a low refusal rate of 4/100 with minimal impact on original model integrity (KL Divergence of 0.0004). It is designed for users seeking a less restrictive model for generating diverse content, particularly for image prompts.
Loading preview...
Overview
This model, Qwen3-4b-Z-Image-Turbo-AbliteratedV1, is a 4 billion parameter text encoder derived from the Tongyi-MAI/Z-Image-Turbo base model. Its primary distinction lies in its "abliterated" nature, meaning it has been specifically modified to significantly reduce content refusal rates, both for image generation prompts and general conversational queries. The modification process involved applying a "p-e-w heretic method" across 1000 trials.
Key Capabilities
- Reduced Refusal Rates: Engineered to minimize instances where the model declines to generate content, achieving a reported refusal rate of only 4/100 in torture tests.
- Minimal Lobotomy: Despite the modifications, the model maintains a very low KL Divergence of 0.0004, indicating that the core capabilities and integrity of the original model are largely preserved.
- Flexible Content Generation: Designed to provide users with greater freedom in generating desired outputs without encountering common content restrictions.
Good For
- Unrestricted Image Prompting: Ideal for users who require a text encoder that is less prone to refusing image generation requests.
- Creative Applications: Suitable for creative tasks where a wider range of outputs is desired without encountering built-in content filters.
- Exploratory Content Creation: Useful for developers and creators looking to push the boundaries of what can be generated by an LLM.