Unchained is a Qwen Image 2.1 LoRA trained on the base model to reduce the restrictions imposed by its learned concepts and improve its handling of a range of popular concepts and anatomical features. It is designed to expand the model's response to prompts while preserving compatibility with the original model's capabilities.

The LoRA works with both the base model and its Turbo variant, and supports image editing workflows. Recommended settings are CFG 5, 25 steps, the ER_SDE sampler, and the Bong Tangent scheduler. These settings provide a good starting point, although results may vary depending on the model variant, prompt structure, and other components in the workflow.

Qwen Image 2.1 is highly sensitive to prompt wording. Attributes such as skin color, lighting, and other visual characteristics can have a much stronger effect than expected. For example, requesting a pink skin part may produce an excessively saturated, unnatural pink rather than a subtle variation. Be precise with color descriptions and avoid unnecessarily strong modifiers when aiming for natural results. Small changes in wording can produce significant differences, so adjust prompts carefully.

All example images include their original prompts and workflows. Download an example to inspect the exact configuration and reproduce the results in your own setup.