Lightweight post-processing node designed for enhancing Qwen3-4B conditioning in Z-Image Turbo workflows within ComfyUI. It refines the text encoder's 2560-dimensional conditioning by improving coherence, detail retention, and adherence to prompts.
- Offers per-token normalization and optional self-attention for improved prompt consistency.
- Features a two-layer MLP refiner with customizable blend strength to adjust detail and sharpness.
- Includes an advanced version with additional controls for literal prompt adherence and detail enhancement.
Context
This tool, known as the Capitan Conditioning Enhancer, serves as a post-processing node specifically for enhancing the conditioning output from the Qwen3-4B text encoder in ComfyUI. Its primary purpose is to refine the generated outputs to improve their visual coherence and adherence to user-defined prompts.
Key Features & Benefits
The Capitan Conditioning Enhancer provides several practical features that significantly improve the quality of generated images. The per-token normalization ensures that each token's contribution is balanced, leading to cleaner and more stable outputs. Additionally, the optional self-attention mechanism allows for better cohesion across distant parts of prompts, enhancing the overall unity of complex scenes.
Advanced Functionalities
The tool includes a two-layer MLP (Multi-Layer Perceptron) refiner that allows users to adjust the hidden layer width, providing flexibility in detail enhancement. The advanced version of the enhancer introduces parameters like detail boost and preserve original, which enable users to fine-tune the balance between high-frequency detail and the stability of the original embeddings. This is particularly useful for achieving a high degree of literal prompt adherence while managing potential noise artifacts.
Practical Benefits
By integrating the Capitan Conditioning Enhancer into their workflow, users can expect improved control over the quality and detail of their generated images. The ability to adjust parameters like enhance strength and normalization contributes to a more efficient workflow, allowing for rapid iterations and fine-tuning of outputs. This tool enhances the user’s ability to create visually appealing images that closely align with their intended prompts.
Credits/Acknowledgments
The Capitan Conditioning Enhancer was developed by capitan01R and is released under the MIT License. The original functionality remains intact, with updates providing additional options for advanced users seeking to maximize detail retention and prompt fidelity.
