BennyDaBall/Z-Image-Engineer-V6
BennyDaBall/Z-Image-Engineer-V6 is a 4 billion parameter Qwen text encoder, fine-tuned from Tongyi-MAI/Z-Image-Turbo. It is optimized for dual-role performance as a local prompt-enhancement model and a merged HF text encoder for Z-Image workflows. This model excels at transforming minimal seed prompts into rich, structured visual narratives by adding explicit scene composition, lighting, and texture details. It is designed for private, local workflows within environments like LM Studio and ComfyUI.
Loading preview...
Z-Image-Engineer V6: Enhanced Prompting and Text Encoding
Z-Image-Engineer V6 is a 4 billion parameter Qwen text encoder, fine-tuned from Tongyi-MAI/Z-Image-Turbo by BennyDaBall. This model is engineered for a dual role: enhancing image prompts and serving as a dedicated text encoder within Z-Image workflows. It significantly expands simple prompts into detailed visual narratives, incorporating elements like scene composition, lighting, material textures, and depth separation, while removing generic prompt filler.
Key Capabilities
- Prompt Enhancement: Transforms basic concepts into descriptive, high-fidelity visual prompts locally.
- Text Encoder Swap: Replaces the default Z-Image Qwen text encoder to generate distinct conditioning from the same seed.
- Hybrid Mode: Can rewrite a prompt and then encode the enhanced prompt, effectively guiding the image generation process.
- Private Local Workflow: Designed for use with LM Studio, ComfyUI, and
llama.cpp, ensuring no external API logs or telemetry.
Under the Hood: SMART DoRA Training
V6 utilizes a SMART DoRA (Weight-Decomposed Low-Rank Adaptation) training system. This method allows for precise adapter updates by separating directional and magnitude adjustments. SMART incorporates auxiliary regularizers to prevent repetitive outputs and superficial patterns:
- Entropic: Increases output diversity.
- Holographic: Improves foreground/background hierarchy.
- Topological: Stabilizes coherent latent trajectories.
- Manifold: Regulates overall weight distributions for stable behavior.
The model underwent a multi-stage refinement pipeline, including a base pass, retention pass, supervised refinement (SceneClean SFT32), and anti-repeat binary refinement (AntiRepeat Binary24), culminating in a blended composite architecture.
Integration
Z-Image-Engineer V6 can be used directly in LM Studio for prompt enhancement or integrated into ComfyUI via the ComfyUI-Z-Engineer custom node, which supports both text encoding and local prompt enhancement functionalities.