Z-Image-Engineer V6 開源
- URL: https://www.facebook.com/share/p/192JmpGLQt/?mibextid=wwXIfr
- Date Saved: 2026-06-11
- Source: Facebook (Tonyaigirls page)
- Tags: creative-ai, ai-tools
Summary
Tonyaigirls Facebook page announced the open source release of Z-Image-Engineer V6 (skipping V5, jumping directly from V4 to V6).
What is Z-Image-Engineer V6?
A SMART DoRA fine-tuned 4B Qwen text encoder based on Tongyi-MAI/Z-Image-Turbo. It’s designed for prompt engineering/enhancement for Z-Image Turbo image generation workflows.
Key details:
- Created by: BennyDaBall (HuggingFace/GitHub)
- Base model: Tongyi-MAI/Z-Image-Turbo (Qwen text encoder)
- License: Apache 2.0
- Released: 2026-06-06
- Downloads: ~916 (GGUF) / ~16K (V4 was popular)
Training Method: SMART DoRA
- DoRA (direction/magnitude-separated adapter updates)
- Entropic regularization (less repetition)
- Holographic regularization (cleaner features)
- Topological regularization (coherent latent trajectories)
- Manifold regularization (stable weight behavior)
- Final blend: 25% style-restoration / 75% anti-repeat DoRA
Use Cases
- LM Studio — load GGUF, use as prompt enhancer directly
- ComfyUI — replace stock Z-Image text encoder via CLIPLoaderGGUF node (set model type to
lumina2) - llama.cpp — standard GGUF loading
Quantization Options
- F16 (7.5 GiB) down to Q3_K_M (1.9 GiB)
- Recommended daily-driver: Q5_K_M (2.7 GiB)