README.md

---
license: apache-2.0
language:
  - en
base_model:
  - Tongyi-MAI/Z-Image-Turbo
library_name: transformers
pipeline_tag: text-generation
tags:
  - text-generation
  - prompt-engineering
  - image-generation
  - z-image
  - z-image-turbo
  - qwen3
  - text-encoder
  - comfyui
  - lm-studio
  - conversational
---

# Z-Image-Engineer V6 (4B)

**Follow me on X [@BennyDaBall_OG](https://x.com/BennyDaBall_OG) !**

## Model Metadata

| Key | Value |
|---|---|
| **License** | Apache-2.0 |
| **Language** | English (`en`) |
| **Base Model** | `Tongyi-MAI/Z-Image-Turbo` |
| **Library** | `transformers` |
| **Pipeline Tag** | `text-generation` |
| **Format** | HF Safetensors |

---

The **Z-Engineer** returns, fully rebuilt around the **SMART DoRA** training system for Z-Image Turbo.

Yes, we jump from V4 to V6. Unlike the usual guy math, this one actually brought the extra two inches.

**Z-Image-Engineer V6** is a fine-tuned 4B Qwen text encoder (`Tongyi-MAI/Z-Image-Turbo`) optimized for dual-role performance: a local prompt-enhancement model and a merged HF text encoder for Z-Image workflows. The [ComfyUI-Z-Engineer](https://github.com/BennyDaBall930/ComfyUI-Z-Engineer) node runs both roles fully inside ComfyUI from this release.

![Z-Image-Engineer V6 simple A/B with rewrites](evidence/gallery_z_image_engineer_v6_simple_ab_with_rewrites_CONTACT.png)

---

## What is Z-Image-Engineer V6?

V6 transforms minimal seed prompts into rich, highly structured visual narratives. It adds explicit scene composition, lighting direction, material texture, and depth separation while stripping out empty prompt sludge like *"8k, masterpiece, trending on ArtStation."*

It can also be used directly as a Z-Image text encoder. This repo contains the merged HF safetensors. The GGUF quantized release lives in the companion repo: [Z-Image-Engineer-V6-GGUF](https://huggingface.co/BennyDaBall/Z-Image-Engineer-V6-GGUF).

### Key Use Cases

- **Prompt Enhancement:** Upgrade simple concepts into descriptive, high-fidelity visual prompts locally.
- **Text Encoder Swap:** Replace the stock Z-Image Qwen text encoder to generate different conditioning from the same seed.
- **Hybrid Mode:** Use V6 to rewrite your prompt, then use V6 again to encode it. It writes the scene and drives the image model.
- **Private Local Workflow:** Built for LM Studio, ComfyUI, and `llama.cpp`. No API logs, no external telemetry.

---

## Under the Hood: SMART DoRA

V4 pioneered SMART training. V6 adapts that system into a **Weight-Decomposed Low-Rank Adaptation (DoRA)** framework.

DoRA provides surgical adapter updates by decoupling directional and magnitude adjustments. SMART adds auxiliary pressure so the model does not collapse into repetitive prompt loops or superficial sentence patterns.

| Regularizer | What it Does | Why it Matters |
|---|---|---|
| **Entropic** | Broadens output probability diversity. | Reduces repetitive loops and generic vocabulary. |
| **Holographic** | Enforces structured, depth-wise feature logic. | Improves foreground/background hierarchy. |
| **Topological** | Stabilizes coherent latent trajectories. | Keeps prompts flowing naturally instead of stalling out. |
| **Manifold** | Regulates overall weight distributions. | Keeps model behavior stable under high-pressure refinement. |

### The Refinement Pipeline

V6 was not a simple one-and-done training run. The final architecture is a blended composite:

1. **Base Pass:** Master-corpus SMART DoRA training on the native Z-Image Turbo text encoder.
2. **Retention Pass:** Preservation pressure for numbers, color accuracy, text signage, named objects, actions, and spatial tracking.
3. **SceneClean SFT32:** Supervised refinement to restore the cinematic V4/base-V6 voice.
4. **AntiRepeat Binary24:** Binary anti-repeat refinement to reduce loops, abrupt fragments, and bad endings.
5. **Final Blend:** A 25% style-restoration / 75% anti-repeat DoRA adapter blend, balancing vivid descriptions with tighter syntax.

---

## Quick Start

### LM Studio: Prompt Enhancement

Use this merged HF release directly where supported, or download a GGUF quant from [Z-Image-Engineer-V6-GGUF](https://huggingface.co/BennyDaBall/Z-Image-Engineer-V6-GGUF) for LM Studio. No complex system prompt is required.

```text
Enhance this image prompt for Z-Image Turbo: a unicorn
```

The comparison examples were generated from direct LM Studio user requests like this, with no separate system prompt. `V6_SYSTEM_PROMPT.md` is included only as an optional preset for people who want a stricter prompt-only chat setup.

### ComfyUI: Text Encoder + Local Prompt Enhancer

Use the [ComfyUI-Z-Engineer](https://github.com/BennyDaBall930/ComfyUI-Z-Engineer) custom node (v2.0+). It loads this repo's sharded safetensors release directly and runs V6 as both the Z-Image text encoder and an in-ComfyUI prompt enhancer - no LM Studio or external server required.

1. Download this repo into `ComfyUI/models/text_encoders/Z-Image-Engineer-V6/` (the three `model-0000X-of-00003.safetensors` shards plus `model.safetensors.index.json`).
2. Add **Z-Engineer CLIP Loader (Safetensors / Shards)** and pick `Z-Image-Engineer-V6/` from the dropdown.
3. Wire `clip` into your Z-Image **CLIP Text Encode** - V6 replaces the stock Qwen text encoder.
4. Optional: add **Z-Engineer Prompt Enhancer (Local)** with the same `clip` to rewrite seed prompts in-process; the enhanced prompt is previewed right on the node.

A ready-made workflow ships with the node repo: `example_workflows/z_image_turbo_z_engineer.json`.

Prefer smaller files? Use a quant from [Z-Image-Engineer-V6-GGUF](https://huggingface.co/BennyDaBall/Z-Image-Engineer-V6-GGUF) with the node's **Z-Engineer CLIP Loader (GGUF)** instead.

### Verified Image Settings

```text
UNET: z_image_turbo_bf16.safetensors
VAE: ae.safetensors
Text Encoder: Z-Image-Engineer-V6 (this repo's sharded safetensors, or a GGUF quant)
Resolution: 1024x1024
Steps: 8
CFG: 1.0
Sampler: res_multistep
Scheduler: simple
Shift: 3.0
```

---

## Training Specifics

| Parameter | Specification |
|---|---|
| **Base Text Encoder** | `Tongyi-MAI/Z-Image-Turbo/text_encoder` |
| **Tokenizer** | `Tongyi-MAI/Z-Image-Turbo/tokenizer` |
| **Method** | SMART DoRA / PEFT Adapter Training |
| **Rank / Alpha / Dropout** | 64 / 64 / 0.03 |
| **Target Modules** | `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `down_proj`, `up_proj` |
| **Refinement Stack** | Supervised Style SFT + Binary Anti-Repeat |
| **Final Packaging** | Merged HF safetensors |

---

## GGUF Quantization Ladder

The quantized release is separate on purpose:

[BennyDaBall/Z-Image-Engineer-V6-GGUF](https://huggingface.co/BennyDaBall/Z-Image-Engineer-V6-GGUF)

That repo contains the full GGUF ladder: F16, Q8_0, Q6_K, Q5_K_M, Q4_K_M, Q3_K_M, and MXFP4.

---

## Verification & Proof

The bundled comparison image is:

```text
evidence/gallery_z_image_engineer_v6_simple_ab_with_rewrites_CONTACT.png
```

It compares foundational prompts across four isolated control paths:

1. Stock Encoder + Raw Prompt
2. V6 Encoder + Raw Prompt
3. Stock Encoder + V6 LM Studio Rewrite
4. V6 Encoder + V6 LM Studio Rewrite

---

## Disclaimer & Acknowledgements

This model is a prompt engineer and text encoder. Diffusion is still diffusion; structural expansion improves compositional adherence, but it does not mathematically guarantee a perfect seed every single time. Use creative judgment locally.

- **Tongyi-MAI** for the Z-Image Turbo ecosystem.
- **Qwen** for the adaptable text encoder backbone.
- The open-source maintainers behind **LM Studio**, **ComfyUI**, **llama.cpp**, **PEFT**, and **Transformers**.
- My local power utility provider, for sustaining the research grid.

**Built & trained locally with care by BennyDaBall.**

**Follow me on X [@BennyDaBall_OG](https://x.com/BennyDaBall_OG) !**