Z-Image-Engineer-V6-GGUF

Z-Image-Engineer V6 GGUF

Follow me on X @BennyDaBall_OG !

GGUF quantized release for Z-Image-Engineer V6.

The main repo contains the merged HF safetensors. This repo contains the quant ladder for the ComfyUI-Z-Engineer node, LM Studio, ComfyUI CLIPLoaderGGUF, llama.cpp-style loaders, and local prompt-enhancement workflows.

---

What is this?

Z-Image-Engineer V6 is a SMART DoRA fine-tuned 4B Qwen text encoder from Tongyi-MAI/Z-Image-Turbo.

Use these GGUF files when you want:

• ComfyUI Z-Image text-encoder replacement and in-ComfyUI prompt enhancement through ComfyUI-Z-Engineer (no external server needed)
• LM Studio prompt enhancement
• ComfyUI text-encoder loading through plain CLIPLoaderGGUF
• smaller local files than the merged HF safetensors
• the same V6 prompt style and conditioning behavior in a quantized format

---

Quantization Ladder

Full recursive validation hashes are in HASHES.sha256.

---

Quick Start

LM Studio

Download a GGUF quant, load it, and prompt it directly:

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 (recommended: ComfyUI-Z-Engineer)

1. Install the ComfyUI-Z-Engineer custom node (v2.0+).
2. Place a GGUF file into ComfyUI/models/text_encoders/.
3. Add Z-Engineer CLIP Loader (GGUF) and pick the quant - use the clip output where the stock Z-Image Qwen text encoder would normally go.
4. Optional: add Z-Engineer Prompt Enhancer (Local) with the same clip to rewrite seed prompts in-process, previewed on the node. No LM Studio or external server required.

A ready-made workflow ships with the node repo: example_workflows/z_image_turbo_z_engineer.json. With ComfyUI-GGUF installed the quant stays quantized in VRAM.

Alternative without the node: add a plain CLIPLoaderGGUF node (ComfyUI-GGUF), set model type to lumina2, and use it as the text encoder only.

Verified image settings:

UNET: z_image_turbo_bf16.safetensors
VAE: ae.safetensors
Text Encoder: Z-Image-Engineer-V6-Q8_0.gguf
Resolution: 1024x1024
Steps: 8
CFG: 1.0
Sampler: res_multistep
Scheduler: simple
Shift: 3.0

---

NVFP4 (RTX 50 / Blackwell)

By request! Z-Image-Engineer-V6-NVFP4.gguf is a tuned NVFP4 build, not a straight dump:

• NVFP4 on 216 of 252 block tensors, with the most sensitive ones (attn_v + ffn_down on 18 layers, plus the embeddings) kept at Q6_K.
• Error-minimizing block-scale search during quantization, same standard NVFP4 format.
• Per-tensor 8-bit activation flags, so Blackwell runs it at higher quality by default.

Quality lands right next to Q4_K_M (mean KLD vs F16: 0.058 vs Q4_K_M's 0.051) at the same size, and well ahead of MXFP4 (0.307).

Speed on my 5090:

• LM Studio / llama.cpp: about 290 tok/s vs about 190 tok/s for Q8_0. This is where NVFP4 shines.
• ComfyUI text encoding: same speed as the other quants.
• ComfyUI in-graph prompt enhancing: works, but slower than Q8_0 (about 5 vs 14 tok/s), because ComfyUI-GGUF expands the weights on every token. For fast rewriting inside ComfyUI, use Q8_0 or point the enhancer at LM Studio.

Requirements:

• LM Studio: CUDA 12 llama.cpp runtime 2.56 or newer. Runs on any recent llama.cpp build.
• ComfyUI: needs the gguf Python package 0.19.0 or newer (pip install -U "gguf>=0.19.0" in ComfyUI's Python), then restart ComfyUI.

---

SMART DoRA

V6 was trained with BennyDaBall's SMART DoRA system:

• DoRA for direction/magnitude-separated adapter updates.
• Entropic regularization for less repetition and broader output variety.
• Holographic regularization for cleaner depth-wise feature structure.
• Topological regularization for more coherent latent trajectories.
• Manifold regularization for stable weight behavior during refinement.

The final V6 build used master-corpus SMART DoRA training, retention pressure, SceneClean SFT32 style restoration, AntiRepeat Binary24 refinement, and a 25% style-restoration / 75% anti-repeat DoRA blend.

---

Related Repos

• Main merged HF release: BennyDaBall/Z-Image-Engineer-V6
• ComfyUI custom node (GGUF + shard loaders, local prompt enhancer with preview): ComfyUI-Z-Engineer

---

Acknowledgements

• 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.

Built & trained locally with care by BennyDaBall.

Follow me on X @BennyDaBall_OG !