Z-Engineer CLIP Loader (Safetensors / Shards)

The published Z-Image-Engineer-V6 release doesn't come as one tidy file. It's three safetensors shards plus an index JSON, the way HuggingFace dumps any big model, and ComfyUI's stock text-encoder loader has no idea what to do with that. This node does. It's the fidelity path in the Z-Engineer pack - full FP16 weights, no quantization - and it's also the one you use when you grab the V6 release directly and want it to just work.

The Z-Engineer CLIP Loader (Safetensors / Shards) loads a single .safetensors or a sharded HuggingFace folder under models/text_encoders as a Z-Image text encoder, and gives you a clip output that feeds CLIP Text Encode exactly like any other text encoder would. Remember what Z-Image's encoder actually is: a Qwen3-4B LLM, not a CLIP from 2021. This loader is what lets that LLM play the role of text encoder in your graph.

How it works

The model_name dropdown walks models/text_encoders and clip and lists both loose files and folders. A folder containing model.safetensors.index.json (or model-0000X-of-0000N files) is recognized as a shard directory, and the loader merges its shards into a single state dict before handing the weights to ComfyUI's normal text-encoder loading. Single-file safetensors skip the merge step. Weights sit on the offload device until first use, so loading doesn't spike VRAM while another job shares the GPU, and the model participates in ComfyUI's model management like any other.

The only required input is model_name; device (default or cpu) is an advanced toggle you can ignore. One output: clip.

Install

The pack is a standard custom-node install - Manager search "ComfyUI Z-Engineer," or:

cd ComfyUI/custom_nodes
git clone https://github.com/BennyDaBall930/ComfyUI-Z-Engineer.git
pip install -r ComfyUI-Z-Engineer/requirements.txt

Restart, then set up the model folder from BennyDaBall/Z-Image-Engineer-V6:

ComfyUI/models/text_encoders/Z-Image-Engineer-V6/
├── model-00001-of-00003.safetensors
├── model-00002-of-00003.safetensors
├── model-00003-of-00003.safetensors
└── model.safetensors.index.json

It shows up in the dropdown as Z-Image-Engineer-V6/. No extra dependencies beyond the pack's own requirements - requests and gguf, which are in requirements.txt. You do need a ComfyUI with native Z-Image support (v0.3.75+).

The one thing to plan around: VRAM

This is the full-fidelity path, and fidelity has a bill. The README's number is about 9 GB during generation for the FP16 shards. If that fits your card, take it - you never have to think about quant artifacts in the encoder. If it doesn't, the GGUF loader is the same model at 3–4 GB for a Q4_K_M, and the quant lives in the prompt encoder, not the part of the pipeline that's sensitive to it.

Where people get burned

• Partial downloads. The loader errors out when a shard referenced by the index is missing. Make sure all three shards and model.safetensors.index.json are in the folder before you queue.
• Wrong location. The dropdown only sees models/text_encoders and clip. A folder sitting in models/checkpoints is invisible, and an empty dropdown means the file isn't where this node looks.
• "CLIP doesn't support text generation." That one comes from the enhancer, not this loader - but it means you wired a stock Z-Image CLIP into the local enhancer instead of one loaded through this pack's loaders. Load it here, then share that same clip output.

Wire clip into CLIP Text Encode for the Z-Image Turbo pipeline. Add the Z-Engineer Prompt Enhancer (Local) on the same clip and it doubles as your prompt writer.