⚗️🔬| Load Z-Image Model (safetensors / gguf)

Z-Image is two model files plus a VAE, and this node loads the big one: the 6B diffusion model itself. Load Z-Image Model (safetensors / gguf) is the experimental loader that handles both formats, with the same GGUF-detection trick as its sibling text-encoder loader. Pick a checkpoint from the dropdown, get a MODEL out, feed it to a sampler. That's the entire contract.

If that sounds thin, here's why it exists. Z-Image checkpoints ship in several flavors - full BF16 (12.3 GB), Int8-ConvRot (6.17 GB), GGUF Q5 (5.19 GB), and FP8 - and they're not all loadable with the same generic node. The author's README is pretty opinionated about the trade-offs, and it's worth trusting: Int8-ConvRot is his current recommendation for "best balance of speed, quality, and compatibility for older hardware," GGUF has the best size-to-quality ratio but ComfyUI's support is "non-native and problematic" (the author's phrase, and not entirely wrong), and FP8 is "NOT RECOMMENDED, use only as a last resort" because most FP8 Z-Image files use naive truncation instead of proper scaling.

How it works

Like the Qwen loader, it lists files from your models/diffusion_models/ folder and extends the list with .gguf files when city96's ComfyUI-GGUF is installed. The file_filter toggle does the same job as in the Qwen node: off shows compatible Z-Image files, on shows everything in the folder including files that may not work. It loads the selected checkpoint as a diffusion model and returns MODEL.

Inputs & outputs

• checkpoint (COMBO) - your Z-Image diffusion file.
• file_filter (BOOL) - compatible-only list vs everything.
• Output: MODEL - into a sampler (or the pack's ComfyKitchenAttentionEnabler before one).

Install

cd ComfyUI/custom_nodes
git clone https://github.com/martin-rizzo/ComfyUI-ZImagePowerNodes

Restart ComfyUI. You still need to download the checkpoint into ComfyUI/models/diffusion_models/ yourself - the recommended ConvRot pick is z_image_turbo_int8_convrot_bf16emixed.safetensors, or z_image_turbo-Q5_K_S.gguf for GGUF. For GGUF to appear, install city96/ComfyUI-GGUF first.

The honest take

You can load Z-Image with the stock UNETLoader/diffusion model loader and save yourself a custom node - the loader doesn't add sampling magic, it adds file-discovery convenience and a filtered dropdown. The real value is the pairing with LoadQwen34bModel, which is the one that genuinely needs help finding GGUF text encoders. Grab both if you're assembling a Z-Image workflow from scratch and want the author's file-layout assumptions baked in; it's experimental, so treat the dropdown behavior as "could change," not gospel. If a checkpoint mysteriously won't load, the usual suspects are the wrong file in the folder, a file_filter hiding it, or a GGUF file with no ComfyUI-GGUF installed.