Z-Image Turbo CLIP (CRT AutoDL)

Z-Image Turbo is Alibaba's 6B distilled image model - the one that, as the KB's ecosystem guide puts it, "killed Flux 2 Dev's hype" at launch and became the SDXL successor on consumer cards. It's small, fast, and its text encoder is a 4B-class Qwen3 model, which is the part that trips people up: it's not a classic CLIP you can grab anywhere, it's a specific repackaged file in a specific spot.

This node solves exactly that. It auto-downloads qwen_3_4b_fp8_mixed.safetensors from Comfy-Org's z_image_turbo repo into models/text_encoders/, then loads it as a CLIP configured for the Lumina2 clip type Z-Image expects. No inputs, one CLIP output. It's the "stop hunting for the right file" node.

How it works

First run: it checks if the file exists, downloads it with a progress readout if not, and hands back a ready-to-encode CLIP. Wire that into a CLIP Text Encode, pair it with the model and VAE (which the pack's AutoDL set also covers - ZImageTurbo Model and ZImageTurbo VAE nodes), and you've got a complete Z-Image Turbo stack where every piece auto-downloaded itself into the right folder.

The fp8 quantization matters here. Qwen3-4B in fp16 is chunky; fp8 keeps the text encoder light enough that the whole 6B model stays friendly on the 6–12GB cards Z-Image Turbo is famous for running on. The KB notes 3060-class cards generating in ~30 seconds - that only holds if the text encoder isn't the bottleneck.

Installing it

Standard CRT-Nodes install: ComfyUI Manager → CRT-Nodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes.git
pip install -r requirements.txt

Restart, drop the node in, let the first run download the encoder. The Z-Image pieces are registered under the pack's AutoDL category, so look for ZIMAGETURBO in the menu.

Where people get burned

The classic mistake is grabbing the wrong text encoder - the pack's other AutoDL families use different CLIPs (Krea 2 uses its own Qwen3-VL, FLUX.2 uses Qwen3-8B), and they're not interchangeable with Z-Image's. Use this node's dedicated loader, not a random one from another pack. Also note this only loads the text side: you still need the model and the VAE to actually generate, and Z-Image Turbo wants a flow-match / distilled sampler setup at low steps and CFG 1 - treat the CLIP as one piece of a known-good workflow, not a shortcut past the rest of it.