Z-Image Turbo Model (CRT AutoDL)

Z-Image Turbo is the model that took over local photorealism in late 2025 - fast, Apache-2.0, runs on hardware people actually own, and it's the base every realism family migrated to. The annoying part is getting the weights: they live on Kijai's Z-Image_comfy_fp8_scaled repo under a filename that isn't the friendliest to guess. ZImageTurbo Model (CRT AutoDL) removes that step entirely: on first run it downloads the fp8-scaled turbo checkpoint into models/diffusion_models and loads it, and you've got a MODEL to sample with.

It's the cleanest way to stand up a Z-Image Turbo workflow from a fresh install, and it pairs with the pack's own ZImageTurboVAE and ZImageTurboCLIP AutoDL nodes - download all three, wire them to a sampler, and you have a working Z-Image pipeline with no manual file hunting.

What it does

A fixed-purpose diffusion-model loader. It checks for z-image-turbo_fp8_scaled_e4m3fn_KJ.safetensors in your diffusion models folder; if it's missing, it streams it down from Kijai's repo with a console progress bar, then loads it through ComfyUI's normal diffusion path and returns MODEL.

The three inputs are the family-standard loader controls:

• patch_cublaslinear - toggles Comfy's cublas_ops fast flag. Off by default, leave it off unless you're chasing specific matmul behavior.
• sage_attention - the one that deserves your attention. The dropdown lists eight modes, but the shipped code only implements disabled and auto. auto swaps in sageattention's kernel if the package is installed; any of the named kernel modes in the list raise an error. So: auto if you've installed sageattention, otherwise disabled.
• enable_fp16_accumulation - on by default; fine to leave on.

That's the entire surface. No seed, no scheduler - just "give me the model," which is exactly what an AutoDL loader should be.

Where it fits

Z-Image Turbo's sweet spot is batch work - the CRT pack's KSampler Batch node was built around rendering Z-Image variations in one pass, and the author's own demo posts push exactly that combination. Wire this model into the batch sampler, feed it prompts via the Dynamic Prompt Scheduler, and you've got a variation farm. The KB's photorealism notes add the caveat that Z-Image Turbo's ceiling is LoRA-stacking headroom (about two LoRAs at ≤1.4 strength before things burn), so this is the base for that stack too.

Install and gotchas

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

or ComfyUI Manager → search "CRT-Nodes". The gotchas are the download (multi-GB on first run - it prints progress to the console, don't assume it hung) and the sage_attention trap described above. Also worth knowing: fp8-scaled here means the weights are fp8 but scaled for quality, which is the standard way Z-Image ships to Comfy - it's not a heavy compromise build. And this node loads only the transformer; the VAE and Qwen text encoder come from the sibling AutoDL nodes, so grabbing all three in one go is the intended flow.