unsloth/Z-Image-Turbo-ComfyUI

Z-Image Turbo VAE + Qwen3-4B conditioner, single file (mirror)

The two single-file Z-Image assets that stable-diffusion.cpp needs, mirrored from Comfy-Org/z_image_turbo.

Unsloth Studio fetches these on the no-GPU (native sd.cpp) path: both for Z-Image, and the Qwen3-4B conditioner again for FLUX.2 klein 4B, which shares it. They used to come straight from the community repack; a rename or a takedown there turns every CPU Z-Image and klein-4B load into a 404, so they are mirrored here. The weights are unmodified: byte for byte the files of the same name in the source repo.

File	Size	Upstream model	Licence
split_files/vae/ae.safetensors	0.31 GiB	the FLUX.1 autoencoder, which Tongyi-MAI/Z-Image-Turbo reuses	Apache-2.0
split_files/text_encoders/qwen_3_4b.safetensors	7.49 GiB	Z-Image's Qwen3-4B conditioner	Apache-2.0

Only these two files are mirrored. The source repo also ships the Z-Image diffusion models, a distill LoRA and quantised conditioners; Studio does not use them from there, so they are not here.

Licence

Apache-2.0 for both files. The source repack carries no licence metadata at all, so the terms here come from upstream rather than from the repacker.

Worth naming precisely, because the filename hides it: ae.safetensors is the FLUX.1 autoencoder, not a Tongyi-trained VAE. The file's own safetensors metadata says modelspec.author = "Black Forest Labs" and modelspec.license = "Apache License 2.0", and this mirror is byte-identical, so that metadata is still in there. Z-Image reuses it, and Black Forest Labs ships the same autoencoder in the Apache-2.0 black-forest-labs/FLUX.1-schnell. The conditioner derives from Apache-2.0 Qwen3-4B.

Full text in LICENSE, attribution and the statement of what the repack changed in NOTICE. Not an official Alibaba Group, Black Forest Labs or ComfyUI product, and not endorsed by any of them.