Kelsey1217/vigenflow-npu-weights

VigenFlow NPU weights

This repository keeps the original FLUX.2-klein-4B, FLUX.1-schnell and Z-Image-Turbo packages in their original folders. The same packages are also available in separate model repositories. Both locations are public and ungated; no Hugging Face login is required to download these three models.

Code, releases and setup: github.com/jia1217/vigenflow. The packages contain weights prepared for AMD Ryzen AI 300-series processors (XDNA2), including BFP16 matrix weights in the layout required by the NPU kernels.

Model	Package in this repository	Separate repository	Size
FLUX.2-klein-4B	flux2_klein_4B	FLUX.2-klein-4B-NPU-BFP16	8.0 GiB
FLUX.1-schnell	flux1_schnell	FLUX.1-schnell-NPU-BFP16	26.6 GiB
Z-Image-Turbo	z_image_turbo	Z-Image-Turbo-NPU-BFP16	11.6 GiB

The restored weight shards and index files match the original packages byte for byte. Each folder includes LICENSE, NOTICE and SHA256SUMS alongside its index and shards. FLUX.2-klein-4B text-to-image and image editing use the same weight package.

Download with VigenFlow

Use VigenFlow v0.2.0-beta.1 or a compatible newer release. The original release configuration that uses Kelsey1217/vigenflow-npu-weights works with the restored folders. No repository change is needed for that configuration.

Run from the release directory:

# Ubuntu: download and unpack one model
./prepare/prepare_base.exe --only flux2_klein_4B

# Windows: download and unpack one model
.\prepare\prepare_base.exe --only flux2_klein_4B

Use flux1_schnell or z_image_turbo for the other models, or omit --only to prepare all enabled models. Existing model IDs and local weight-folder names are unchanged.

If you already switched to the separate repositories, that configuration also continues to work. The optional prepare_models.jsonc supplied here selects those separate repositories. Preserve any custom weights_root setting when replacing your configuration.

For a custom configuration that downloads from this combined repository, set packed_repo to Kelsey1217/vigenflow-npu-weights, packed_revision to main, and use these paths in each model entry:

Model ID	packed_path
flux2_klein_4B	flux2_klein_4B
flux1_schnell	flux1_schnell
z_image_turbo	z_image_turbo

Set packed_repo on each model entry if it already has a per-model repository override. The empty packed_path is used only when downloading from the root of a separate model repository.

Krea 2 Turbo

Krea weights are hosted separately in Kelsey1217/Krea-2-Turbo-NPU-BFP16. Accept the Krea license on its model page and follow its download and unpack instructions. Its agreement gate and separate download process still apply.

Format and licensing

The packages use vgf-npu-weights-1: safetensors shards containing U8 byte vectors named by relative file paths, plus model.safetensors.index.json. Unpacking restores the prepared model folder. These archives are not directly loadable as Diffusers checkpoints.

The FLUX models originate from Black Forest Labs; Z-Image-Turbo originates from Tongyi-MAI. Their converted packages remain under Apache 2.0. Retain the LICENSE and NOTICE supplied in each model folder when redistributing them. These companion files are outside the weight shards. Conversion and packing do not replace the upstream terms.

Krea 2 Turbo is a modified version of Krea.ai, Inc.'s model and uses the separate Krea 2 Community License Agreement, subject to the Krea Acceptable Use Policy. This repository's Apache 2.0 license does not apply to Krea weights. The Krea agreement and required Notice accompany that package. These are independent conversions, not official releases or endorsements by the upstream authors.