Qwen Image 2.1 Anime FUSAL GGUF runtime
Built with Qwen. This repository packages a tested LocalAI and stable-diffusion.cpp runtime for MicksHF/Qwen-Image-2.1-Anime-Fusal. It is a multi-file runtime assembly, not a merged checkpoint.
License and use restrictions
The denoiser, VAE, and FUSAL adapter are governed by the Qwen Research License Agreement and are limited to non-commercial research and evaluation unless Qwen grants separate authorization. The Heretic encoder is distributed under Apache 2.0. See LICENSES/README.md, NOTICE, and MODEL_PROVENANCE.json.
The source model includes NSFW training material and is not for all audiences. This repository grants no rights to third-party artwork, characters, trademarks, datasets, or generated outputs. Users are responsible for lawful and responsible use.
Components
| File | Role | Quantization |
|---|---|---|
qwen_image_2.1-Q5_0.gguf | Qwen Image 2.1 diffusion transformer | Q5_0 |
qwen3-vl-8b-heretic-1.3.0-Q4_K_M.gguf | Heretic Qwen3-VL text encoder | Q4_K_M |
qwen_image_2.1_vae_bf16.safetensors | Qwen Image 2.1 VAE | BF16 |
fusal-qwen-image-2.1-stack-exact-r80.safetensors | Exact FUSAL adapter | rank 80, BF16 |
The text encoder does not include vision weights. This package supports text-to-image. It does not support reference-image editing without a compatible Qwen3-VL vision projector or combined encoder.
LocalAI
Use qwen-image-2.1-anime-fusal.yaml or install the matching LocalAI gallery entry.
Prompt format:
fusal style. [TASK: TEXT_TO_IMAGE] your description<lora:fusal-qwen-image-2.1-stack-exact-r80:1.0>
Use exactly one FUSAL adapter at strength 1.0. Recommended settings are 30–40 steps, Euler sampling, and CFG 1.0. Turbo is not supported by this adapter.
Example request:
curl http://127.0.0.1:8080/v1/images/generations \
-H 'Content-Type: application/json' \
-d '{
"model": "qwen-image-2.1-anime-fusal",
"prompt": "fusal style. [TASK: TEXT_TO_IMAGE] An anime character in a snowy forest.<lora:fusal-qwen-image-2.1-stack-exact-r80:1.0>",
"size": "1024x1024",
"step": 30
}'
Validation
Validated on CPU with stable-diffusion.cpp commit 2988060a4338a5dcd94721c964a8e87003c1ac8b and LocalAI. The direct test applied all 448 / 448 LoRA tensors and generated a valid PNG. LocalAI requires a stable-diffusion.cpp revision containing fix 510bccf330a424cca742ca8084d78f4f577c9336 or later for complete fused-MLP LoRA mapping.
The quantized encoder is practical but is not numerically identical to the BF16 encoder used to cache the source model's training conditioning.
Provenance
All weights are byte-identical to the pinned upstream artifacts listed in MODEL_PROVENANCE.json. Verify them with:
sha256sum -c SHA256SUMS