luweigen/qwen-image-2.1-android-q4_0

Qwen Image 2.1 Q4_0 for Android (stable-diffusion.cpp)

Two Qwen Image 2.1 diffusion models merged with LoRAs and quantized to Q4_0 for the Adreno OpenCL backend of stable-diffusion.cpp. They are used by the drawing tab of the Image AI Android app (Snapdragon 8 Elite, about 100–130 s per 512×512 image).

Files

File	Made from	Sampling
qwen-image-2.1-UC-viggle-v0.3-Q4_0.gguf	abenzerps/Qwen-Image-2.1-Uncensored-Quantized qwen-image-2.1-UC-Q4_K_M.gguf (rev 7959e658) + Viggle/Qwen-Image-2.1-viggle-turbo Qwen-Image-2.1-viggle-turbo-v0.3-6step-lora-r128.safetensors (rev 009a44a8)	6 steps, CFG 1, euler, resolution-shifted sigmas (below)
qwen-image-2.1-turbo-UC-Q4_0.gguf	abenzerps/Qwen-Image-2.1-Turbo-Quantized qwen-image-2.1-turbo-Q8_0.gguf (rev 56ae4a3d, from Qwen/Qwen-Image-2.1-Turbo) + abenzerps qwen-image-2.1-uncensored-lora.safetensors (rev 7959e658)	8 steps, CFG 1, euler, --sigmas qwen-image-2.1-turbo

Both were made with stable-diffusion.cpp master-975 (95b3090):

sd-cli -M convert --diffusion-model <base>.gguf --lora-model-dir <lora dir> -p "<lora:<lora name>:1>" --type q4_0 -o <output>.gguf

Other files needed (not in this repository):

• Text encoder: Qwen/Qwen3-VL-8B-Instruct-GGUF Qwen3VL-8B-Instruct-Q4_K_M.gguf
• Decoder: madebyollin/taeqi2_1 taeqi2_1.safetensors (or the full VAE qwen_image_2.1_vae_bf16.safetensors from the abenzerps repository)

Usage

sd-cli --diffusion-model qwen-image-2.1-turbo-UC-Q4_0.gguf --llm Qwen3VL-8B-Instruct-Q4_K_M.gguf --taesd taeqi2_1.safetensors \
  --sigmas qwen-image-2.1-turbo --cfg-scale 1.0 --sampling-method euler -W 512 -H 512 -p "..."

For the Viggle model, pass the Viggle 6-step nodes [1, 0.9375, 0.875, 0.75, 0.5, 0.25] after the resolution shift of the Qwen Image 2.1 scheduler (exp(mu) / (exp(mu) + 1/t - 1), mu linear in (H/16)·(W/16) from 0.5 at 256 to 0.9 at 8192), for example

• 512×512: --sigmas 1.000000,0.962557,0.923057,0.837170,0.631512,0.363570,0.0
• 768×768: --sigmas 1.000000,0.964814,0.927515,0.845775,0.646394,0.378625,0.0

On Adreno OpenCL, add --linear-scale 0.0625 (f16 matrix multiplications overflow otherwise) and --params-backend te=disk to keep the text encoder out of memory while sampling.