SDXL NPU Conversion (Experimental)

This guide converts a Stable Diffusion XL .safetensors checkpoint into an NPU model for evaluation on supported Snapdragon devices.

WARNING

SDXL conversion is experimental and substantially heavier than SD1.5 conversion. High-resolution SDXL workflows can require 100 GB+ disk space and 128 GB of RAM + swap. Read the prerequisites carefully before starting.

Prerequisites

• OS: Linux or WSL
• Qualcomm AI Engine Direct SDK 2.28 — please use v2.28 to avoid potential issues. Download v2.28.0.241029.zip.
• uv — Python environment manager
• RAM + swap: ~128 GB for high-resolution conversion

If you have a CUDA-capable GPU, you can switch the torch dependency in pyproject.toml to the GPU build to speed up data preparation.

Conversion Scripts

INFO

Download the script: convertsdxl.zip.

Environment Setup

cd convertsdxl
uv venv -p 3.10.17
source .venv/bin/activate
uv sync

Set the QNN_SDK_ROOT path inside convert_all_sdxl.sh to your extracted QNN SDK.

Example: export_sdxl.sh

set -e

model_path=~/Downloads/anythingxl.safetensors
model_name=anythingxl
realistic=false                 # set true for realistic-style prompts
scheduler=dpm                   # dpm | lcm | eulera
cfg=5,7                         # 5–7 random per image
steps=15,30                     # 15–30 random per image

# Currently only 8gen3 is built. The 8gen3 build runs on 8 Gen 3, 8 Elite, and 8 Elite Gen 5 / 8 Gen 5.
soc_versions=("8gen3")

uv venv -p 3.10.17 --clear
source .venv/bin/activate
uv sync

realistic_flag=""
if [ "$realistic" = true ]; then
    realistic_flag="--realistic"
fi

# ===== 1024×1024 =====
echo "Processing base resolution: 1024x1024"
python prepare_data_sdxl.py --model_path $model_path $realistic_flag --scheduler $scheduler --cfg $cfg --step $steps
python gen_quant_data_sdxl.py
python export_onnx_sdxl.py --model_path $model_path

for soc in "${soc_versions[@]}"; do
    bash scripts/convert_all_sdxl.sh --min_soc $soc
done

# ===== Package =====
echo "Packaging output files..."
for soc in "${soc_versions[@]}"; do
    touch output/qnn_models_sdxl_${soc}/SDXL
    zip -r ${model_name}_qnn2.28_${soc}.zip output/qnn_models_sdxl_${soc}
done

Key Parameters

Parameter	Notes
scheduler	dpm, lcm, or eulera. Pick one that matches the model's recommended sampler.
cfg	A range used to sample CFG values during quantization data generation. Wider ranges produce more general models but increase quantization noise.
steps	Same idea as cfg — a range that determines how varied the calibration data is.
soc_versions	Currently only 8gen3 is supported. The same binary runs on 8 Gen 3, 8 Elite, and 8 Elite Gen 5 / 8 Gen 5.

Output

You'll get a single zip per SOC tier:

anythingxl_qnn2.28_8gen3.zip

Import it in the app the same way as SD1.5 NPU models. See Custom Models.

Caveats

• Expect generation quality slightly below the source SDXL checkpoint — quantization noise is more visible at SDXL scale than at SD1.5.
• The SDXL marker file inside the output is required so the app loads it as an SDXL model rather than SD1.5.
• Conversion takes considerably longer than SD1.5 — plan for an overnight run on a workstation.