DreamShaper 8 for Snapdragon NPU
DreamShaper 8 for Snapdragon chips
2 min read
This model is a version of the popular DreamShaper 8 image generator (a Stable Diffusion 1.5 fine-tune) optimized specifically to run on neural processing units (NPUs) of mobile platforms. The model files are converted into binary QNN context binaries for the Snapdragon 8 Gen 2 chipset, enabling image generation directly on mobile hardware. The model is aimed at mobile app developers and enthusiasts who want to run AI models locally on smartphones without relying on cloud servers.
Key features
- NPU optimization: files are adapted for the Hexagon NPU of the Snapdragon 8 Gen 2 processor (SM8550).
- Weight quantization: UNet weights are quantized to 8 bits (while keeping 12 sensitive convolutions in 16 bits), and activations and the embedding table are converted to 16-bit format.
- Fixed resolution: the model operates strictly at a 512 × 512 pixel resolution.
- Separate components: the model is split into three binary files (text encoder, UNet, and VAE decoder).
What you can do
- Integrate into a mobile app: run image generation inside an app based on the QNN engine or use it in the LATAS Android app.
- Test local inference: run image generation on a mobile device without an internet connection or cloud GPUs.
Example
The model consists of three files: text_encoder.bin, unet.bin, and vae.bin, which process tokens, latent space, and the image in 16-bit fixed point format via the QNN runtime.
Read more
DetailsRelative UNet error compared to fp32 is 0.9–2.7%, average text encoder similarity is 0.991, and VAE decoder quality reaches 53–56 dB PSNR.
Requires the Qualcomm QNN runtime and a Snapdragon 8 Gen 2 chipset (Hexagon v73).
Distributed under the CreativeML Open RAIL-M license (like the base DreamShaper 8).










Comments