Hello,

Here's the BulkedUp AI R2 SD 1.5 LoRA merged with Fluffusion at 1 strength as a QNN model for Local Dream, an Android Stable Diffusion app with Snapdragon NPU acceleration:
https://github.com/xororz/local-dream

The model was converted to an NPU model with the Github's conversion guide:
https://github.com/xororz/local-dream/tree/master/convert

There are three versions of the QNN model that you may choose:
- 8gen2 for Snapdragon 8 gen 2 processors and up
- 8gen1 for Snapdragon 8 gen 1 processors
- min for Non-flagship Snapdragon chips with Hexagon V68 or above (e.g., Snapdragon 7 Gen 1, 8s Gen 3) (NOTE: this can only gen in 512x512)


With my Snapdragon 8 gen 2 phone, the S23 Ultra, here's some quick benchmarks
512x512 Image gen with NPU (8gen2 version)
- around 6 seconds

832x832 Image gen with NPU (8gen2 version)

- around 39 seconds

512x512 Image gen with GPU
- around 1 minute and 10 seconds

512x512 Image gen with CPU
- around 2 minutes and 50 seconds

The 8gen1 and 8gen2 were converted to support the following resolutions:

512x512, 704x576, 576x704, 640x640, 576x832, 832x576, 704x704, 832x640, 640x832, 832x704, 704x832, 832x832