latentdivergence/lcm-ssd1b-npu-8gen3

lcm-ssd1b-npu-8elite

LCM-SSD-1B compiled for the Qualcomm Hexagon v79 NPU (Snapdragon 8 Elite), as used by Latent Studio. UNet and VAE decoder only - the text encoders and tokenizer are shared with sdxl-lightning-4step-int8.

Measured on a real Snapdragon 8 Gen 3: 58 s end to end for a 1024px image (versus 5 m 40 s for SDXL-Lightning on CPU). VAE decode profiles at 1777 ms.

AI Hub could not profile the UNet on its 8-12GB Galaxy S24, but it runs correctly on a 15.6GB 8 Gen 3 - verified by generating, not inferred.

• —unet/ - 6 inputs: sample [1,4,128,128], timestep [1], encoder_hidden_states [1,77,2048], time_ids [1,6], text_embeds [1,1280], timestep_cond [1,256] (LCM guidance embedding). Output output_0.
• —vae_decoder/ - built from madebyollin/sdxl-vae-fp16-fix. The stock SDXL VAE overflows in fp16 (two convs peak at 640,538 and 106,287 against fp16's 65,504 ceiling) and renders pure black; the fix weights peak at 2,301.

Both are EPContext wrappers: model.onnx names ./model.bin beside it. The architecture is fixed - a v75 binary will not load on v79.