cuesetac/sd15-mediapipe

Stable Diffusion 1.5 for on-device Android generation

Stable Diffusion v1.5 (v1-5-pruned-emaonly), converted for the on-device MediaPipe Image Generator on Android. Used by the Cueset Studio app.

Files

sd15-sdcpp.zip holds sd15-sdcpp/sd15-q8c16.gguf (1.9 GB) for stable-diffusion.cpp. It's the fallback engine for phones without enough memory for MediaPipe. Linear layers are q8_0 and 3×3/1×1 convolution kernels are f16, because they can't be block-quantised. Everything else is f32, including the noise schedule, which breaks sampling at f16. Converted with:

sd-cli -M convert -m v1-5-pruned-emaonly.safetensors -o sd15-q8c16.gguf --type q8_0 \
  --tensor-type-rules "<exact name of each 4-D conv kernel>=f16,..."

MediaPipe

sd15-mediapipe.zip holds one folder, sd15_bins/, with:

• one fp16 .bin file per weight tensor (the VAE encoder is left out because text-to-image doesn't use it)
• bpe_simple_vocab_16e6.txt, the CLIP tokenizer vocab

Pass the unzipped folder to ImageGeneratorOptions.setImageGeneratorModelDirectory().

How it was made

Converted with MediaPipe's image_generator_converter/convert.py from google-ai-edge/mediapipe-samples, changed to read the .safetensors release instead of the pickled .ckpt. The output is the same.

License

This model is a modified copy of Stable Diffusion v1.5 and is distributed under the CreativeML OpenRAIL-M license. Your use is subject to its use-based restrictions (Attachment A). If you redistribute it, you must pass those restrictions on.