README.md

metadata

license: creativeml-openrail-m
base_model:
  - stabilityai/stable-diffusion-2-1-base
tags:
  - stable-diffusion
  - mnn
  - cpu
  - onnx
  - image-generation
inference: false

Stable Diffusion 2.1 (MNN Conversion)

This repository contains a Stable Diffusion 2.1 model converted to the MNN format for efficient inference on CPU.
The base model is stabilityai/stable-diffusion-2-1-base.

Conversion Process

1. Exported the original model to ONNX using MNN:

   cd mnn_path/transformers/diffusion/export
   python onnx_export.py \
       --model_path stabilityai/stable-diffusion-2-1-base \
       --output_path output/path
   

2. Converted each component from ONNX → MNN:

   # Clip (text encoder)
   ./MNNConvert -f ONNX --modelFile ./text_encoder.onnx --MNNModel ./text_encoder.mnn --fp16
   # Unet 
   ./MNNConvert -f ONNX --modelFile ./unet.onnx --MNNModel ./unet.mnn --weightQuantBits 8 --weightQuantAsymmetric --weightQuantBlock 32
   # VAE decoder 
   ./MNNConvert -f ONNX --modelFile ./vae_decoder.onnx --MNNModel ./vae_decoder.mnn --fp16