| --- |
| license: apache-2.0 |
| library_name: keras |
| pipeline_tag: text-to-image |
| tags: |
| - diffusion model |
| - stable diffusion |
| - v1.4 |
| --- |
| |
| This repository hosts the TFLite version of `diffusion model` part of [KerasCV Stable Diffusion](https://github.com/keras-team/keras-cv/tree/master/keras_cv/models/stable_diffusion). |
|
|
| Stable Diffusion consists of `text encoder`, `diffusion model`, `decoder`, and some glue codes to handl inputs and outputs of each part. The TFLite version of `diffusion model` in this repository is built not only with the `diffusion model` itself but also TensorFlow operations that takes `context`, `unconditional context` from `text encoder` and generates `latent`. The `latent` output should be passed down to the `decoder` which is hosted in [this repository](https://huggingface.co/keras-sd/decoder-tflite/tree/main). |
|
|
| TFLite conversion was based on the `SavedModel` from [this repository](https://huggingface.co/keras-sd/tfs-text-encoder/tree/main), and TensorFlow version `>= 2.12-nightly` was used. |
| - NOTE: [Dynamic range quantization](https://www.tensorflow.org/lite/performance/post_training_quant#optimizing_an_existing_model) was used. |
| - NOTE: TensorFlow version `< 2.12-nightly` will fail for the conversion process. |
| - NOTE: For those who wonder how `SavedModel` is constructed, find it in [keras-sd-serving repository](https://github.com/deep-diver/keras-sd-serving). |
