Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,18 @@
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
+
library_name: keras
|
| 4 |
+
pipeline_tag: text-to-image
|
| 5 |
+
tags:
|
| 6 |
+
- diffusion model
|
| 7 |
+
- stable diffusion
|
| 8 |
+
- v1.4
|
| 9 |
---
|
| 10 |
+
|
| 11 |
+
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).
|
| 12 |
+
|
| 13 |
+
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).
|
| 14 |
+
|
| 15 |
+
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.
|
| 16 |
+
- NOTE: [Dynamic range quantization](https://www.tensorflow.org/lite/performance/post_training_quant#optimizing_an_existing_model) was used.
|
| 17 |
+
- NOTE: TensorFlow version `< 2.12-nightly` will fail for the conversion process.
|
| 18 |
+
- NOTE: For those who wonder how `SavedModel` is constructed, find it in [keras-sd-serving repository](https://github.com/deep-diver/keras-sd-serving).
|