app.py

import gradio as gr

from convert import run_conversion
from hub_utils import push_to_hub, save_model_card

PRETRAINED_CKPT = "CompVis/stable-diffusion-v1-4"
DESCRIPTION = """
This Space lets you convert KerasCV Stable Diffusion weights to a format compatible with [Diffusers](https://github.com/huggingface/diffusers) 🧨. This allows users to fine-tune using KerasCV and use the fine-tuned weights in Diffusers taking advantage of its nifty features (like [schedulers](https://huggingface.co/docs/diffusers/main/en/using-diffusers/schedulers), [fast attention](https://huggingface.co/docs/diffusers/optimization/fp16), etc.). Specifically, the Keras weights are first converted to PyTorch and then they are wrapped into a [`StableDiffusionPipeline`](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/overview). This pipeline is then pushed to the Hugging Face Hub given you have provided `your_hf_token`.

## Notes (important)

* The Space downloads a couple of pre-trained weights and runs a dummy inference. Depending, on the machine type, the enture process can take anywhere between 2 - 5 minutes.
* Only Stable Diffusion (v1) is supported as of now. In particular this checkpoint: [`"CompVis/stable-diffusion-v1-4"`](https://huggingface.co/CompVis/stable-diffusion-v1-4).
* Only the text encoder and UNet parameters are converted since only these two elements are generally fine-tuned.
* [This Colab Notebook](https://colab.research.google.com/drive/1RYY077IQbAJldg8FkK8HSEpNILKHEwLb?usp=sharing) was used to develop the conversion utilities initially.
* You can choose NOT to provide `text_encoder_weights` and `unet_weights` in case you don't have any fine-tuned weights. In that case, the original parameters of the respective models (text encoder and UNet) from KerasCV will be used.
* You can provide only `text_encoder_weights` or `unet_weights` or both.
* When providing the weights' links, ensure they're directly downloadable. Internally, the Space uses [`tf.keras.utils.get_file()`](https://www.tensorflow.org/api_docs/python/tf/keras/utils/get_file) to retrieve the weights locally. 
* If you don't provide `your_hf_token` the converted pipeline won't be pushed. 

Check [here](https://github.com/huggingface/diffusers/blob/31be42209ddfdb69d9640a777b32e9b5c6259bf0/examples/dreambooth/train_dreambooth_lora.py#L975) for an example on how you can change the scheduler of an already initialized `StableDiffusionPipeline`.
"""


def run(hf_token, text_encoder_weights, unet_weights, repo_prefix):
    if text_encoder_weights == "":
        text_encoder_weights = None
    if unet_weights == "":
        unet_weights = None
    print(f"unet_weights: {unet_weights}")
    pipeline = run_conversion(text_encoder_weights, unet_weights)
    output_path = "kerascv_sd_diffusers_pipeline"
    pipeline.save_pretrained(output_path)

    weight_paths = []
    if text_encoder_weights is not None:
        weight_paths.append(text_encoder_weights)
    if unet_weights is not None:
        weight_paths.append(unet_weights)
    save_model_card(
        base_model=PRETRAINED_CKPT,
        repo_folder=output_path,
        weight_paths=weight_paths,
    )
    push_str = push_to_hub(hf_token, output_path, repo_prefix)
    return push_str


demo = gr.Interface(
    title="KerasCV Stable Diffusion to Diffusers Stable Diffusion Pipelines 🧨🤗",
    description=DESCRIPTION,
    allow_flagging="never",
    inputs=[
        gr.Text(max_lines=1, label="your_hf_token"),
        gr.Text(max_lines=1, label="text_encoder_weights"),
        gr.Text(max_lines=1, label="unet_weights"),
        gr.Text(max_lines=1, label="output_repo_prefix"),
    ],
    outputs=[gr.Markdown(label="output")],
    fn=run,
)

demo.launch()