app.py

import torch
from safetensors.torch import save_file, load_file
import gradio as gr
import os

def convert_embedding(uploaded_file):
    output_path = "embedding.safetensors"
    file_extension = os.path.splitext(uploaded_file.name)[1]

    #The sample files are probably structured in these ways because the pt files were probably all created with automatic1111, and the safetensors files were probably created with kohya_ss
    #If we learn of other programs that structure the embedding file differently, we'll have to adjust the logic.
    if file_extension == '.pt':
        sd15_embedding = torch.load(uploaded_file.name, map_location=torch.device('cpu'), weights_only=True)
        sd15_tensor = sd15_embedding['string_to_param']['*']
    elif file_extension == '.safetensors':
        loaded_tensors = load_file(uploaded_file.name)
        sd15_tensor = loaded_tensors['emb_params']
    else:
        raise ValueError("Unsupported file format")
    
    num_vectors = sd15_tensor.shape[0]
    clip_g_shape = (num_vectors, 1280)
    clip_l_shape = (num_vectors, 768)
    clip_g = torch.zeros(clip_g_shape, dtype=torch.float16)
    clip_l = torch.zeros(clip_l_shape, dtype=torch.float16)
    clip_l[:sd15_tensor.shape[0], :sd15_tensor.shape[1]] = sd15_tensor.to(dtype=torch.float16)
    save_file({"clip_g": clip_g, "clip_l": clip_l}, output_path)

    # Return the path to the converted file for download
    return output_path

iface = gr.Interface(
    fn=convert_embedding,
    inputs=gr.File(label="Upload SD1.5 embedding"),
    outputs=gr.File(label="Download converted SDXL safetensors embedding"),
    title="SD1.5 to SDXL Embedding Converter",
    description="Upload an SD1.5 embedding file to convert it to SDXL."
)

if __name__ == "__main__":
    iface.launch()