generate-embeddingXL.py

#!/bin/env python

""" Work in progress

Similar to generate-embedding.py, but outputs in the format
that SDXL models expect. I hope.

Also tries to load the SDXL base text encoder specifically.
Requires you populate the two paths mentioned immediately below this comment section.

You can get them from:
https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main/text_encoder_2

(rename diffusion_pytorch_model.safetensors to text_encoder_xl.safetensors)


Plan:
   Take input for a single word or phrase.
   Save out calculations, to "generatedXL.safetensors"

   Note that you can generate an embedding from two words, or even more
   
   I could also include a "clip_l" key, but..
   Meh. 
"""

model_path = "text_encoder_xl.safetensors"
model_config = "text_encoder_2_config.json"

import sys
import torch
from transformers import CLIPProcessor, CLIPTextModel, CLIPTextModelWithProjection
from safetensors.torch import save_file

# 1. Load the pretrained model
# Note that it doesnt like a leading "/" in the name!!


model=None
processor=None

device=torch.device("cuda")

# Note the default, required 2 pathnames
def initXLCLIPmodel():
    global model
    print("loading",model_path)
    model = CLIPTextModelWithProjection.from_pretrained(model_path,config=model_config,local_files_only=True,use_safetensors=True)
    model.to(device)

# a bit wierd, but SDXL seems to still use this tokeninzer
def initCLIPprocessor():
    global processor
    CLIPname= "openai/clip-vit-large-patch14"
    print("getting processor from",CLIPname)
    processor = CLIPProcessor.from_pretrained(CLIPname)

def embed_from_text(text):
    global processor,model
    if processor == None:
        initCLIPprocessor()
        initXLCLIPmodel()
    print("getting tokens")
    inputs = processor(text=text, return_tensors="pt")
    inputs.to(device)

    print("getting embeddings?")
    outputs = model(**inputs)
    print("finalizing")
    embeddings = outputs.text_embeds
    return embeddings



##########################################

word = input("type a phrase to generate an embedding for: ")

emb = embed_from_text(word)
#embs=emb.unsqueeze(0) # stupid matrix magic to make it the required shape
embs=emb

print("Shape of result = ",embs.shape)
# Note that programs like shapes such as
#  torch.Size([1, 768])

output = "generatedXL.safetensors"
# if single word used, then rename output file
if all(char.isalpha() for char in word):
    output=f"{word}XL.safetensors"
print(f"Saving to {output}...")
save_file({"clip_g": embs}, output)

# technically we are saving a shape ([1][1280])
# whereas official XL embeddings files, are 
# (clip_g) shape ([8][1280])
# (clip_l) shape ([8][768])