app.py

import os
import torch
import random
from huggingface_hub import snapshot_download
from kolors.pipelines.pipeline_stable_diffusion_xl_chatglm_256 import StableDiffusionXLPipeline
from kolors.models.modeling_chatglm import ChatGLMModel
from kolors.models.tokenization_chatglm import ChatGLMTokenizer
from diffusers import UNet2DConditionModel, AutoencoderKL
from diffusers import EulerDiscreteScheduler
import gradio as gr

# Download the model files
ckpt_dir = snapshot_download(repo_id="Kwai-Kolors/Kolors")

# Function to load models
def load_models():
    text_encoder = ChatGLMModel.from_pretrained(
        os.path.join(ckpt_dir, 'text_encoder'),
        torch_dtype=torch.float16).half()
    tokenizer = ChatGLMTokenizer.from_pretrained(os.path.join(ckpt_dir, 'text_encoder'))
    vae = AutoencoderKL.from_pretrained(os.path.join(ckpt_dir, "vae"), revision=None).half()
    scheduler = EulerDiscreteScheduler.from_pretrained(os.path.join(ckpt_dir, "scheduler"))
    unet = UNet2DConditionModel.from_pretrained(os.path.join(ckpt_dir, "unet"), revision=None).half()

    return StableDiffusionXLPipeline(
        vae=vae,
        text_encoder=text_encoder,
        tokenizer=tokenizer,
        unet=unet,
        scheduler=scheduler,
        force_zeros_for_empty_prompt=False
    )

# Create a global variable to hold the pipeline
pipe = load_models()

def generate_image(prompt, negative_prompt, height, width, num_inference_steps, guidance_scale, num_images_per_prompt, use_random_seed, seed, progress=gr.Progress(track_tqdm=True)):
    if use_random_seed:
        seed = random.randint(0, 2**32 - 1)
    else:
        seed = int(seed)  # Ensure seed is an integer
    
    # Move the model to the CPU for inference and clear unnecessary variables
    with torch.no_grad():
        generator = torch.Generator().manual_seed(seed)
        result = pipe(
            prompt=prompt,
            negative_prompt=negative_prompt,
            height=height,
            width=width,
            num_inference_steps=num_inference_steps,
            guidance_scale=guidance_scale,
            num_images_per_prompt=num_images_per_prompt,
            generator=generator
        )
        image = result.images
    
    return image, seed

# Gradio interface
iface = gr.Interface(
    fn=generate_image,
    inputs=[
        gr.Textbox(label="Prompt"),
        gr.Textbox(label="Negative Prompt")
    ],
    additional_inputs=[
        gr.Slider(512, 2048, 1024, step=64, label="Height"),
        gr.Slider(512, 2048, 1024, step=64, label="Width"),
        gr.Slider(20, 50, 20, step=1, label="Number of Inference Steps"),
        gr.Slider(1, 20, 5, step=0.5, label="Guidance Scale"),
        gr.Slider(1, 4, 1, step=1, label="Number of images per prompt"),
        gr.Checkbox(label="Use Random Seed", value=True),
        gr.Number(label="Seed", value=0, precision=0)
    ],
    additional_inputs_accordion=gr.Accordion(label="Advanced settings", open=False),
    outputs=[
        gr.Gallery(label="Result", elem_id="gallery", show_label=False),
        gr.Number(label="Seed Used")
    ],
    title="Kolors",
    theme='bethecloud/storj_theme',
)

iface.launch()