modules/model.py

# modules/model.py
import os
import torch
from diffusers import StableDiffusionPipeline, StableDiffusionXLPipeline
from transformers import AutoencoderKL
import spaces

def get_checkpoints(folder):
    checkpoints = []
    for file in os.listdir(folder):
        if file.endswith(('.safetensors', '.ckpt', '.pt', '.pth')):
            checkpoints.append(file)
    return checkpoints

def load_model(checkpoint, vae, checkpoint_folder, vae_folder):
    # Memilih pipeline yang sesuai
    if "sdxl" in checkpoint.lower():
        pipeline_class = StableDiffusionXLPipeline
    else:
        pipeline_class = StableDiffusionPipeline

    # Load checkpoint
    if checkpoint in get_checkpoints(checkpoint_folder):
        checkpoint_path = os.path.join(checkpoint_folder, checkpoint)
        try:
            model = pipeline_class.from_single_file(checkpoint_path, torch_dtype=torch.float16)
        except Exception as e:
            model = pipeline_class.from_pretrained(checkpoint_path, torch_dtype=torch.float16)
    else:
        if checkpoint.startswith("http"):
            try:
                model = pipeline_class.from_single_file(checkpoint, torch_dtype=torch.float16)
            except Exception as e:
                model = pipeline_class.from_pretrained(checkpoint, torch_dtype=torch.float16)
        else:
            model = pipeline_class.from_pretrained(checkpoint, torch_dtype=torch.float16)

    # Load VAE
    if vae != "none":
        if vae in get_checkpoints(vae_folder):
            vae_path = os.path.join(vae_folder, vae)
            vae_model = AutoencoderKL.from_pretrained(vae_path, torch_dtype=torch.float16)
        else:
            vae_model = AutoencoderKL.from_pretrained(vae, torch_dtype=torch.float16)
        model.vae = vae_model

    return model

def get_model_and_vae_options():
    checkpoint_folder = "../models/checkpoint/"
    vae_folder = "../models/vae/"
    model_file = "../models/model.py"

    # Membaca model dan VAE dari models/model.py
    exec(open(model_file).read())

    # Mendapatkan daftar checkpoint dan VAE dari folder
    checkpoints = get_checkpoints(checkpoint_folder)
    vae_files = get_checkpoints(vae_folder)

    # Menggabungkan daftar checkpoint, model Diffusers, dan VAE
    all_models = checkpoints + diffusers
    all_vaes = ["none"] + vae_files + vae
    
    # Mengubah format dropdown
    formatted_models = [os.path.basename(model) if not model.startswith("http") else model for model in all_models]
    formatted_vaes = [os.path.basename(vae) if not vae.startswith("http") else vae for vae in all_vaes]

    return formatted_models, formatted_vaes

@spaces.GPU()
def generate_image(text, neg_prompt, width, height, scheduler, num_steps, num_images, cfg_scale, seed, model, vae):
    checkpoint_folder = "../models/checkpoint/"
    vae_folder = "../models/vae/"
    model = load_model(model, vae, checkpoint_folder, vae_folder)
    images = model([text], height=height, width=width, num_inference_steps=num_steps, guidance_scale=cfg_scale, negative_prompt=neg_prompt)
    return images.images