app.py

import subprocess
subprocess.run(['sh', './spaces.sh'])

import os
# Environment variable setup
os.putenv('PYTORCH_NVML_BASED_CUDA_CHECK','1')
os.putenv('TORCH_LINALG_PREFER_CUSOLVER','1')
os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'expandable_segments:True,pinned_use_background_threads:True'
os.environ["SAFETENSORS_FAST_GPU"] = "1"
os.putenv('HF_HUB_ENABLE_HF_TRANSFER','1')

import spaces
import gradio as gr
import numpy as np
import random
import datetime
import threading
import io

# --- New GCS Imports ---
from google.oauth2 import service_account
from google.cloud import storage

import torch
# Torch performance settings
torch.backends.cuda.matmul.allow_tf32 = False
torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = False
torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = False
torch.backends.cudnn.allow_tf32 = False
torch.backends.cudnn.deterministic = False
torch.backends.cudnn.benchmark = False
torch.set_float32_matmul_precision("highest")

from diffusers import StableDiffusion3Pipeline, SD3Transformer2DModel, AutoencoderKL
from PIL import Image
from image_gen_aux import UpscaleWithModel

# --- GCS Configuration ---
# Make sure to set these secrets in your Hugging Face Space settings
GCS_BUCKET_NAME = os.getenv("GCS_BUCKET_NAME")
GCS_SA_KEY = os.getenv("GCS_SA_KEY") # The full JSON key content as a string

# Initialize GCS client if credentials are available
gcs_client = None
if GCS_SA_KEY and GCS_BUCKET_NAME:
    try:
        credentials_info = eval(GCS_SA_KEY) # Using eval is safe here if you trust the secret source
        credentials = service_account.Credentials.from_service_account_info(credentials_info)
        gcs_client = storage.Client(credentials=credentials)
        print("✅ GCS Client initialized successfully.")
    except Exception as e:
        print(f"❌ Failed to initialize GCS client: {e}")

def upload_to_gcs(image_object, filename):
    if not gcs_client:
        print("⚠️ GCS client not initialized. Skipping upload.")
        return
    try:
        print(f"--> Starting GCS upload for {filename}...")
        bucket = gcs_client.bucket(GCS_BUCKET_NAME)
        blob = bucket.blob(f"stablediff/{filename}")
        img_byte_arr = io.BytesIO()
        image_object.save(img_byte_arr, format='PNG', optimize=False, compress_level=0)
        img_byte_arr = img_byte_arr.getvalue()
        blob.upload_from_string(img_byte_arr, content_type='image/png')
        print(f"✅ Successfully uploaded {filename} to GCS.")
    except Exception as e:
        print(f"❌ An error occurred during GCS upload: {e}")

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

pipe = StableDiffusion3Pipeline.from_pretrained(
    "ford442/stable-diffusion-3.5-large-bf16",
    trust_remote_code=True,
    transformer=None, # Load transformer separately
    use_safetensors=True
)

ll_transformer=SD3Transformer2DModel.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='transformer').to(device, dtype=torch.bfloat16)
pipe.transformer=ll_transformer
pipe.load_lora_weights("ford442/sdxl-vae-bf16", weight_name="LoRA/UltraReal.safetensors")
pipe.to(device=device, dtype=torch.bfloat16)

upscaler_2 = UpscaleWithModel.from_pretrained("Kim2091/ClearRealityV1").to(device)

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 4096

@spaces.GPU(duration=45)
def generate_images_30(prompt, neg_prompt_1, neg_prompt_2, neg_prompt_3, width, height, guidance, steps, progress=gr.Progress(track_tqdm=True)):
    seed = random.randint(0, MAX_SEED)
    generator = torch.Generator(device=device).manual_seed(seed)
    print('-- generating image --')
    sd_image = pipe(
        prompt=prompt, prompt_2=prompt, prompt_3=prompt,
        negative_prompt=neg_prompt_1, negative_prompt_2=neg_prompt_2, negative_prompt_3=neg_prompt_3,
        guidance_scale=guidance, num_inference_steps=steps,
        width=width, height=height, generator=generator,
        max_sequence_length=512
    ).images[0]
    print('-- got image --')
    with torch.no_grad():
        upscale = upscaler_2(sd_image, tiling=True, tile_width=256, tile_height=256)
        upscale2 = upscaler_2(upscale, tiling=True, tile_width=256, tile_height=256)
    print('-- got upscaled image --')
    downscaled_upscale = upscale2.resize((upscale2.width // 4, upscale2.height // 4), Image.LANCZOS)
    return sd_image, downscaled_upscale, prompt

@spaces.GPU(duration=70)
def generate_images_60(prompt, neg_prompt_1, neg_prompt_2, neg_prompt_3, width, height, guidance, steps, progress=gr.Progress(track_tqdm=True)):
    seed = random.randint(0, MAX_SEED)
    generator = torch.Generator(device=device).manual_seed(seed)
    print('-- generating image --')
    sd_image = pipe(
        prompt=prompt, prompt_2=prompt, prompt_3=prompt,
        negative_prompt=neg_prompt_1, negative_prompt_2=neg_prompt_2, negative_prompt_3=neg_prompt_3,
        guidance_scale=guidance, num_inference_steps=steps,
        width=width, height=height, generator=generator,
        max_sequence_length=512
    ).images[0]
    print('-- got image --')
    with torch.no_grad():
        upscale = upscaler_2(sd_image, tiling=True, tile_width=256, tile_height=256)
        upscale2 = upscaler_2(upscale, tiling=True, tile_width=256, tile_height=256)
    print('-- got upscaled image --')
    downscaled_upscale = upscale2.resize((upscale2.width // 4, upscale2.height // 4), Image.LANCZOS)
    return sd_image, downscaled_upscale, prompt

@spaces.GPU(duration=110)
def generate_images_100(prompt, neg_prompt_1, neg_prompt_2, neg_prompt_3, width, height, guidance, steps, progress=gr.Progress(track_tqdm=True)):
    seed = random.randint(0, MAX_SEED)
    generator = torch.Generator(device=device).manual_seed(seed)
    print('-- generating image --')
    sd_image = pipe(
        prompt=prompt, prompt_2=prompt, prompt_3=prompt,
        negative_prompt=neg_prompt_1, negative_prompt_2=neg_prompt_2, negative_prompt_3=neg_prompt_3,
        guidance_scale=guidance, num_inference_steps=steps,
        width=width, height=height, generator=generator,
        max_sequence_length=512
    ).images[0]
    print('-- got image --')
    with torch.no_grad():
        upscale = upscaler_2(sd_image, tiling=True, tile_width=256, tile_height=256)
        upscale2 = upscaler_2(upscale, tiling=True, tile_width=256, tile_height=256)
    print('-- got upscaled image --')
    downscaled_upscale = upscale2.resize((upscale2.width // 4, upscale2.height // 4), Image.LANCZOS)
    return sd_image, downscaled_upscale, prompt

def run_inference_and_upload_30(prompt, neg_prompt_1, neg_prompt_2, neg_prompt_3, width, height, guidance, steps, save_consent, progress=gr.Progress(track_tqdm=True)):
    sd_image, upscaled_image, expanded_prompt = generate_images_30(prompt, neg_prompt_1, neg_prompt_2, neg_prompt_3, width, height, guidance, steps, progress)
    if save_consent:
        print("✅ User consented to save. Preparing uploads...")
        timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
        sd_filename = f"sd35ll_{timestamp}.png"
        upscale_filename = f"sd35ll_upscale_{timestamp}.png"
        sd_thread = threading.Thread(target=upload_to_gcs, args=(sd_image, sd_filename))
        upscale_thread = threading.Thread(target=upload_to_gcs, args=(upscaled_image, upscale_filename))
        sd_thread.start()
        upscale_thread.start()
    else:
        print("ℹ️ User did not consent to save. Skipping upload.")
    return sd_image, expanded_prompt

def run_inference_and_upload_60(prompt, neg_prompt_1, neg_prompt_2, neg_prompt_3, width, height, guidance, steps, save_consent, progress=gr.Progress(track_tqdm=True)):
    sd_image, upscaled_image, expanded_prompt = generate_images_60(prompt, neg_prompt_1, neg_prompt_2, neg_prompt_3, width, height, guidance, steps, progress)
    if save_consent:
        print("✅ User consented to save. Preparing uploads...")
        timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
        sd_filename = f"sd35ll_{timestamp}.png"
        upscale_filename = f"sd35ll_upscale_{timestamp}.png"
        sd_thread = threading.Thread(target=upload_to_gcs, args=(sd_image, sd_filename))
        upscale_thread = threading.Thread(target=upload_to_gcs, args=(upscaled_image, upscale_filename))
        sd_thread.start()
        upscale_thread.start()
    else:
        print("ℹ️ User did not consent to save. Skipping upload.")
    return sd_image, expanded_prompt

def run_inference_and_upload_100(prompt, neg_prompt_1, neg_prompt_2, neg_prompt_3, width, height, guidance, steps, save_consent, progress=gr.Progress(track_tqdm=True)):
    sd_image, upscaled_image, expanded_prompt = generate_images_100(prompt, neg_prompt_1, neg_prompt_2, neg_prompt_3, width, height, guidance, steps, progress)
    if save_consent:
        print("✅ User consented to save. Preparing uploads...")
        timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
        sd_filename = f"sd35ll_{timestamp}.png"
        upscale_filename = f"sd35ll_upscale_{timestamp}.png"
        sd_thread = threading.Thread(target=upload_to_gcs, args=(sd_image, sd_filename))
        upscale_thread = threading.Thread(target=upload_to_gcs, args=(upscaled_image, upscale_filename))
        sd_thread.start()
        upscale_thread.start()
    else:
        print("ℹ️ User did not consent to save. Skipping upload.")
    return sd_image, expanded_prompt
    
css = """
#col-container {margin: 0 auto;max-width: 640px;}
body{background-color: blue;}
"""
with gr.Blocks(theme=gr.themes.Origin(), css=css) as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown(" # StableDiffusion 3.5 Large with UltraReal lora test")
        expanded_prompt_output = gr.Textbox(label="Prompt", lines=1)
        with gr.Row():
            prompt = gr.Text(
                label="Prompt", show_label=False, max_lines=1,
                placeholder="Enter your prompt", container=False,
            )
            run_button_30 = gr.Button("Run30", scale=0, variant="primary")
            run_button_60 = gr.Button("Run60", scale=0, variant="primary")
            run_button_100 = gr.Button("Run100", scale=0, variant="primary")
        result = gr.Image(label="Result", show_label=False, type="pil")
        save_consent_checkbox = gr.Checkbox(
            label="✅ Anonymously upload result to a public gallery",
            value=True, # Default to not uploading
            info="Check this box to help us by contributing your image."
        )
        with gr.Accordion("Advanced Settings", open=True):
            negative_prompt_1 = gr.Text(label="Negative prompt 1", max_lines=1, placeholder="Enter a negative prompt", value="bad anatomy, poorly drawn hands, distorted face, blurry, out of frame, low resolution, grainy, pixelated, disfigured, mutated, extra limbs, bad composition")
            negative_prompt_2 = gr.Text(label="Negative prompt 2", max_lines=1, placeholder="Enter a second negative prompt", value="unrealistic, cartoon, anime, sketch, painting, drawing, illustration, graphic, digital art, render, 3d, blurry, deformed, disfigured, poorly drawn, bad anatomy, mutated, extra limbs, ugly, out of frame, bad composition, low resolution, grainy, pixelated, noisy, oversaturated, undersaturated, (worst quality, low quality:1.3), (bad hands, missing fingers:1.2)")
            negative_prompt_3 = gr.Text(label="Negative prompt 3", max_lines=1, placeholder="Enter a third negative prompt", value="(worst quality, low quality:1.3), (bad anatomy, bad hands, missing fingers, extra digit, fewer digits:1.2), (blurry:1.1), cropped, watermark, text, signature, logo, jpeg artifacts, (ugly, deformed, disfigured:1.2), (poorly drawn:1.2), mutated, extra limbs, (bad proportions, gross proportions:1.2), (malformed limbs, missing arms, missing legs, extra arms, extra legs:1.2), (fused fingers, too many fingers, long neck:1.2), (unnatural body, unnatural pose:1.1), out of frame, (bad composition, poorly composed:1.1), (oversaturated, undersaturated:1.1), (grainy, pixelated:1.1), (low resolution, noisy:1.1), (unrealistic, distorted:1.1), (extra fingers, mutated hands, poorly drawn hands, bad hands:1.3), (missing fingers:1.3)")
            with gr.Row():
                width = gr.Slider(label="Width", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=768)
                height = gr.Slider(label="Height", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=768)
            with gr.Row():
                guidance_scale = gr.Slider(label="Guidance scale", minimum=0.0, maximum=30.0, step=0.1, value=4.2)
                num_inference_steps = gr.Slider(label="Inference steps", minimum=1, maximum=150, step=1, value=60)

        run_button_30.click(
            fn=run_inference_and_upload_30,
            inputs=[
                prompt,
                negative_prompt_1,
                negative_prompt_2,
                negative_prompt_3,
                width,
                height,
                guidance_scale,
                num_inference_steps,
                save_consent_checkbox # Pass the checkbox value
            ],
            outputs=[result, expanded_prompt_output],
        )

        run_button_60.click(
            fn=run_inference_and_upload_60,
            inputs=[
                prompt,
                negative_prompt_1,
                negative_prompt_2,
                negative_prompt_3,
                width,
                height,
                guidance_scale,
                num_inference_steps,
                save_consent_checkbox # Pass the checkbox value
            ],
            outputs=[result, expanded_prompt_output],
        )

        run_button_100.click(
            fn=run_inference_and_upload_100,
            inputs=[
                prompt,
                negative_prompt_1,
                negative_prompt_2,
                negative_prompt_3,
                width,
                height,
                guidance_scale,
                num_inference_steps,
                save_consent_checkbox # Pass the checkbox value
            ],
            outputs=[result, expanded_prompt_output],
        )
        

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