app.py

import os
import gradio as gr
import numpy as np
import random
import spaces
import torch
import json
import logging
from diffusers import DiffusionPipeline, AutoencoderTiny, AutoencoderKL, AutoPipelineForImage2Image
from huggingface_hub import login
from diffusers.utils import load_image
import time
from datetime import datetime
from io import BytesIO
import torch.nn.functional as F
from PIL import Image, ImageFilter
import time
import boto3
from io import BytesIO
import re
import json
import random
import string
from diffusers import FluxPipeline
from huggingface_hub import hf_hub_download
from diffusers.quantizers import PipelineQuantizationConfig
from diffusers import (FluxPriorReduxPipeline, FluxInpaintPipeline, FluxFillPipeline, FluxPipeline)




# Login Hugging Face Hub
HF_TOKEN = os.environ.get("HF_TOKEN")
login(token=HF_TOKEN)
import diffusers




# init
dtype = torch.bfloat16
device = "cuda:0"




base_model = "black-forest-labs/FLUX.1-dev"




txt2img_pipe = FluxPipeline.from_pretrained(base_model, torch_dtype=dtype)
txt2img_pipe = txt2img_pipe.to(device)




MAX_SEED = 2**32 - 1




class calculateDuration:
    def __init__(self, activity_name=""):
        self.activity_name = activity_name


    def __enter__(self):
        self.start_time = time.time()
        self.start_time_formatted = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime(self.start_time))
        print(f"Activity: {self.activity_name}, Start time: {self.start_time_formatted}")
        return self
    
    def __exit__(self, exc_type, exc_value, traceback):
        self.end_time = time.time()
        self.elapsed_time = self.end_time - self.start_time
        self.end_time_formatted = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime(self.end_time))
        
        if self.activity_name:
            print(f"Elapsed time for {self.activity_name}: {self.elapsed_time:.6f} seconds")
        else:
            print(f"Elapsed time: {self.elapsed_time:.6f} seconds")
        


def safe_trim_for_clip(text: str, max_words: int = 77) -> str:
    tokens = re.split(r"\s+", text.strip())
    if len(tokens) <= max_words:
        return text
    return " ".join(tokens[:max_words])




def upload_image_to_r2(image, account_id, access_key, secret_key, bucket_name):
    with calculateDuration("Upload images"):
        connectionUrl = f"https://{account_id}.r2.cloudflarestorage.com"
        s3 = boto3.client(
            's3',
            endpoint_url=connectionUrl,
            region_name='auto',
            aws_access_key_id=access_key,
            aws_secret_access_key=secret_key
        )
    
        current_time = datetime.now().strftime("%Y/%m/%d/%H/%M/%S")
        image_file = f"generated_images/{current_time}/{random.randint(0, MAX_SEED)}.png"
        buffer = BytesIO()
        image.save(buffer, "PNG")
        buffer.seek(0)
        s3.upload_fileobj(buffer, bucket_name, image_file)
        print("upload finish", image_file)
        
        # Generate thumbnail
        thumbnail = image.copy()
        thumbnail_width = 256
        aspect_ratio = image.height / image.width
        thumbnail_height = int(thumbnail_width * aspect_ratio)
        thumbnail = thumbnail.resize((thumbnail_width, thumbnail_height), Image.LANCZOS)
        
        thumbnail_file = image_file.replace(".png", "_thumbnail.png")
        
        thumbnail_buffer = BytesIO()
        thumbnail.save(thumbnail_buffer, "PNG")
        thumbnail_buffer.seek(0)
        s3.upload_fileobj(thumbnail_buffer, bucket_name, thumbnail_file)
        print("upload thumbnail finish", thumbnail_file)
    return image_file




def generate_random_4_digit_string():
    return ''.join(random.choices(string.digits, k=4))




@spaces.GPU(duration=120)
def run_lora(
    prompt, 
    image_url, 
    lora_strings_json, 
    image_strength,  
    cfg_scale, 
    steps, 
    randomize_seed, 
    seed, 
    width, 
    height, 
    upload_to_r2, 
    account_id, 
    access_key, 
    secret_key, 
    bucket, 
    progress=gr.Progress(track_tqdm=True)
):
    print("run_lora", prompt, lora_strings_json, cfg_scale, steps, width, height)
    gr.Info("Starting process")
    pipe = txt2img_pipe
    device = pipe.device
    print(device)


    # ========== Seed ==========
    if randomize_seed:
        with calculateDuration("Set random seed"):
            seed = random.randint(0, MAX_SEED)
    generator = torch.Generator(device=device).manual_seed(seed)


    # ========== LoRA ==========
    gr.Info("Start to load LoRA ...")
    with calculateDuration("Unloading LoRA"):
        try:
            pipe.unload_lora_weights()
        except Exception as _:
            pass


    adapter_names = []
    adapter_weights = []
    if lora_strings_json:
        try:
            lora_configs = json.loads(lora_strings_json)
        except Exception as _:
            lora_configs = None
            gr.Warning("Parse lora config json failed")
            print("parse lora config json failed")


        if lora_configs:
            with calculateDuration("Loading LoRA weights"):
                for lora_info in lora_configs:
                    repo = lora_info.get("repo")
                    weights = lora_info.get("weights")
                    adapter_name = lora_info.get("adapter_name") or f"adp_{generate_random_4_digit_string()}"
                    weight = float(lora_info.get("adapter_weight", 1.0))
                    if not (repo and weights):
                        print(f"skip invalid lora entry: {lora_info}")
                        continue
                    try:
                        weight_path = hf_hub_download(repo_id=repo, filename=weights)
                        pipe.load_lora_weights(weight_path, adapter_name=adapter_name, prefix=None)
                        adapter_names.append(adapter_name)
                        adapter_weights.append(weight)
                    except Exception as e:
                        print(f"load lora error for {repo}/{weights}: {e}")


                if adapter_names:
                    pipe.set_adapters(adapter_names, adapter_weights=adapter_weights)
                    try:
                        active = pipe.get_active_adapters() if hasattr(pipe, "get_active_adapters") else []
                        print("Active adapters:", active)
                    except Exception as e:
                        print("Active adapters query failed:", e)
                    
                    lora_layer_count = 0
                    for name, module in pipe.transformer.named_modules():
                        attrs = dir(module)
                        if any(a.startswith("lora_") for a in attrs) or "lora" in module.__class__.__name__.lower():
                            lora_layer_count += 1
                    print(f"[DEBUG] transformer LoRA layers: {lora_layer_count}")
                    
                    if lora_layer_count == 0:
                        gr.Warning("LoRA seems not injected (0 layers on transformer). Check whether the LoRA is trained for FLUX and `prefix=None` is set.")


    pipe.enable_vae_slicing()
    clip_side_prompt = safe_trim_for_clip(prompt, max_words=77) 
    init_image = None
    error_message = ""
    image = None
    
    try:
        gr.Info("Start to generate images ...")
        joint_attention_kwargs = {"scale": 1}
        image = pipe(
            prompt=prompt,
            num_inference_steps=int(steps),
            guidance_scale=float(cfg_scale),
            width=int(width),
            height=int(height),
            max_sequence_length=512,
            generator=generator,
            joint_attention_kwargs=joint_attention_kwargs
        ).images[0]
    except Exception as e:
        error_message = str(e)
        gr.Error(error_message)
        print("fatal error", e)
        image = None


    result = {"status": "failed", "message": error_message} if image is None else {"status": "success", "message": "Image generated but not uploaded", "seed": seed}
    
    if image is not None and upload_to_r2:
        try:
            url = upload_image_to_r2(image, account_id, access_key, secret_key, bucket)
            result = {"status": "success", "message": "upload image success", "url": url, "seed": seed}
        except Exception as e:
            err = f"Upload failed: {e}"
            gr.Warning(err)
            print(err)
            result = {"status": "success", "message": "generated but upload failed", "seed": seed}


    gr.Info("Completed！")
    progress(100, "Completed!")
    
    # CHANGED: Return both image AND json
    return image, json.dumps(result)




# Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("# flux-dev-multi-lora")
    with gr.Row():
        
        with gr.Column():
            prompt = gr.Text(label="Prompt", placeholder="Enter prompt", lines=10)
            lora_strings_json = gr.Text(
                label="LoRA Configs (JSON List String)", 
                placeholder='[{"repo": "lora_repo1", "weights": "weights1.safetensors", "adapter_name": "adapter1", "adapter_weight": 1.0}]', 
                lines=5
            )
            image_url = gr.Text(label="Image url", placeholder="Enter image url to enable image to image model", lines=1)
            run_button = gr.Button("Run", scale=0)


            with gr.Accordion("Advanced Settings", open=False):
                with gr.Row():
                    seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, randomize=True)
                    randomize_seed = gr.Checkbox(label="Randomize seed", value=True)


                with gr.Row():
                    width = gr.Slider(label="Width", minimum=256, maximum=1536, step=64, value=1024)
                    height = gr.Slider(label="Height", minimum=256, maximum=1536, step=64, value=1024)


                with gr.Row():
                    image_strength = gr.Slider(label="Denoise Strength", info="Lower means more image influence", minimum=0.1, maximum=1.0, step=0.01, value=0.75)
                    cfg_scale = gr.Slider(label="CFG Scale", minimum=1, maximum=20, step=0.5, value=3.5)
                    steps = gr.Slider(label="Steps", minimum=1, maximum=50, step=1, value=28) 


                upload_to_r2 = gr.Checkbox(label="Upload to R2", value=False)
                account_id = gr.Textbox(label="Account Id", placeholder="Enter R2 account id")
                access_key = gr.Textbox(label="Access Key", placeholder="Enter R2 access key here")
                secret_key = gr.Textbox(label="Secret Key", placeholder="Enter R2 secret key here")
                bucket = gr.Textbox(label="Bucket Name", placeholder="Enter R2 bucket name here")
        
        with gr.Column():
            # CHANGED: Add image output
            output_image = gr.Image(label="Generated Image", type="pil")
            json_text = gr.Text(label="Result JSON")
            
    gr.Markdown("**Disclaimer:**")
    gr.Markdown(
        "This demo is only for research purpose. This space owner cannot be held responsible for the generation of NSFW (Not Safe For Work) content through the use of this demo. Users are solely responsible for any content they create, and it is their obligation to ensure that it adheres to appropriate and ethical standards. This space owner provides the tools, but the responsibility for their use lies with the individual user."
    )
    
    inputs = [
        prompt,
        image_url,
        lora_strings_json,
        image_strength,
        cfg_scale,
        steps,
        randomize_seed,
        seed,
        width,
        height,
        upload_to_r2,
        account_id,
        access_key,
        secret_key,
        bucket
    ]


    # CHANGED: Two outputs now
    outputs = [output_image, json_text]


    run_button.click(
        fn=run_lora,
        inputs=inputs,
        outputs=outputs
    )
    
try:
    demo.queue().launch()
except:
    print("demo exception ...")