app.py

import os
import gradio as gr
import json
import logging
import torch
from PIL import Image
from os import path
from torchvision import transforms
from dataclasses import dataclass
import math
from typing import Callable
import spaces
from diffusers import DiffusionPipeline, AutoencoderTiny, AutoencoderKL
from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
from transformers import CLIPModel, CLIPProcessor, CLIPTextModel, CLIPTokenizer, CLIPConfig, T5EncoderModel, T5Tokenizer
from diffusers.models.transformers import FluxTransformer2DModel
import copy
import random
import time
import safetensors.torch
from tqdm import tqdm
from safetensors.torch import load_file
from huggingface_hub import HfFileSystem, ModelCard
from huggingface_hub import login, hf_hub_download
hf_token = os.environ.get("HF_TOKEN")
login(token=hf_token)

cache_path = path.join(path.dirname(path.abspath(__file__)), "models")
os.environ["TRANSFORMERS_CACHE"] = cache_path
os.environ["HF_HUB_CACHE"] = cache_path
os.environ["HF_HOME"] = cache_path

#torch.set_float32_matmul_precision("medium")

# Load LoRAs from JSON file
with open('loras.json', 'r') as f:
    loras = json.load(f)

# Initialize the base model
dtype = torch.bfloat16
base_model = "shuttleai/shuttle-jaguar"
pipe = DiffusionPipeline.from_pretrained(base_model, torch_dtype=dtype).to("cuda")
#pipe.vae = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=torch.float16).to("cuda")
torch.cuda.empty_cache()

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

model_id = ("zer0int/LongCLIP-GmP-ViT-L-14")
config = CLIPConfig.from_pretrained(model_id)
config.text_config.max_position_embeddings = 248
clip_model = CLIPModel.from_pretrained(model_id, torch_dtype=torch.bfloat16, config=config, ignore_mismatched_sizes=True)
clip_processor = CLIPProcessor.from_pretrained(model_id, padding="max_length", max_length=248)
pipe.tokenizer = clip_processor.tokenizer
pipe.text_encoder = clip_model.text_model
pipe.tokenizer_max_length = 248
pipe.text_encoder.dtype = torch.bfloat16
#pipe.text_encoder_2 = t5.text_model

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()
        return self
    
    def __exit__(self, exc_type, exc_value, traceback):
        self.end_time = time.time()
        self.elapsed_time = self.end_time - self.start_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 update_selection(evt: gr.SelectData, width, height):
    selected_lora = loras[evt.index]
    new_placeholder = f"Prompt with activator word(s): '{selected_lora['trigger_word']}'! "
    lora_repo = selected_lora["repo"]
    lora_trigger = selected_lora['trigger_word']
    updated_text = f"### Selected: [{lora_repo}](https://huggingface.co/{lora_repo}). Prompt using: '{lora_trigger}'!"
    if "aspect" in selected_lora:
        if selected_lora["aspect"] == "portrait":
            width = 768
            height = 1024
        elif selected_lora["aspect"] == "landscape":
            width = 1024
            height = 768
    return (
        gr.update(placeholder=new_placeholder),
        updated_text,
        evt.index,
        width,
        height,
    )

@spaces.GPU(duration=50)
def generate_image(prompt, trigger_word, steps, seed, cfg_scale, width, height, lora_scale, progress): 
    pipe.to("cuda")
    generator = torch.Generator(device="cuda").manual_seed(seed)
    
    with calculateDuration("Generating image"):
        # Generate image
        image = pipe(
            prompt=f"{prompt} {trigger_word}",
            num_inference_steps=steps,
            guidance_scale=cfg_scale,
            width=width,
            height=height,
            generator=generator,
            joint_attention_kwargs={"scale": lora_scale},
        ).images[0]
    return image

def run_lora(prompt, cfg_scale, steps, selected_index, randomize_seed, seed, width, height, lora_scale, progress=gr.Progress(track_tqdm=True)):
    if selected_index is None:
        raise gr.Error("You must select a LoRA before proceeding.")

    selected_lora = loras[selected_index]
    lora_path = selected_lora["repo"]
    trigger_word = selected_lora['trigger_word']

    # Load LoRA weights
    with calculateDuration(f"Loading LoRA weights for {selected_lora['title']}"):
        if "weights" in selected_lora:
            pipe.load_lora_weights(lora_path, weight_name=selected_lora["weights"])
        else:
            pipe.load_lora_weights(lora_path)
        
    # Set random seed for reproducibility
    with calculateDuration("Randomizing seed"):
        if randomize_seed:
            seed = random.randint(0, MAX_SEED)
    
    image = generate_image(prompt, trigger_word, steps, seed, cfg_scale, width, height, lora_scale, progress)
    pipe.to("cpu")
    pipe.unload_lora_weights()
    return image, seed  

run_lora.zerogpu = True

css = '''
#gen_btn{height: 100%}
#title{text-align: center}
#title h1{font-size: 3em; display:inline-flex; align-items:center}
#title img{width: 100px; margin-right: 0.5em}
#gallery .grid-wrap{height: 10vh}
'''
with gr.Blocks(theme=gr.themes.Soft(), css=css) as app:
    title = gr.HTML(
        """<h1><img src="https://huggingface.co/spaces/multimodalart/flux-lora-the-explorer/resolve/main/flux_lora.png" alt="LoRA"> SOONfactory </h1>""",
        elem_id="title",
    )
    	    # Info blob stating what the app is running
    info_blob = gr.HTML(
        """<div id="info_blob"> Img. Manufactory Running On: the Shuttle Jaguar Model (a fast open offshoot of FLUX). Nearly all of the LoRA adapters accessible via this space were trained by us in an extensive progression of inspired experiments and conceptual mini-projects. Check out our poetry translations at WWW.SILVERagePOETS.com Find our music on SoundCloud @ AlekseyCalvin & YouTube @ SilverAgePoets / AlekseyCalvin!  </div>"""
    )

        # Info blob stating what the app is running
    info_blob = gr.HTML(
        """<div id="info_blob"> To reinforce/focus in selected fine-tuned LoRAs (Low-Rank Adapters), add special “trigger" words/phrases to your prompts. </div>"""
    )
    selected_index = gr.State(None)
    with gr.Row():
        with gr.Column(scale=3):
            prompt = gr.Textbox(label="Prompt", lines=1, placeholder="Select LoRa/Style & type prompt!")
        with gr.Column(scale=1, elem_id="gen_column"):
            generate_button = gr.Button("Generate", variant="primary", elem_id="gen_btn")
    with gr.Row():
        with gr.Column(scale=3):
            selected_info = gr.Markdown("")
            gallery = gr.Gallery(
                [(item["image"], item["title"]) for item in loras],
                label="LoRA Inventory",
                allow_preview=False,
                columns=3,
                elem_id="gallery"
            )
            
        with gr.Column(scale=4):
            result = gr.Image(label="Generated Image")

    with gr.Row():
        with gr.Accordion("Advanced Settings", open=True):
            with gr.Column():
                with gr.Row():
                    cfg_scale = gr.Slider(label="CFG Scale", minimum=0, maximum=20, step=.1, value=1.0)
                    steps = gr.Slider(label="Steps", minimum=1, maximum=50, step=1, value=9)
                
                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():
                    randomize_seed = gr.Checkbox(True, label="Randomize seed")
                    seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, randomize=True)
                    lora_scale = gr.Slider(label="LoRA Scale", minimum=0, maximum=2.5, step=0.01, value=1.0)

    gallery.select(
        update_selection,
        inputs=[width, height],
        outputs=[prompt, selected_info, selected_index, width, height]
    )

    gr.on(
        triggers=[generate_button.click, prompt.submit],
        fn=run_lora,
        inputs=[prompt, cfg_scale, steps, selected_index, randomize_seed, seed, width, height, lora_scale],
        outputs=[result, seed]
    )

app.queue(default_concurrency_limit=2).launch(show_error=True)
app.launch()