app.py

import gradio as gr
import json
import logging
import argparse
import torch
import os
from os import path
from PIL import Image
import numpy as np
import spaces
import copy
import random
import time
from typing import Any, Dict, List, Optional, Union
from huggingface_hub import hf_hub_download
from diffusers import DiffusionPipeline, FluxTransformer2DModel, FluxPipeline, AutoencoderTiny
import safetensors.torch
from safetensors.torch import load_file
from custom_pipeline import FluxWithCFGPipeline
from transformers import CLIPModel, CLIPProcessor, CLIPConfig
import gc

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.backends.cuda.matmul.allow_tf32 = True

dtype = torch.bfloat16
pipe = FluxWithCFGPipeline.from_pretrained(
    "ostris/OpenFLUX.1", torch_dtype=dtype
).to("cuda")
pipe.vae = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=dtype).to("cuda")

pipe.to("cuda")
clipmodel = 'norm'
if clipmodel == "long":
    model_id = "zer0int/LongCLIP-GmP-ViT-L-14"
    config = CLIPConfig.from_pretrained(model_id)
    maxtokens = 77
if clipmodel == "norm":
    model_id = "zer0int/CLIP-GmP-ViT-L-14"
    config = CLIPConfig.from_pretrained(model_id)
    maxtokens = 77
clip_model = CLIPModel.from_pretrained(model_id, torch_dtype=torch.bfloat16, config=config, ignore_mismatched_sizes=True).to("cuda")
clip_processor = CLIPProcessor.from_pretrained(model_id, padding="max_length", max_length=maxtokens, ignore_mismatched_sizes=True, return_tensors="pt", truncation=True)
config.text_config.max_position_embeddings = 77

pipe.tokenizer = clip_processor.tokenizer
pipe.text_encoder = clip_model.text_model
pipe.tokenizer_max_length = maxtokens
pipe.text_encoder.dtype = torch.bfloat16
torch.cuda.empty_cache()

# Load LoRAs from JSON file
with open('loras.json', 'r') as f:
    loras = json.load(f)
    
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"Type a prompt for {selected_lora['title']}"
    lora_repo = selected_lora["repo"]
    updated_text = f"### Selected: [{lora_repo}](https://huggingface.co/{lora_repo}) ✨"
    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=70)
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"]
    if(trigger_word):
        if "trigger_position" in selected_lora:
            if selected_lora["trigger_position"] == "prepend":
                prompt_mash = f"{trigger_word} {prompt}"
            else:
                prompt_mash = f"{prompt} {trigger_word}"
        else:
            prompt_mash = f"{trigger_word} {prompt}"
    else:
        prompt_mash = prompt

    # 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"], adapter_name="soonr")
            pipe.load_lora_weights("ostris/OpenFLUX.1", weight_name="openflux1-v0.1.0-fast-lora.safetensors", adapter_name="fast")
            pipe.set_adapters(["fast", "soonr"], adapter_weights=[1.0, lora_scale])
        else:
            pipe.load_lora_weights(lora_path, adapter_name="choice")
            pipe.load_lora_weights("ostris/OpenFLUX.1", weight_name="openflux1-v0.1.0-fast-lora.safetensors", adapter_name="fast")
            pipe.set_adapters(["fast", "choice"], adapter_weights=[1.0, lora_scale])
        
    # 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
#pipe.load_lora_weights("ostris/OpenFLUX.1", weight_name="openflux1-v0.1.0-fast-lora.safetensors", adapter_name="fast")
#pipe.set_adapters("fast")
#pipe.set_adapters(["fast", "toy"], adapter_weights=[0.5, 1.0])
#pipe.fuse_lora(adapter_names=["fast"], lora_scale=1.0)

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"> SOON®'s Activist & Futurealist LoRa-stocked Img Manufactory (now running on Ostris' OpenFLUX.1 model + their Fast LoRA co-activated, + using zer0int's fine-tuned CLIP-GmP-ViT-L-14! ('normal' version w/max length of 77 tokens)) </div>"""
    )

        # Info blob stating what the app is running
    info_blob = gr.HTML(
        """<div id="info_blob">Prephrase prompts w/: 1-2. HST style autochrome photo |3. RCA style Communist poster |4. SOTS art |5. HST Austin Osman Spare style |6. Mayakovsky |7-8. Tsvetaeva |9. Akhmatova |10. Mandelshtam |11-13. Blok  |14. LEN Lenin |15. Trotsky |16. Rosa Luxemburg |17-30. HST style |31. How2Draw |32. Propaganda Poster |33. TOK hybrid |34. 2004 photo |35. Unexpected Photo of |36. flmft |37. Yearbook photo |38. TOK portra |39. pficonics |40. retrofuturism |41. wh3r3sw4ld0  |42. amateur photo shot on phone |43. crisp photo |44-45. IMG_1099.CR2 photo |46. vintage cover art </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=1, maximum=20, step=0.5, value=2.5)
                    steps = gr.Slider(label="Steps", minimum=1, maximum=50, step=1, value=5)
                
                with gr.Row():
                    width = gr.Slider(label="Width", minimum=256, maximum=1536, step=64, value=768)
                    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=2.0, step=0.01, value=0.8)

    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=None).launch(show_error=True)
app.launch()