app.py

import os
import gradio as gr
import json
import logging
import torch
from os import path
from PIL import Image
import spaces
from diffusers import DiffusionPipeline, AutoPipelineForText2Image
from diffusers import StableDiffusion3Pipeline, FlowMatchEulerDiscreteScheduler, SD3Transformer2DModel # pip install diffusers>=0.31.0
from transformers import CLIPModel, CLIPProcessor, CLIPTextModel, CLIPTokenizer, CLIPConfig, T5EncoderModel, T5Tokenizer
import copy
import random
import time
from huggingface_hub import HfFileSystem, ModelCard
from huggingface_hub import login, hf_hub_download
import safetensors.torch
from safetensors.torch import load_file
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("high")

#torch._inductor.config.conv_1x1_as_mm = True
#torch._inductor.config.coordinate_descent_tuning = True
#torch._inductor.config.epilogue_fusion = False
#torch._inductor.config.coordinate_descent_check_all_directions = True


# Load LoRAs from JSON file
with open('loras.json', 'r') as f:
    loras = json.load(f)
    
# Initialize the base model
#base_model = "stabilityai/stable-diffusion-3.5-large"
pipe = AutoPipelineForText2Image.from_pretrained("ariG23498/sd-3.5-merged", torch_dtype=torch.bfloat16)

clipmodel = 'norm'
if clipmodel == "long":
    model_id = "zer0int/LongCLIP-GmP-ViT-L-14"
    config = CLIPConfig.from_pretrained(model_id)
    maxtokens = 248
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)

pipe.tokenizer = clip_processor.tokenizer
pipe.text_encoder = clip_model.text_model
pipe.tokenizer_max_length = maxtokens
pipe.text_encoder.dtype = torch.bfloat16


#pipe.transformer.to(memory_format=torch.channels_last)
#pipe.vae.to(memory_format=torch.channels_last)

#pipe.transformer = torch.compile(pipe.transformer, mode="max-autotune", fullgraph=True)
#pipe.vae.decode = torch.compile(pipe.vae.decode, mode="max-autotune", fullgraph=True)

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 infer(prompt, negative_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}",
            negative_prompt=negative_prompt,
            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, negative_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 = infer(prompt, negative_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/AlekseyCalvin/StabledHSTorY_SD3.5_LoRA_V2_rank256/resolve/main/acs62v.png" alt="LoRA">Stabled HSTorY S.D.3.5 LoRAs</h1>""",
        elem_id="title",
    )
    	    # Info blob stating what the app is running
    info_blob = gr.HTML(
        """<div id="info_blob">SOON®'s curated LoRa Gallery & Art Manufactory.|Now testing HST-triggered historic photo-trained LoRAs for Stable Diffusion 3.5.</div>"""
    )

        # Info blob stating what the app is running
    info_blob = gr.HTML(
        """<div id="info_blob">Prephrase prompts w/: "HST style autochrome photo" </div>"""
    )
    selected_index = gr.State(None)
    with gr.Row():
        with gr.Column(scale=2):
            prompt = gr.Textbox(label="Prompt", lines=1, placeholder="Select LoRa/Style & type prompt!")
        with gr.Column(scale=2):
            negative_prompt = gr.Textbox(label="Negative Prompt", lines=1, placeholder="What to exclude!")
        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=1.7)
                    steps = gr.Slider(label="Steps", minimum=1, maximum=50, step=1, value=7)
                
                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=3.0, 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, negative_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()