app.py

import os
import torch
import torchvision
import torchvision.transforms as transforms
from torch.utils.data import Dataset, DataLoader
import gradio as gr
import sys
import uuid
import tqdm 
import random
sys.path.append(os.path.abspath(os.path.join("", "..")))
import gc
import warnings
warnings.filterwarnings("ignore")
from PIL import Image
import numpy as np
from editing import get_direction, debias
from sampling import sample_weights
from lora_w2w import LoRAw2w
from transformers import CLIPTextModel
from lora_w2w import LoRAw2w
from diffusers import AutoencoderKL, DDPMScheduler, DiffusionPipeline, UNet2DConditionModel, LMSDiscreteScheduler
from transformers import AutoTokenizer, PretrainedConfig
from diffusers import (
    AutoencoderKL,
    DDPMScheduler,
    DiffusionPipeline,
    DPMSolverMultistepScheduler,
    UNet2DConditionModel,
    PNDMScheduler, 
    StableDiffusionPipeline
)
from huggingface_hub import snapshot_download
import spaces
from gradio_imageslider import ImageSlider


models_path = snapshot_download(repo_id="Snapchat/w2w")

device = "cuda"
pretrained_model_name_or_path = "stablediffusionapi/realistic-vision-v51" 
revision = None
weight_dtype = torch.bfloat16
# Load scheduler, tokenizer and models.
pipe = StableDiffusionPipeline.from_pretrained("stablediffusionapi/realistic-vision-v51", 
                                                torch_dtype=torch.float16,safety_checker = None,
                                                requires_safety_checker = False).to(device)
noise_scheduler = pipe.scheduler
del pipe
tokenizer = AutoTokenizer.from_pretrained(
            pretrained_model_name_or_path, subfolder="tokenizer", revision=revision
        )
text_encoder = CLIPTextModel.from_pretrained(
            pretrained_model_name_or_path, subfolder="text_encoder", revision=revision
        )
vae = AutoencoderKL.from_pretrained(pretrained_model_name_or_path, subfolder="vae", revision=revision)
unet = UNet2DConditionModel.from_pretrained(
            pretrained_model_name_or_path, subfolder="unet", revision=revision
        )
unet.requires_grad_(False)
unet.to(device, dtype=weight_dtype)
vae.requires_grad_(False)
    
text_encoder.requires_grad_(False)
vae.requires_grad_(False)
vae.to(device, dtype=weight_dtype)
text_encoder.to(device, dtype=weight_dtype)
print("")


mean = torch.load(f"{models_path}/files/mean.pt", map_location=torch.device('cpu')).bfloat16().to(device)
std = torch.load(f"{models_path}/files/std.pt", map_location=torch.device('cpu')).bfloat16().to(device)
v = torch.load(f"{models_path}/files/V.pt", map_location=torch.device('cpu')).bfloat16().to(device)
proj = torch.load(f"{models_path}/files/proj_1000pc.pt", map_location=torch.device('cpu')).bfloat16().to(device)
df = torch.load(f"{models_path}/files/identity_df.pt")
weight_dimensions = torch.load(f"{models_path}/files/weight_dimensions.pt")
pinverse = torch.load(f"{models_path}/files/pinverse_1000pc.pt", map_location=torch.device('cpu')).bfloat16().to(device)

        
young = get_direction(df, "Young", pinverse, 1000, device)
young = debias(young, "Male", df, pinverse, device)
young = debias(young, "Pointy_Nose", df, pinverse, device)
young = debias(young, "Wavy_Hair", df, pinverse, device)
young = debias(young, "Chubby", df, pinverse, device)
young = debias(young, "No_Beard", df, pinverse, device)
young = debias(young, "Mustache", df, pinverse, device)
    
pointy = get_direction(df, "Pointy_Nose", pinverse, 1000, device)
pointy = debias(pointy, "Young", df, pinverse, device)
pointy = debias(pointy, "Male", df, pinverse, device)
pointy = debias(pointy, "Wavy_Hair", df, pinverse, device)
pointy = debias(pointy, "Chubby", df, pinverse, device)
pointy = debias(pointy, "Heavy_Makeup", df, pinverse, device)
        
wavy = get_direction(df, "Wavy_Hair", pinverse, 1000, device)
wavy = debias(wavy, "Young", df, pinverse, device)
wavy = debias(wavy, "Male", df, pinverse, device)
wavy = debias(wavy, "Pointy_Nose", df, pinverse, device)
wavy = debias(wavy, "Chubby", df, pinverse, device)
wavy = debias(wavy, "Heavy_Makeup", df, pinverse, device)
        
thick = get_direction(df, "Bushy_Eyebrows", pinverse, 1000, device)
thick = debias(thick, "Male", df, pinverse, device)
thick = debias(thick, "Young", df, pinverse, device)
thick = debias(thick, "Pointy_Nose", df, pinverse, device)
thick = debias(thick, "Wavy_Hair", df, pinverse, device)
thick = debias(thick, "Mustache", df, pinverse, device)
thick = debias(thick, "No_Beard", df, pinverse, device)
thick = debias(thick, "Sideburns", df, pinverse, device)
thick = debias(thick, "Big_Nose", df, pinverse, device)
thick = debias(thick, "Big_Lips", df, pinverse, device)
thick = debias(thick, "Black_Hair", df, pinverse, device)
thick = debias(thick, "Brown_Hair", df, pinverse, device)
thick = debias(thick, "Pale_Skin", df, pinverse, device)
thick = debias(thick, "Heavy_Makeup", df, pinverse, device)



@torch.no_grad()
@spaces.GPU
def sample_then_run(net):
    device = "cuda"
    # get mean and standard deviation for each principal component
    m = torch.mean(proj, 0)
    standev = torch.std(proj, 0)
    
    # sample
    sample = torch.zeros([1, 10000]).to(device)

    sample_seed = random.randint(0, 100000)
    torch.manual_seed(sample_seed)
    
    #only first 1000 PCs
    for i in range(1000):
        sample[0, i] = torch.normal(m[i], standev[i], (1,1))

    
    net = "model_"+str(uuid.uuid4())[:4]+".pt"
    torch.save(sample, net)
    
    image = prompt = "sks person"
    negative_prompt = "low quality, blurry, unfinished, nudity, weapon"
    seed = 5
    cfg = 3.0
    steps = 25
    image = inference(net, prompt, negative_prompt, cfg, steps, seed)
    return net,net,image, None

@torch.no_grad()
@spaces.GPU()
def inference(net, prompt, negative_prompt, guidance_scale, ddim_steps, seed):
    device = "cuda"
    mean.to(device)
    std.to(device)
    v.to(device)
    
    weights = torch.load(net).to(device)
    network = LoRAw2w(weights, mean, std, v[:, :10000], 
                    unet,
                    rank=1,
                    multiplier=1.0,
                    alpha=27.0,
                    train_method="xattn-strict"
                ).to(device, torch.bfloat16)
        

        
    generator = torch.Generator(device=device).manual_seed(seed)
    latents = torch.randn(
                (1, unet.in_channels, 512 // 8, 512 // 8),
                generator = generator,
                device = device
            ).bfloat16()
           
        
    text_input = tokenizer(prompt, padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt")
        
    text_embeddings = text_encoder(text_input.input_ids.to(device))[0]
        
    max_length = text_input.input_ids.shape[-1]
    uncond_input = tokenizer(
                                    [negative_prompt], padding="max_length", max_length=max_length, return_tensors="pt"
                                )
    uncond_embeddings = text_encoder(uncond_input.input_ids.to(device))[0]
    text_embeddings = torch.cat([uncond_embeddings, text_embeddings]).bfloat16()
    noise_scheduler.set_timesteps(ddim_steps) 
    latents = latents * noise_scheduler.init_noise_sigma
            
    for i,t in enumerate(tqdm.tqdm(noise_scheduler.timesteps)):
        latent_model_input = torch.cat([latents] * 2)
        latent_model_input = noise_scheduler.scale_model_input(latent_model_input, timestep=t)
        with torch.no_grad():
            with network:
                noise_pred = unet(latent_model_input, t, encoder_hidden_states=text_embeddings, timestep_cond= None).sample
            
        #guidance
        noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
        noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
        latents = noise_scheduler.step(noise_pred, t, latents).prev_sample
            
    latents = 1 / 0.18215 * latents
    image = vae.decode(latents).sample
    image = (image / 2 + 0.5).clamp(0, 1)
    image = image.detach().cpu().float().permute(0, 2, 3, 1).numpy()[0]
        
    image = Image.fromarray((image * 255).round().astype("uint8"))

        
    return image




@torch.no_grad()
@spaces.GPU()
def edit_inference(net, prompt, negative_prompt, guidance_scale, ddim_steps, seed, start_noise, a1, a2, a3, a4, input_image): 
    device = "cuda"
    mean.to(device)
    std.to(device)
    v.to(device)
    

    
    weights = torch.load(net).to(device)
    network = LoRAw2w(weights, mean, std, v[:, :10000], 
                    unet,
                    rank=1,
                    multiplier=1.0,
                    alpha=27.0,
                    train_method="xattn-strict"
                ).to(device, torch.bfloat16)
    
 

    
            
    #pad to same number of PCs
    pcs_original = weights.shape[1]
    pcs_edits = young.shape[1]
    padding =  torch.zeros((1,pcs_original-pcs_edits)).to(device)
    young_pad = torch.cat((young.to(device), padding), 1)
    pointy_pad = torch.cat((pointy.to(device), padding), 1)
    wavy_pad = torch.cat((wavy.to(device), padding), 1)
    thick_pad = torch.cat((thick.to(device), padding), 1)
            
        
    edited_weights = weights+a1*1e6*young_pad+a2*1e6*pointy_pad+a3*1e6*wavy_pad+a4*2e6*thick_pad

    generator = torch.Generator(device=device).manual_seed(seed)
    latents = torch.randn(
                (1, unet.in_channels, 512 // 8, 512 // 8),
                generator = generator,
                device = device
            ).bfloat16()
           
        
    text_input = tokenizer(prompt, padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt")
        
    text_embeddings = text_encoder(text_input.input_ids.to(device))[0]
        
    max_length = text_input.input_ids.shape[-1]
    uncond_input = tokenizer(
                                    [negative_prompt], padding="max_length", max_length=max_length, return_tensors="pt"
                                )
    uncond_embeddings = text_encoder(uncond_input.input_ids.to(device))[0]
    text_embeddings = torch.cat([uncond_embeddings, text_embeddings]).bfloat16()
    noise_scheduler.set_timesteps(ddim_steps) 
    latents = latents * noise_scheduler.init_noise_sigma
           
        
            
    for i,t in enumerate(tqdm.tqdm(noise_scheduler.timesteps)):
        latent_model_input = torch.cat([latents] * 2)
        latent_model_input = noise_scheduler.scale_model_input(latent_model_input, timestep=t)
                
        if t>start_noise:
            pass
        elif t<=start_noise:
            network.proj = torch.nn.Parameter(edited_weights)
            network.reset()
        with torch.no_grad():
            with network:
                noise_pred = unet(latent_model_input, t, encoder_hidden_states=text_embeddings, timestep_cond= None).sample
                    
                
        #guidance
        noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
        noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
        latents = noise_scheduler.step(noise_pred, t, latents).prev_sample
            
    latents = 1 / 0.18215 * latents
    image = vae.decode(latents).sample
    image = (image / 2 + 0.5).clamp(0, 1)
    image = image.detach().cpu().float().permute(0, 2, 3, 1).numpy()[0] 
    image = Image.fromarray((image * 255).round().astype("uint8"))

    return net, image


class CustomImageDataset(Dataset):
    def __init__(self, images, transform=None):
        self.images = images
        self.transform = transform
        
    def __len__(self):
        return len(self.images)
    
    def __getitem__(self, idx):
        image = self.images[idx]
        if self.transform:
            image = self.transform(image)
        return image
        
@spaces.GPU(duration=180)
def invert(image, mask, pcs=10000, epochs=400, weight_decay = 1e-10, lr=1e-1):
    device = "cuda"
    mean.to(device)
    std.to(device)
    v.to(device)
    
            
    weights = torch.zeros(1,pcs).bfloat16().to(device)
    network = LoRAw2w( weights, mean, std, v[:, :pcs], 
                                unet,
                                rank=1,
                                multiplier=1.0,
                                alpha=27.0,
                                train_method="xattn-strict"
                            ).to(device, torch.bfloat16)
    
    ### load mask
    mask = transforms.Resize((64,64), interpolation=transforms.InterpolationMode.BILINEAR)(mask)
    mask = torchvision.transforms.functional.pil_to_tensor(mask).unsqueeze(0).to(device).bfloat16()[:,0,:,:].unsqueeze(1)
    ### check if an actual mask was draw, otherwise mask is just all ones
    if torch.sum(mask) == 0:
        mask = torch.ones((1,1,64,64)).to(device).bfloat16()
                
    ### single image dataset
    image_transforms = transforms.Compose([transforms.Resize(512, interpolation=transforms.InterpolationMode.BILINEAR),
                                                        transforms.RandomCrop(512),
                                                        transforms.ToTensor(),
                                                        transforms.Normalize([0.5], [0.5])])
        
        
    train_dataset = CustomImageDataset(image, transform=image_transforms)
    train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=1, shuffle=True) 
        
    ### optimizer 
    optim = torch.optim.Adam(network.parameters(), lr=lr, weight_decay=weight_decay)    
        
    ### training loop
    unet.train()
    for epoch in tqdm.tqdm(range(epochs)):
        for batch in train_dataloader:
            ### prepare inputs
            batch = batch.to(device).bfloat16()
            latents = vae.encode(batch).latent_dist.sample()
            latents = latents*0.18215
            noise = torch.randn_like(latents)
            bsz = latents.shape[0]
                 
            timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device)
            timesteps = timesteps.long()
            noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
            text_input = tokenizer("sks person", padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt")
            text_embeddings = text_encoder(text_input.input_ids.to(device))[0]
        
            ### loss + sgd step
            with network:
                model_pred = unet(noisy_latents, timesteps, text_embeddings).sample
                loss = torch.nn.functional.mse_loss(mask*model_pred.float(), mask*noise.float(), reduction="mean")
                optim.zero_grad()
                loss.backward()
                optim.step()


    #pad to 10000 PCs
    pcs_original = weights.shape[1]
    padding =  torch.zeros((1,10000-pcs_original)).to(device)
    weights = network.proj.detach()
    weights = torch.cat((weights, padding), 1)

    net = "model_"+str(uuid.uuid4())[:4]+".pt"
    torch.save(weights, net)
    
    return net


@spaces.GPU(duration=180)
def run_inversion(net, dict, pcs, epochs, weight_decay,lr):
    init_image = dict["background"].convert("RGB").resize((512, 512))
    mask = dict["layers"][0].convert("RGB").resize((512, 512))
    
    net = invert([init_image], mask, pcs, epochs, weight_decay,lr)
    
    #sample an image
    prompt = "sks person"
    negative_prompt = "low quality, blurry, unfinished, nudity, weapon"
    seed = 5
    cfg = 3.0
    steps = 25
    image = inference( net, prompt, negative_prompt, cfg, steps, seed)
    
    return net, net, image, None
    
    
@spaces.GPU
def file_upload(file, net):
    device="cuda"
    weights = torch.load(file.name).to(device)
        
    #pad to 10000 Principal components to keep everything consistent
    pcs = weights.shape[1]
    padding =  torch.zeros((1,10000-pcs)).to(device)
    weights = torch.cat((weights, padding), 1)
    net = "model_"+str(uuid.uuid4())[:4]+".pt"
    torch.save(weights, net)
       
    image = prompt = "sks person"
    negative_prompt = "low quality, blurry, unfinished, nudity, weapon"
    seed = 5
    cfg = 3.0
    steps = 25
    image = inference(net, prompt, negative_prompt, cfg, steps, seed)
    return net, image, None
    
help_text1 = """
            <b>Instructions</b>:
            1. To get results faster without waiting for a GPU, you can duplicate into a private space with an A100 GPU.
            2. To begin, you will have to get an identity-encoding model. You can either sample one from *weights2weights* space by clicking `🎲 Sample new identity` or by uploading an image and clicking `⬆️ Insert identity` to insert the identity into a model. You can optionally draw over the head to define a mask in the image for better results. Sampling a model takes around 10 seconds and inserting an identity takes around 15 seconds per 100 epochs. After this is done, you can optionally download this model for later use. A model can be uploaded in the \"Uploading a model\" tab in the `Advanced Options`.   
            3. After getting a model, an image of the identity will be displayed on the left as a reference. You can sample from the model by changing seeds as well as prompts and then clicking `🖼️ Generate` to generate images in the right column. Make sure to include \"sks person\" in your prompt to keep the same identity.
            4. The identity in the model can be edited by changing the sliders for various attributes. After clicking `🖼️ Generate`, you can see how the identity has changed and the effects are maintained across different seeds and prompts. 
            """
help_text2 = """<b>Tips</b>:
            1. Identity Editing and Generation
                * If you are interested in preserving more of the image during identity-editing (i.e., where the same seed and prompt results in the same image with only the identity changed), you can play with the "Injection Step" parameter in the \"Editing and Generation\" tab in the `Advanced Options`. 
                    * During the first *n* timesteps, the original model's weights will be used, and then the edited weights will be set during the remaining steps. Values closer to 1000 will set the edited weights early, having a more pronounced effect, which may disrupt some semantics and structure of the generated image. Lower values will set the edited weights later, better preserving image context. We notice that around 600-800 tends to produce the best results. Larger values in the range (700-1000) are helpful for more global attribute changes, while smaller (400-700) can be used for more finegrained edits. Although it is not always needed.
                * You can play around with negative prompts, number of inference steps, and CFG in the \"Editing and Generation\" tab in the `Advanced Options` to affect the ultimate image quality.
                * Sometimes the identity will not be perfectly consistent (e.g., there might be small variations of the face) when you use some seeds or prompts. This is a limitation of our method as well as an open-problem in personalized models. 
            2. Inserting Identity
                * To obtain the best results for inserting an identity, upload a high resolution photo of the face with minimal occlusion. It is recommended to draw over the face and hair to define a mask. But inversion should still work generally for non-closeup face shots.
                * For inserting a realistic photo of an identity, typically 800 epochs with lr=1e-1 and 10,000 principal components (PCs) works well. If the resulting generations have artifacted and unrealstic textures, there is probably overfitting and you may want to reduce the number of epochs or learning rate, or play with weight decay. If the generations do not look like the identity in the input photo, then you may want to increase the number of epochs.
                * For inserting out-of-distribution identities, such as artistic renditions of people or non-humans (e.g. the ones shown in the paper), it is recommended to use 1000 PCs, lr=1 or 2, and train for 400-800 epochs.
                * Note that if you change the number of PCs, you will probably need to change the learning rate. For less PCs, higher learning rates are typically required."""     

    
intro = """
<div style="display: flex;align-items: center;justify-content: center">
        <h1 style="margin-left: 12px;text-align: center;margin-bottom: 7px;display: inline-block"><em>weights2weights</em> Demo</h1>
        <h3 style="display: inline-block;margin-left: 10px;margin-top: 12px;font-weight: 500">(from "Interpreting the Weight Space of Customized Diffusion Models")</h3>
    </div>
    <p style="font-size: 0.95rem;margin: 0rem;line-height: 1.2em;margin-top:1em;display: inline-block">
        <a href="https://snap-research.github.io/weights2weights/" target="_blank">Project Page</a> | <a href="https://arxiv.org/abs/2406.09413" target="_blank">Paper</a>
         |  <a href="https://github.com/snap-research/weights2weights" target="_blank">Code</a> |
         <a href="https://huggingface.co/snap-research/weights2weights" target="_blank">Model Card</a> |
        <a href="https://huggingface.co/spaces/Snapchat/w2w-demo?duplicate=true" target="_blank" style="
            display: inline-block;
        ">
        <img style="margin-top: -1em;margin-bottom: 0em;position: absolute;" src="https://bit.ly/3CWLGkA" alt="Duplicate Space"></a>
    </p>
    """
    
    
    
with gr.Blocks(css="style.css") as demo:
    net = gr.State()  
    gr.HTML(intro)
        
    gr.Markdown(""" **Getting Started:** Sample a random identity or insert the identity from an image to get a model (or upload a previously downloaded model using the `Uploading a model` option in `Advanced Options`). 
    **What Can You Do?** Generate new images of the identity & edit the identity 👩->🧑‍🦳. **See further instructions and tips at the bottom of the page.**""")
    with gr.Column():
            with gr.Row():
                with gr.Column(): 
                    gr.Markdown("""❶ Either a) sample a random identity by clicking `🎲 Sample random identity` **or** b) upload an image (optional - draw a mask over the head) then click `⬆️ Insert identity`""")
                    input_image = gr.ImageEditor(elem_id="image_upload", type='pil', label="Reference Identity",
                                               width=512, height=512)    

                   
                   
                    with gr.Row():
                        sample = gr.Button("🎲 Sample new identity")
                        invert_button = gr.Button("⬆️ Insert identity")
    
    
    
                with gr.Column():
                    gr.Markdown("""❷ Generate new images of the identity from the model by entering different seeds and prompts and clicking `🖼️ Generate`. Edit the encoded identity with the sliders and then visualize the edits with the generate button.""")
                    gallery = gr.Image(label="Generated Image",height=512, width=512, interactive=False)
                    submit = gr.Button("🖼️ Generate")
    
                        
            prompt = gr.Textbox(label="Prompt",
                                                info="Make sure to include 'sks person'" ,
                                                placeholder="sks person", 
                                                value="sks person")
            seed = gr.Number(value=5, label="Seed", precision=0, interactive=True)
                
                    
            # Editing 
            with gr.Column():
                with gr.Row():
                    a1 = gr.Slider(label="- Young +", value=0, step=0.001, minimum=-1, maximum=1, interactive=True)
                    a2 = gr.Slider(label="- Pointy Nose +", value=0, step=0.001, minimum=-1, maximum=1, interactive=True)       
                with gr.Row():        
                    a3 = gr.Slider(label="- Curly Hair +", value=0, step=0.001, minimum=-1, maximum=1, interactive=True)
                    a4 = gr.Slider(label="- Thick Eyebrows +", value=0, step=0.001, minimum=-1, maximum=1, interactive=True)
    
                
            with gr.Accordion("Advanced Options ⬇️", open=False):
                with gr.Tab("Editing and Generation"):
                    with gr.Row(): 
                        cfg= gr.Slider(label="CFG", value=3.0, step=0.1, minimum=0, maximum=10, interactive=True)
                        steps = gr.Slider(label="Inference Steps",  value=25, step=1, minimum=0, maximum=100, interactive=True)
                    with gr.Row():
                        negative_prompt = gr.Textbox(label="Negative Prompt", placeholder="low quality, blurry, unfinished, nudity, weapon", value="low quality, blurry, unfinished, nudity, weapon")
                        injection_step = gr.Slider(label="Injection Step",  value=800, step=1, minimum=0, maximum=1000, interactive=True)
                                
                with gr.Tab("Inserting Identity"):
                    with gr.Row():     
                        lr = gr.Number(value=1e-1, label="Learning Rate", interactive=True)
                        pcs = gr.Slider(label="# Principal Components", value=10000, step=1, minimum=1, maximum=10000, interactive=True)
                    with gr.Row():
                        epochs = gr.Slider(label="Epochs", value=800, step=1, minimum=1, maximum=2000, interactive=True)
                        weight_decay = gr.Number(value=1e-10, label="Weight Decay", interactive=True)
                
                with gr.Tab("Uploading a model"):
                    gr.Markdown("""<div style="text-align: justify;">Upload a model below downloaded from this demo.""")
    
                    file_input = gr.File(label="Upload Model", container=True)
    
         
    with gr.Accordion("💡Instructions & Tips", open=False):
        gr.Markdown(help_text1)
        gr.Markdown(help_text2)
    
    gr.Markdown("""<div style="text-align: justify;">❸ After sampling a new model or inserting an identity, you can download the model below.""")
    
    with gr.Row():
        file_output = gr.File(label="Download Model", container=True, interactive=False)
            
    
    
    
        invert_button.click(fn=run_inversion,
                        inputs=[net, input_image, pcs, epochs, weight_decay,lr], 
                        outputs = [net, file_output, input_image, gallery])
            
                
        sample.click(fn=sample_then_run,inputs = [net], outputs=[net, file_output, input_image, gallery])
            
        submit.click(
                fn=edit_inference, inputs=[net, prompt, negative_prompt, cfg, steps, seed, injection_step, a1, a2, a3, a4, input_image], outputs=[net, gallery]
            )
        file_input.change(fn=file_upload, inputs=[file_input, net], outputs = [net, input_image, gallery])
    
    

    

        
demo.queue().launch()