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 tqdm
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 utils import load_models
from editing import get_direction, debias
from sampling import sample_weights
from lora_w2w import LoRAw2w
from huggingface_hub import snapshot_download
import spaces

global device
global generator 
global unet
global vae 
global text_encoder
global tokenizer
global noise_scheduler
global network
device = "cuda"
#generator = torch.Generator(device=device)

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

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)

unet, vae, text_encoder, tokenizer, noise_scheduler = load_models(device)


global young
global pointy
global wavy 
global thick

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)

def sample_model():
    global unet
    del unet
    global network
    mean.to(device)
    std.to(device)
    v.to(device)
    proj.to(device)
    unet, _, _, _, _ = load_models(device)
    network = sample_weights(unet, proj, mean, std, v[:, :1000], device, factor = 1.00)
 
@torch.no_grad()
@spaces.GPU
def inference( prompt, negative_prompt, guidance_scale, ddim_steps, seed):
    global device
    #global generator 
    global unet
    global vae 
    global text_encoder
    global tokenizer
    global noise_scheduler
    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])
    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 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(prompt, negative_prompt, guidance_scale, ddim_steps, seed, start_noise, a1, a2, a3, a4):
    global device
    #global generator 
    global network
    global unet
    global vae 
    global text_encoder
    global tokenizer
    global noise_scheduler
    global young
    global pointy
    global wavy
    global thick
    
    original_weights = network.proj.clone()
    
    #pad to same number of PCs
    pcs_original = original_weights.shape[1]
    pcs_edits = young.shape[1]
    padding =  torch.zeros((1,pcs_original-pcs_edits)).to(device)
    young_pad = torch.cat((young, padding), 1)
    pointy_pad = torch.cat((pointy, padding), 1)
    wavy_pad = torch.cat((wavy, padding), 1)
    thick_pad = torch.cat((thick, padding), 1)
    

    edited_weights = original_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])
    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 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"))

    #reset weights back to original 
    network.proj = torch.nn.Parameter(original_weights)
    network.reset()

    return image

@spaces.GPU
def sample_then_run():
    sample_model()    
    prompt = "sks person"
    negative_prompt = "low quality, blurry, unfinished, nudity, weapon"
    seed = 5
    cfg = 3.0
    steps = 25
    image = inference( prompt, negative_prompt, cfg, steps, seed)
    torch.save(network.proj, "model.pt" )
    return image, "model.pt"

#@spaces.GPU
def start_items():
    print("Starting items")
    global young
    global pointy
    global wavy 
    global thick
    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)

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
def invert(image, mask, pcs=10000, epochs=400, weight_decay = 1e-10, lr=1e-1):
    global unet
    del unet
    global network
    unet, _, _, _, _ = load_models(device)
    
    proj = torch.zeros(1,pcs).bfloat16().to(device)
    network = LoRAw2w( proj, 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()

    ### return optimized network
    return network


@spaces.GPU
def run_inversion(dict, pcs, epochs, weight_decay,lr):
    global network
    init_image = dict["image"].convert("RGB").resize((512, 512))
    mask = dict["mask"].convert("RGB").resize((512, 512))
    network = invert([init_image], mask, pcs, epochs, weight_decay,lr)


    #sample an image
    prompt = "sks person"
    negative_prompt = "low quality, blurry, unfinished, nudity"
    seed = 5
    cfg = 3.0
    steps = 25
    image = inference( prompt, negative_prompt, cfg, steps, seed)
    torch.save(network.proj, "model.pt" )
    return image, "model.pt"
    
    
@spaces.GPU
def file_upload(file):
    global unet
    del unet
    global network
    global device

    
    
    proj = torch.load(file.name).to(device)

    #pad to 10000 Principal components to keep everything consistent
    pcs = proj.shape[1]
    padding =  torch.zeros((1,10000-pcs)).to(device)
    proj = torch.cat((proj, padding), 1)

    unet, _, _, _, _ = load_models(device)


    network = LoRAw2w( proj, mean, std, v[:, :10000], 
                        unet,
                        rank=1,
                        multiplier=1.0,
                        alpha=27.0,
                        train_method="xattn-strict"
                    ).to(device, torch.bfloat16)
        
    
    prompt = "sks person"
    negative_prompt = "low quality, blurry, unfinished, nudity"
    seed = 5
    cfg = 3.0
    steps = 25
    image = inference( prompt, negative_prompt, cfg, steps, seed)
    return image
    
    

    
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: 6px;font-weight: 500">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./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:
    gr.HTML(intro)
    
    gr.Markdown("""<div style="text-align: justify;"> In this demo, you can get an identity-encoding model by sampling or inverting. To use a model previously downloaded from this demo see \"Uploading a model\" in the Advanced Options. Next, you can generate new images from it, or edit the identity encoded in the model and generate images from the edited model. We provide detailed instructions and tips at the bottom of the page.""")
    with gr.Column():
            with gr.Row():
                with gr.Column(): 
                    gr.Markdown("""1) Either sample a new model, or upload an image (optionally draw a mask over the head) and click `invert`.""")
                    sample = gr.Button("🎲 Sample New Model")
                    input_image = gr.ImageEditor(elem_id="image_upload", type='pil', label="Reference Identity",
                                           width=512, height=512)    
                  
                    with gr.Row():
                        invert_button = gr.Button("⬆️ Invert")



                with gr.Column():
                    gr.Markdown("""2) Generate images of the sampled/inverted identity or edit the identity with the sliders and generate new images with various prompts and seeds.""")
                    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("Inversion"):
                    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("Sampling"):
                    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("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)

     


    gr.Markdown("""<div style="text-align: justify;"> After sampling a new model or inverting, you can download the model below.""")

    with gr.Row():
        file_output = gr.File(label="Download Sampled/Inverted Model", container=True, interactive=False)
        



        invert_button.click(fn=run_inversion,
                    inputs=[input_image, pcs, epochs, weight_decay,lr], 
                    outputs = [input_image, file_output])
        
            
        sample.click(fn=sample_then_run, outputs=[input_image, file_output])
        
        submit.click(
            fn=edit_inference, inputs=[prompt, negative_prompt, cfg, steps, seed, injection_step, a1, a2, a3, a4], outputs=[gallery]
        )
        file_input.change(fn=file_upload, inputs=file_input, outputs = gallery)



    help_text1 = """
        <b>Instructions</b>:
        1. To get results faster without waiting in queue, 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 Model` or by uploading an image and clicking `invert` to invert 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 inversion takes around 2 minutes. 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 right. You can sample from the model by changing seeds as well as prompts and then clicking `Generate`. 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. Editing and Identity 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 \"Sampling\" 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 \"Sampling\" 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. Inversion 
            * To obtain the best results for inversion, 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 inverting 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 input photo, then you may want to increase the number of epochs.
            * For inverting 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, and train for 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.""" 
    
        
    gr.Markdown(help_text1)
    gr.Markdown(help_text2)
    #demo.load(fn=start_items)
demo.queue().launch()