app.py

import gradio as gr
import numpy as np
import torch
import requests 
import random
import os
import sys
import pickle
from PIL import Image

from tqdm.auto import tqdm
from datetime import datetime

import diffusers
from diffusers import DDIMScheduler
from transformers import CLIPTextModel, CLIPTokenizer
import torch.nn.functional as F

from utils import preprocess_mask, process_sketch, process_prompts, process_example


#################################################
#################################################
canvas_html = "<div id='canvas-root' style='max-width:400px; margin: 0 auto'></div>"
load_js = """
async () => {
const url = "https://huggingface.co/datasets/radames/gradio-components/raw/main/sketch-canvas.js"
fetch(url)
  .then(res => res.text())
  .then(text => {
    const script = document.createElement('script');
    script.type = "module"
    script.src = URL.createObjectURL(new Blob([text], { type: 'application/javascript' }));
    document.head.appendChild(script);
  });
}
"""

get_js_colors = """
async (canvasData) => {
  const canvasEl = document.getElementById("canvas-root");
  return [canvasEl._data]
}
"""

css = '''
#color-bg{display:flex;justify-content: center;align-items: center;}
.color-bg-item{width: 100%; height: 32px}
#main_button{width:100%}
<style>
'''


#################################################
#################################################
global sreg, creg, sizereg, COUNT, creg_maps, sreg_maps, pipe, text_cond

sreg = 0
creg = 0
sizereg = 0
COUNT = 0
reg_sizes = {}
creg_maps = {}
sreg_maps = {}
text_cond = 0
device="cuda"
MAX_COLORS = 12

pipe = diffusers.StableDiffusionPipeline.from_pretrained(
        "runwayml/stable-diffusion-v1-5",
        variant="fp16").to(device)

pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
pipe.scheduler.set_timesteps(50)
timesteps = pipe.scheduler.timesteps
sp_sz = pipe.unet.sample_size

with open('./valset.pkl', 'rb') as f:
    val_prompt = pickle.load(f)


#################################################
#################################################
def mod_forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None, temb=None):

    residual = hidden_states

    if self.spatial_norm is not None:
        hidden_states = self.spatial_norm(hidden_states, temb)

    input_ndim = hidden_states.ndim

    if input_ndim == 4:
        batch_size, channel, height, width = hidden_states.shape
        hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)

    batch_size, sequence_length, _ = (hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape)
    attention_mask = self.prepare_attention_mask(attention_mask, sequence_length, batch_size)

    if self.group_norm is not None:
        hidden_states = self.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

    query = self.to_q(hidden_states)

    global sreg, creg, COUNT, creg_maps, sreg_maps, reg_sizes, text_cond
    
    sa_ = True if encoder_hidden_states is None else False
    encoder_hidden_states = text_cond if encoder_hidden_states is not None else hidden_states
        
    if self.norm_cross:
        encoder_hidden_states = self.norm_encoder_hidden_states(encoder_hidden_states)

    key = self.to_k(encoder_hidden_states)
    value = self.to_v(encoder_hidden_states)

    query = self.head_to_batch_dim(query)
    key = self.head_to_batch_dim(key)
    value = self.head_to_batch_dim(value)
    
    if COUNT/32 < 50*0.3:
        
        dtype = query.dtype
        if self.upcast_attention:
            query = query.float()
            key = key.float()
            
        sim = torch.baddbmm(torch.empty(query.shape[0], query.shape[1], key.shape[1], 
                                        dtype=query.dtype, device=query.device),
                            query, key.transpose(-1, -2), beta=0, alpha=self.scale)
        
        treg = torch.pow(timesteps[COUNT//32]/1000, 5)
        
        ## reg at self-attn
        if sa_:
            min_value = sim[int(sim.size(0)/2):].min(-1)[0].unsqueeze(-1)
            max_value = sim[int(sim.size(0)/2):].max(-1)[0].unsqueeze(-1)  
            mask = sreg_maps[sim.size(1)].repeat(self.heads,1,1)
            size_reg = reg_sizes[sim.size(1)].repeat(self.heads,1,1)
            
            sim[int(sim.size(0)/2):] += (mask>0)*size_reg*sreg*treg*(max_value-sim[int(sim.size(0)/2):])
            sim[int(sim.size(0)/2):] -= ~(mask>0)*size_reg*sreg*treg*(sim[int(sim.size(0)/2):]-min_value)
            
        ## reg at cross-attn
        else:
            min_value = sim[int(sim.size(0)/2):].min(-1)[0].unsqueeze(-1)
            max_value = sim[int(sim.size(0)/2):].max(-1)[0].unsqueeze(-1)  
            mask = creg_maps[sim.size(1)].repeat(self.heads,1,1)
            size_reg = reg_sizes[sim.size(1)].repeat(self.heads,1,1)
            
            sim[int(sim.size(0)/2):] += (mask>0)*size_reg*creg*treg*(max_value-sim[int(sim.size(0)/2):])
            sim[int(sim.size(0)/2):] -= ~(mask>0)*size_reg*creg*treg*(sim[int(sim.size(0)/2):]-min_value)

        attention_probs = sim.softmax(dim=-1)
        attention_probs = attention_probs.to(dtype)
            
    else:
        attention_probs = self.get_attention_scores(query, key, attention_mask)
           
    COUNT += 1
            
    hidden_states = torch.bmm(attention_probs, value)
    hidden_states = self.batch_to_head_dim(hidden_states)

    # linear proj
    hidden_states = self.to_out[0](hidden_states)
    # dropout
    hidden_states = self.to_out[1](hidden_states)

    if input_ndim == 4:
        hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)

    if self.residual_connection:
        hidden_states = hidden_states + residual

    hidden_states = hidden_states / self.rescale_output_factor

    return hidden_states

for _module in pipe.unet.modules():
    if _module.__class__.__name__ == "Attention":
        _module.__class__.__call__ = mod_forward

        
#################################################
#################################################
def process_generation(binary_matrixes, seed, creg_, sreg_, sizereg_, bsz, master_prompt, *prompts):

    global creg, sreg, sizereg
    creg, sreg, sizereg = creg_, sreg_, sizereg_
    
    clipped_prompts = prompts[:len(binary_matrixes)]
    prompts = [master_prompt] + list(clipped_prompts)
    layouts = torch.cat([preprocess_mask(mask_, sp_sz, sp_sz, device) for mask_ in binary_matrixes])
    
    text_input = pipe.tokenizer(prompts, padding="max_length", return_length=True, return_overflowing_tokens=False, 
                                max_length=pipe.tokenizer.model_max_length, truncation=True, return_tensors="pt")
    cond_embeddings = pipe.text_encoder(text_input.input_ids.to(device))[0]

    uncond_input = pipe.tokenizer([""]*bsz, padding="max_length", max_length=pipe.tokenizer.model_max_length,
                                  truncation=True, return_tensors="pt")
    uncond_embeddings = pipe.text_encoder(uncond_input.input_ids.to(device))[0]

    
    ###########################
    ###### prep for sreg ###### 
    ###########################
    global sreg_maps, reg_sizes
    sreg_maps = {}
    reg_sizes = {}
    
    for r in range(4):
        res = int(sp_sz/np.power(2,r))
        layouts_s = F.interpolate(layouts,(res, res),mode='nearest')
        layouts_s = (layouts_s.view(layouts_s.size(0),1,-1)*layouts_s.view(layouts_s.size(0),-1,1)).sum(0).unsqueeze(0).repeat(bsz,1,1)
        reg_sizes[np.power(res, 2)] = 1-sizereg*layouts_s.sum(-1, keepdim=True)/(np.power(res, 2))
        sreg_maps[np.power(res, 2)] = layouts_s


    ###########################
    ###### prep for creg ######
    ###########################
    pww_maps = torch.zeros(1,77,sp_sz,sp_sz).to(device)
    for i in range(1,len(prompts)):
        wlen = text_input['length'][i] - 2
        widx = text_input['input_ids'][i][1:1+wlen]
        for j in range(77):
            try:
                if (text_input['input_ids'][0][j:j+wlen] == widx).sum() == wlen:
                    pww_maps[:,j:j+wlen,:,:] = layouts[i-1:i]
                    cond_embeddings[0][j:j+wlen] = cond_embeddings[i][1:1+wlen]
                    break
            except:
                raise gr.Error("Please check whether every segment prompt is included in the full text !")
                return
    
    global creg_maps
    creg_maps = {}
    for r in range(4):
        res = int(sp_sz/np.power(2,r))
        layout_c = F.interpolate(pww_maps,(res,res),mode='nearest').view(1,77,-1).permute(0,2,1).repeat(bsz,1,1)
        creg_maps[np.power(res, 2)] = layout_c


    ###########################    
    #### prep for text_emb ####
    ###########################
    global text_cond
    text_cond = torch.cat([uncond_embeddings, cond_embeddings[:1].repeat(bsz,1,1)])    
    
    global COUNT
    COUNT = 0
    
    if seed == -1:
        latents = torch.randn(bsz,4,sp_sz,sp_sz).to(device)
    else:
        latents = torch.randn(bsz,4,sp_sz,sp_sz, generator=torch.Generator().manual_seed(seed)).to(device)
        
    image = pipe(prompts[:1]*bsz, latents=latents).images

    return(image)


#################################################
#################################################
### define the interface
with gr.Blocks(css=css) as demo:
    binary_matrixes = gr.State([])
    color_layout = gr.State([])
    gr.Markdown('''## DenseDiffusion: Dense Text-to-Image Generation with Attention Modulation''')
    gr.Markdown('''
    #### 😺 Instruction to generate images 😺 <br>
    (1) Create the image layout. <br>
    (2) Label each segment with a text prompt. <br>
    (3) Adjust the full text. The default full text is automatically concatenated from each segment's text. The default one works well, but refineing the full text will further improve the result. <br>
    (4) Check the generated images, and tune the hyperparameters if needed. <br>
    &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; - w<sup>c</sup> : The degree of attention modulation at cross-attention layers. <br>
    &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; - w<sup>s</sup> : The degree of attention modulation at self-attention layers. <br>
    ''')
    
    with gr.Row():
        with gr.Box(elem_id="main-image"):
            canvas_data = gr.JSON(value={}, visible=False)
            canvas = gr.HTML(canvas_html)
            button_run = gr.Button("(1) I've finished my sketch ! 😺", elem_id="main_button", interactive=True)
      
            prompts = []
            colors = []
            color_row = [None] * MAX_COLORS
            with gr.Column(visible=False) as post_sketch:
                for n in range(MAX_COLORS):
                    if n == 0 :
                        with gr.Row(visible=False) as color_row[n]:
                            colors.append(gr.Image(shape=(100, 100), label="background", type="pil", image_mode="RGB", width=100, height=100))
                            prompts.append(gr.Textbox(label="Prompt for the background (white region)", value=""))
                    else:
                        with gr.Row(visible=False) as color_row[n]:
                            colors.append(gr.Image(shape=(100, 100), label="segment "+str(n), type="pil", image_mode="RGB", width=100, height=100))
                            prompts.append(gr.Textbox(label="Prompt for the segment "+str(n)))
                    
                get_genprompt_run = gr.Button("(2) I've finished segment labeling ! 😺", elem_id="prompt_button", interactive=True)
                
            with gr.Column(visible=False) as gen_prompt_vis:
                general_prompt = gr.Textbox(value='', label="(3) Textual Description for the entire image", interactive=True)
                with gr.Accordion("(4) Tune the hyperparameters", open=False):
                    creg_ = gr.Slider(label=" w\u1D9C (The degree of attention modulation at cross-attention layers) ", minimum=0, maximum=2., value=1.0, step=0.1)
                    sreg_ = gr.Slider(label=" w \u02E2 (The degree of attention modulation at self-attention layers) ", minimum=0, maximum=2., value=0.3, step=0.1)
                    sizereg_ = gr.Slider(label="The degree of mask-area adaptive adjustment", minimum=0, maximum=1., value=1., step=0.1)
                    bsz_ = gr.Slider(label="Number of Samples to generate", minimum=1, maximum=4, value=1, step=1)
                    seed_ = gr.Slider(label="Seed", minimum=-1, maximum=999999999, value=-1, step=1)
                    
                final_run_btn = gr.Button("Generate ! 😺")
                
                layout_path = gr.Textbox(label="layout_path", visible=False)
                all_prompts = gr.Textbox(label="all_prompts", visible=False)
                
        with gr.Column():
            out_image = gr.Gallery(label="Result", columns=2, height='auto')
            
    button_run.click(process_sketch, inputs=[canvas_data], outputs=[post_sketch, binary_matrixes, *color_row, *colors], _js=get_js_colors, queue=False)
    
    get_genprompt_run.click(process_prompts, inputs=[binary_matrixes, *prompts], outputs=[gen_prompt_vis, general_prompt], queue=False)
    
    final_run_btn.click(process_generation, inputs=[binary_matrixes, seed_, creg_, sreg_, sizereg_, bsz_, general_prompt, *prompts], outputs=out_image)
    
    gr.Examples(
        examples=[['0.png', '***'.join([val_prompt[0]['textual_condition']] + val_prompt[0]['segment_descriptions']), 381940206],
                  ['1.png', '***'.join([val_prompt[1]['textual_condition']] + val_prompt[1]['segment_descriptions']), 307504592],
                  ['5.png', '***'.join([val_prompt[5]['textual_condition']] + val_prompt[5]['segment_descriptions']), 114972190]],
        inputs=[layout_path, all_prompts, seed_],
        outputs=[post_sketch, binary_matrixes, *color_row, *colors, *prompts, gen_prompt_vis, general_prompt, seed_],
        fn=process_example,
        run_on_click=True,
        label='😺 Examples 😺',
    )
    
    demo.load(None, None, None, _js=load_js)
    
demo.launch(debug=True)