app.py

from asyncio import constants
import gradio as gr
import requests
import os 
import re
import random
from words import *
from base64 import b64decode
from PIL import Image
import io
import numpy as np


# GPT-J-6B API
API_URL = "https://api-inference.huggingface.co/models/EleutherAI/gpt-j-6B"
#HF_TOKEN = os.environ["HF_TOKEN"]
#headers = {"Authorization": f"Bearer {HF_TOKEN}"}

prompt = """

Bilbo is a hobbit rogue who wears a brown cloak and carries a ring.



Bremen is a human wizard, he wears a blue robe and carries a wand.

"""

examples = [["river"], ["night"], ["trees"],["table"],["laughs"]]


def npc_randomize():
    #name is a random combination of syllables
    name =""
    for i in range(random.randint(2,4)):
        name += random.choice(constants)
        name += random.choice(vowels)
        if random.random()<0.5:
            name += random.choice(constants)
        if random.random()<0.1:
            name += random.choice(seperators)
    #capitalize first letter
    name = name[0].upper() + name[1:]
    race=random.choice(races)
    characterClass=random.choice(classes)
    pronoun=random.choices(["he","she","they"],weights=[0.45,0.45,0.1],k=1)[0]
    return name,race,characterClass,pronoun


def genericDescription():
    
    desc=" wears a {color} {outfit}".format(color=random.choice(colors),outfit=random.choice(outfits))
    if random.random()<0.5:
        desc+=" and a {color} {outfit}".format(color=random.choice(colors),outfit=random.choice(outfits))
    
    if random.random()<0.5:
        desc+=" and carries a {weapon}".format(weapon=random.choice(weapons))
    elif random.random()<0.5:
        desc+=" and carries a {weapon} and a {object}".format(weapon=random.choice(weapons),object=random.choice(objects))
    else:
        desc+=" and carries two {weapon}s".format(weapon=random.choice(weapons))
        
    return desc


def npc_generate(name,race,characterClass,pronoun):

  desc="{name} is a {race} {characterClass}, {pronoun}".format(name=name,race=race,characterClass=characterClass,pronoun=pronoun)

  p = prompt + "\n"+desc
  print(f"*****Inside desc_generate - Prompt is :{p}")
  json_ = {"inputs": p,
            "parameters":
            {
            "top_p": 0.9,
          "temperature": 1.1,
          "max_new_tokens": 50,
          "return_full_text": False,
          }}
  #response = requests.post(API_URL, headers=headers, json=json_)
  response = requests.post(API_URL, json=json_)
  output = response.json()
  print(f"If there was an error? Reason is : {output}")


  #error handling
  if "error" in output:
    print("using fallback description method!")
    #fallback method
    longDescription=genericDescription()
  else:
    output_tmp = output[0]['generated_text']
    print(f"GPTJ response without splits is: {output_tmp}")
    if "\n\n" not in output_tmp:
        if output_tmp.find('.') != -1:
            idx = output_tmp.find('.')
            longDescription = output_tmp[:idx+1]
        else:
            idx = output_tmp.rfind('\n')
            longDescription = output_tmp[:idx]
    else:
        longDescription = output_tmp.split("\n\n")[0] # +"."
    longDescription = longDescription.replace('?','')
    print(f"longDescription being returned is: {longDescription}")
  return desc+longDescription

def desc_to_image(desc):
  print("*****Inside desc_to_image")
  desc = " ".join(desc.split('\n'))
  desc = desc + ", character art, concept art, artstation"
  steps, width, height, images, diversity = '50','256','256','1',15
  iface = gr.Interface.load("spaces/multimodalart/latentdiffusion")
  print("about to die",iface,dir(iface))

  prompt = re.sub(r'[^a-zA-Z0-9 ,.]', '', desc)
  print("about to die",prompt)


  img=iface(desc, steps, width, height, images, diversity)[0]
  return img

def desc_to_image_dalle(desc):
  print("*****Inside desc_to_image")
  desc = " ".join(desc.split('\n'))
  desc = desc + ", character art, concept art, artstation"
  steps, width, height, images, diversity = '50','256','256','1',15
  #iface = gr.Interface.load("huggingface/flax-community/dalle-mini")#this isn't a real interface
  iface = gr.Interface.load("spaces/multimodalart/rudalle")
  print("about to die",iface,dir(iface))

  prompt = re.sub(r'[^a-zA-Z0-9 ,.]', '', desc)
  print("about to die",prompt)

  model='Realism'
  aspect_ratio = 'Square'


  #img=iface(desc,model,aspect_ratio)[0]
  result=iface(desc,"Square","Realism")
  print(f"result is: {result}")
  return result[0]


def desc_to_image_cf(desc):
  cf = gr.Interface.load("spaces/Gradio-Blocks/clip-guided-faces")

  print("about to die",cf)

  text=desc
  init_image=None 
  skip_timesteps=0
  clip_guidance_scale=600
  tv_scale=0
  range_scale=0
  init_scale=0
  seed=0
  image_prompts=None
  timestep_respacing= 25
  cutn=16
  im_prompt_weight =1
  result = cf.fns[0].fn(text, init_image, skip_timesteps, clip_guidance_scale, tv_scale, range_scale, init_scale, seed, image_prompts,timestep_respacing, cutn, im_prompt_weight)
  

  #convert result from dataurl to image
  img=result[0]
  header, encoded = img.split(",", 1)
  data = b64decode(encoded)
  image = Image.open(io.BytesIO(data))
  image_np = np.array(image)
  return image_np


demo = gr.Blocks()

with demo:
  gr.Markdown("<h1><center>NPC Generator</center></h1>")
  gr.Markdown(
        "based on <a href=https://huggingface.co/spaces/Gradio-Blocks/GPTJ6B_Poetry_LatentDiff_Illustration> Gradio poetry generator</a>."
        "<div>first input name, race and class (or generate them randomly)</div>"
        "<div>Next, use GPT-J to generate a short description</div>"
        "<div>Finally, Generate an illustration 🎨 provided by <a href=https://huggingface.co/spaces/multimodalart/latentdiffusion>Latent Diffusion model</a>.</div>"
        #"<div>Or using <a href=https://huggingface.co/spaces/multimodalart/rudalle> Rudalle model</a>.</div>"
        "<div>Or using <a href=https://huggingface.co/spaces/Gradio-Blocks/clip-guided-faces> clip-guides faces</a>.</div>"
    )
  
  with gr.Row():
    b0 = gr.Button("Randomize name,race and class")
    b1 = gr.Button("Generate NPC Description")
    b2 = gr.Button("Generate Portrait (latent diffusion)")
    b3 = gr.Button("Generate Portrait (clip-faces)")
  
  with gr.Row():  
    input_name = gr.Textbox(label="name",placeholder="Drizzt")
    input_race = gr.Textbox(label="race",placeholder="dark elf")
    input_class = gr.Textbox(label="class",placeholder="ranger")
    input_pronoun = gr.Textbox(label="pronoun",placeholder="he")

  with gr.Row():
    desc_txt = gr.Textbox(label="description",lines=7)
    output_image = gr.Image(label="portrait",type="filepath", shape=(256,256))
  
  b0.click(npc_randomize,inputs=[],outputs=[input_name,input_race,input_class,input_pronoun])
  b1.click(npc_generate, inputs=[ input_name,input_race,input_class,input_pronoun], outputs=desc_txt)
  b2.click(desc_to_image, desc_txt, output_image)
  b3.click(desc_to_image_cf, desc_txt, output_image)
  #examples=examples

demo.launch(enable_queue=True, debug=True)