Promptist: 强化学习用于自动提示优化

新闻

• [演示发布] 2022年12月： Demo at HuggingFace Space
• [模型发布] 2022年12月：link
• [论文发布] 2022年12月： Optimizing Prompts for Text-to-Image Generation

• 语言模型作为提示界面，将用户输入优化为模型推荐的提示。

• 通过强化学习学习语言模型以实现自动提示优化。

为 Stable Diffusion v1.4 加载预训练模型

您可以在 https://huggingface.co/spaces/microsoft/Promptist 上尝试在线演示。

[注意] HuggingFace Space上的在线演示使用CPU，因此生成速度较慢。请在本地加载带有GPU的模型以加快生成速度。

import gradio as grad
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

def load_prompter():
  prompter_model = AutoModelForCausalLM.from_pretrained("microsoft/Promptist")
  tokenizer = AutoTokenizer.from_pretrained("gpt2")
  tokenizer.pad_token = tokenizer.eos_token
  tokenizer.padding_side = "left"
  return prompter_model, tokenizer

prompter_model, prompter_tokenizer = load_prompter()

def generate(plain_text):
    input_ids = prompter_tokenizer(plain_text.strip()+" Rephrase:", return_tensors="pt").input_ids
    eos_id = prompter_tokenizer.eos_token_id
    outputs = prompter_model.generate(input_ids, do_sample=False, max_new_tokens=75, num_beams=8, num_return_sequences=8, eos_token_id=eos_id, pad_token_id=eos_id, length_penalty=-1.0)
    output_texts = prompter_tokenizer.batch_decode(outputs, skip_special_tokens=True)
    res = output_texts[0].replace(plain_text+" Rephrase:", "").strip()
    return res

txt = grad.Textbox(lines=1, label="Initial Text", placeholder="Input Prompt")
out = grad.Textbox(lines=1, label="Optimized Prompt")
examples = ["A rabbit is wearing a space suit", "Several railroad tracks with one train passing by", "The roof is wet from the rain", "Cats dancing in a space club"]

grad.Interface(fn=generate,
               inputs=txt,
               outputs=out,
               title="Promptist Demo",
               description="Promptist is a prompt interface for Stable Diffusion v1-4 (https://huggingface.co/CompVis/stable-diffusion-v1-4) that optimizes user input into model-preferred prompts.",
               examples=examples,
               allow_flagging='never',
               cache_examples=False,
               theme="default").launch(enable_queue=True, debug=True)