Add disco diffusion ernievil vitb16

[JunnYu]

PR types

New features

PR changes

Models

Description

• 添加ErnieViL模型及使用该模型+disco diffusion生成图像。
• 该PR依赖上一下CLIP的PR。

import os
os.environ["CUDA_VISIBLE_DEVICES"] = "5"

from paddlenlp.transformers import ErnieViLForImageGeneration, ErnieViLTokenizer

# Initialize the model and tokenizer
model_name_or_path = 'disco_diffusion_ernie_vil-2.0-base-zh'
model = ErnieViLForImageGeneration.from_pretrained(model_name_or_path)
tokenizer = ErnieViLTokenizer.from_pretrained(model_name_or_path)
model.eval()

# Prepare the model inputs.
text_prompt = "孤舟蓑笠翁，独钓寒江雪，油画"
style = None
artist = None
text_prompt = model.preprocess_text_prompt(text_prompt)
tokenized_inputs = tokenizer(text_prompt, return_tensors="pd", padding="max_length", max_length=64)

images = model.generate(**tokenized_inputs)
images[0].save("figure.png")

孤舟蓑笠翁，独钓寒江雪，油画

在宁静的风景中画了一幅美丽的建筑画，由齐白石所作

两个黄鹂鸣翠柳，一行白鹭上青天

该文件里面包含了中英文的dd内容。
test_ernie_vil_clip_dd.tar.gz

[rainyfly]

LGTM for diffusion_generate process.

[guoshengCS]

LGTM