add disco diffusion stable diffusion into taskflow

[JunnYu]

PR types

New features

PR changes

Taskflow

Description

• 给taskflow接入disco diffusion 和 stable diffusion

>>> from paddlenlp import Taskflow
# 默认模型为 pai-painter-painting-base-zh，
>>> text_to_image = Taskflow("text_to_image", model="pai-painter-painting-base-zh", num_return_images=2)
# 单条输入， 默认返回2张图片。
>>> images = text_to_image("风阁水帘今在眼，且来先看早梅红")
# [[<PIL.Image.Image image mode=RGB size=256x256>], [<PIL.Image.Image image mode=RGB size=256x256>]]
>>> images[0][0].save("painting-figure-1.png")
>>> images[0][1].save("painting-figure-2.png")
>>> images[0][0].argument
# argument表示生成该图片所使用的参数
# {'input': '风阁水帘今在眼，且来先看早梅红',
#  'batch_size': 1,
#  'seed': 2414128200,
#  'temperature': 1.0,
#  'top_k': 32,
#  'top_p': 1.0,
#  'condition_scale': 10.0,
#  'num_return_images': 2,
#  'use_faster': False,
#  'use_fp16_decoding': False,
#  'image_index_in_returned_images': 0}
#
# 多条输入， 返回值解释：[[第一个文本返回的第一张图片, 第一个文本返回的第二张图片], [第二个文本返回的第一张图片, 第二个文本返回的第二张图片]]
>>> image_list = text_to_image(["风阁水帘今在眼，且来先看早梅红", "见说春风偏有贺，露花千朵照庭闹"])
# [[<PIL.Image.Image image mode=RGB size=256x256>, <PIL.Image.Image image mode=RGB size=256x256>],
#  [<PIL.Image.Image image mode=RGB size=256x256>, <PIL.Image.Image image mode=RGB size=256x256>]]
>>> for batch_index, batch_image in enumerate(image_list):
# len(batch_image) == 2 (num_return_images)
>>>     for return_image_index, each_image in enumerate(batch_image):
>>>         each_image.save(f"painting-figure_{batch_index}_{return_image_index}.png")

支持多种模型

EasyNLP仓库中的pai-painter模型

>>> text_to_image = Taskflow("text_to_image", model="pai-painter-commercial-base-zh", num_return_images=2)
>>> image_list = text_to_image(["女童套头毛衣打底衫秋冬针织衫童装儿童内搭上衣", "春夏真皮工作鞋女深色软皮久站舒适上班面试职业皮鞋"])
>>> for batch_index, batch_image in enumerate(image_list):
>>>     # len(batch_image) == 2 (num_return_images)
>>>     for return_image_index, each_image in enumerate(batch_image):
>>>         each_image.save(f"commercial-figure_{batch_index}_{return_image_index}.png")

DALLE-mini模型

>>> text_to_image = Taskflow("text_to_image", model="dalle-mini", num_return_images=2)
>>> image_list = text_to_image(["New York Skyline with 'Google Research Pizza Cafe' written with fireworks on the sky.", "Dali painting of WALL·E"])
>>> for batch_index, batch_image in enumerate(image_list):
>>>     # len(batch_image) == 2 (num_return_images)
>>>     for return_image_index, each_image in enumerate(batch_image):
>>>         each_image.save(f"dalle-mini-figure_{batch_index}_{return_image_index}.png")

Disco Diffusion模型

# 注意，该模型生成速度较慢，最好返回1张图片。
>>> text_to_image = Taskflow("text_to_image", model="disco_diffusion_ernie_vil-2.0-base-zh", num_return_images=1)
>>> image_list = text_to_image("一幅美丽的睡莲池塘的画，由Adam Paquette在artstation上所做。")
>>> for batch_index, batch_image in enumerate(image_list):
>>>     for return_image_index, each_image in enumerate(batch_image):
>>>         each_image.save(f"disco_diffusion_ernie_vil-2.0-base-zh-figure_{batch_index}_{return_image_index}.png")

Stable Diffusion模型

>>> text_to_image = Taskflow("text_to_image", model="CompVis/stable-diffusion-v1-4", mode="text2image", num_return_images=2)
>>> prompt = [
    "In the morning light,Chinese ancient buildings in the mountains,Magnificent and fantastic John Howe landscape,lake,clouds,farm,Fairy tale,light effect,Dream,Greg Rutkowski,James Gurney,artstation",
    "clouds surround the mountains and Chinese palaces,sunshine,lake,overlook,overlook,unreal engine,light effect,Dream，Greg Rutkowski,James Gurney,artstation"
    ]
>>> image_list = text_to_image(prompt)
>>> for batch_index, batch_image in enumerate(image_list):
>>>     # len(batch_image) == 2 (num_return_images)
>>>     for return_image_index, each_image in enumerate(batch_image):
>>>         each_image.save(f"stable-diffusion-figure_{batch_index}_{return_image_index}.png")

支持复现生成结果 (以Stable Diffusion模型为例)

>>> from paddlenlp import Taskflow
>>> text_to_image = Taskflow("text_to_image", model="CompVis/stable-diffusion-v1-4", mode="text2image", num_return_images=2)
>>> prompt = [
    "In the morning light,Chinese ancient buildings in the mountains,Magnificent and fantastic John Howe landscape,lake,clouds,farm,Fairy tale,light effect,Dream,Greg Rutkowski,James Gurney,artstation",
    ]
>>> image_list = text_to_image(prompt)
>>> for batch_index, batch_image in enumerate(image_list):
>>>     # len(batch_image) == 2 (num_return_images)
>>>     for return_image_index, each_image in enumerate(batch_image):
>>>         each_image.save(f"stable-diffusion-figure_{batch_index}_{return_image_index}.png")
# 如果我们想复现promt[0]文本的第二张返回的结果，我们可以首先查看生成该图像所使用的参数信息。
>>> each_image.argument
# {'mode': 'text2image',
#  'seed': 2389376819,
#  'height': 512,
#  'width': 512,
#  'num_inference_steps': 50,
#  'guidance_scale': 7.5,
#  'latents': None,
#  'num_return_images': 1,
#  'input': 'In the morning light,Chinese ancient buildings in the mountains,Magnificent and fantastic John Howe landscape,lake,clouds,farm,Fairy tale,light effect,Dream,Greg Rutkowski,James Gurney,artstation'}
# 通过set_argument设置该参数。
>>> text_to_image.set_argument(each_image.argument)
>>> new_image = text_to_image(each_image.argument["input"])
# 查看生成图片的结果，可以发现最终结果与之前的图片相一致。
>>> new_image[0][0]

[wawltor]

LGTM