Найден корректный robots.txt
Файл управления сканированием сайта роботами
?
Robots.txt указывает поисковым роботам, какие страницы сканировать, а какие — нет. Ошибки в файле могут случайно закрыть важные разделы от индексации.
Robots.txt настроен корректно. Размер файла: 28873 байт. Загружен за: 0сек.
Проверяемая страница не запрещена в robots.txt.
Цепочка редиректов для файла robots.txt:
http://parti.research.google/robots.txt
302 Found
https://sites.research.google/parti/
200 OK
Показать содержимое robots.txt
<!DOCTYPE html>
<meta charset="utf-8">
<head>
<link rel="icon" type="image/x-icon" href="assets/favicon.ico">
<title>Parti: Pathways Autoregressive Text-to-Image Model</title>
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.6.0/jquery.min.js"></script>
<link rel="stylesheet" href="style.css">
<script src="registry.js"></script>
<script src="main.js"></script>
</head>
<body>
<!-- Banner -->
<div id="banner">
<div id="banner-text" class="white-on-black">
<h1 style="font-size: 20vw; line-height: 1.0;">
Parti
</h1>
<h1 style="font-size: 2.5vw; line-height: 1.0;">
Pathways Autoregressive Text-to-Image Model
</h1>
</div>
<div id="banner-images"></div>
</div>
<!-- Introduction -->
<div class="table">
<div class="row">
<div class="col-12">
<p>
<a href="https://arxiv.org/abs/2206.10789" class="button black-on-white">Research paper</a>
<a href="https://github.com/google-research/parti" class="button black-on-white">GitHub repository</a>
<br>
<br>
<br>
</p>
<h1>Introduction</h1>
<p>
We introduce the <i>Pathways Autoregressive Text-to-Image</i> model (Parti), an autoregressive text-to-image generation model that achieves high-fidelity photorealistic image generation and supports content-rich synthesis involving complex compositions and world knowledge. Recent advances with diffusion models for text-to-image generation, such as Google’s <a href="https://imagen.research.google/">Imagen</a>, have also shown impressive capabilities and state-of-the-art performance on research benchmarks. Parti and Imagen are complementary in exploring two different families of generative models – autoregressive and diffusion, respectively – opening exciting opportunities for combinations of these two powerful models.
<br>
<br>
Parti treats text-to-image generation as a sequence-to-sequence modeling problem, analogous to machine translation – this allows it to benefit from advances in large language models, especially capabilities that are unlocked by scaling data and model sizes. In this case, the target outputs are sequences of image tokens instead of text tokens in another language. Parti uses the powerful image tokenizer, <a href="https://ai.googleblog.com/2022/05/vector-quantized-image-modeling-with.html">ViT-VQGAN</a>, to encode images as sequences of discrete tokens, and takes advantage of its ability to reconstruct such image token sequences as high quality, visually diverse images.
<br>
<br>
We observed the following results:
<ul>
<li>Consistent quality improvements by scaling Parti’s encoder-decoder up to 20 billion parameters.</li>
<li>State-of-the-art zero-shot FID score of 7.23 and finetuned FID score of 3.22 on MS-COCO.</li>
<li>Effectiveness across a wide variety of categories and difficulty aspects in our analysis on Localized Narratives and PartiPrompts, our new holistic benchmark of 1600+ English prompts that we release as part of this work.</li>
</ul>
<br>
We also explore and highlight limitations of our models, giving key example areas of focus for further improvements.
<br>
<br>
<img style="max-width: 1024px; margin: auto; display: block;" src="assets/parti_overview.jpg">
<br>
<br>
</p>
</div>
</div>
</div>
<!-- Scaling and compositionality -->
<div class="table white-on-black">
<div class="row">
<div class="col-12">
<h1>Scaling from 350M to 20B parameters</h1>
<p>
Parti is implemented in <a href="https://github.com/tensorflow/lingvo">Lingvo</a> and scaled with <a href="https://ai.googleblog.com/2021/12/general-and-scalable-parallelization.html">GSPMD</a> on <a href="https://cloud.google.com/blog/topics/tpus/google-showcases-cloud-tpu-v4-pods-for-large-model-training">TPU v4</a> hardware for both training and inference, which allowed us to train a 20B parameter model that achieves record performance on multiple benchmarks.
<br>
<br>
We perform detailed comparisons of four scales of Parti models – 350M, 750M, 3B and 20B – and observe:
<ul>
<li>Consistent and substantial improvements in model capabilities and output image quality.</li>
<li>When comparing the 3B and 20B models, human evaluators preferred the latter most of the time, specifically:</li>
<ul>
<li>63.2% for image realism/quality</li>
<li>75.9% for image-text match</li>
</ul>
<li>The 20B model especially excels at prompts that are abstract, require world knowledge, specific perspectives, or writing and symbol rendering.</li>
</ul>
<br>
Click on one of the following prompts to compare Parti models across scales:
<br>
<br>
</p>
<p id="scaling-prompt" class="selection">
<span class="options">
<span value="kangaroo" class="option selected">
A portrait photo of a kangaroo wearing an orange hoodie and blue sunglasses standing on the grass in front of the Sydney Opera House holding a sign on the chest that says Welcome Friends!
</span>
<br>
<span value="dl" class="option">
A green sign that says "Very Deep Learning" and is at the edge of the Grand Canyon. Puffy white clouds are in the sky.
</span>
<br>
<span value="astronaut" class="option">
A photo of an astronaut riding a horse in the forest. There is a river in front of them with water lilies.
</span>
<br>
<span value="map" class="option">
A map of the United States made out of sushi. It is on a table next to a glass of red wine.
</span>
<br>
<span value="apple" class="option">
A squirrel gives an apple to a bird
</span>
<br>
<span value="violin" class="option">
The back of a violin
</span>
<br>
<span value="infinity" class="option">
Infinity
</span>
<br>
</span>
</p>
</div>
</div>
<div id="scaling-images" class="row">
<div class="col-3">
<h1 style="text-align: center; font-size: xxx-large;">350M</h1>
<img id="scaling-image-0" src="paper_images_green_watermark_outputs/figures/scaling_comparison/kangaroo_0.jpg">
<br>
<br>
</div>
<div class="col-3">
<h1 style="text-align: center; font-size: xxx-large;">750M</h1>
<img id="scaling-image-1" src="paper_images_green_watermark_outputs/figures/scaling_comparison/kangaroo_1.jpg">
<br>
<br>
</div>
<div class="col-3">
<h1 style="text-align: center; font-size: xxx-large;">3B</h1>
<img id="scaling-image-2" src="paper_images_green_watermark_outputs/figures/scaling_comparison/kangaroo_2.jpg">
<br>
<br>
</div>
<div class="col-3">
<h1 style="text-align: center; font-size: xxx-large;">20B</h1>
<img id="scaling-image-3" src="paper_images_green_watermark_outputs/figures/scaling_comparison/kangaroo_3.jpg">
<br>
<br>
</div>
</div>
<div class="row">
<div class="col-12">
<h1>Composing real-world knowledge</h1>
<p>
Text-to-image generation is most interesting when it allows us to create scenes that have never been seen.
<br>
<br>
We find that Parti can manage long, complex prompts that require it to:
<ul>
<li>Accurately reflect world knowledge</li>
<li>Compose many participants and objects, with fine-grained details and interactions</li>
<li>Adhere to a specific image format and style</li>
</ul>
<br>
In the following examples of prompts and output images, we show how Parti responds to changes in participants, activities, descriptions, locations, and format.
<br>
<br>
</p>
</div>
</div>
<div class="row">
<div class="col-6">
<img id="raccoon-image" src="paper_images_green_watermark_outputs/figures/cherries/raccoons/hokusai.jpg">
<p id="raccoon-prompt" class="selection">
A raccoon wearing formal clothes, wearing a tophap and holding a cane.
The raccoon is holding a garbage bag. Oil painting in the style of
<span class="options">
<span value="rembrandt" class="option">Rembrandt</span>
<span value="vangogh" class="option">Vincent Van Gogh</span>
<span value="hokusai" class="option selected">Hokusai</span>
<span value="pixelart" class="option">pixel art</span>
<span value="cubism" class="option">abstract cubism</span>
<span value="egyptian" class="option">Egyptian tomb heiroglyphics</span>
</span>
</p>
<br>
<br>
</div>
<div class="col-6">
<img id="tiger-image" src="paper_images_green_watermark_outputs/figures/cherries/tigers/tiger3.jpg">
<p id="tiger-prompt" class="selection">
Portrait of a tiger wearing a train conductor’s hat and holding a skateboard that has a yin-yang symbol on it.
<span class="options">
<span value="tiger1" class="option">photograph</span>
<span value="tiger2" class="option">comic book illustration</span>
<span value="tiger3" class="option selected">oil painting</span>
<span value="tiger4" class="option">marble statue</span>
<span value="tiger5" class="option">charcoal sketch</span>
<span value="tiger6" class="option">woodcut</span>
<span value="tiger7" class="option">child’s crayon drawing</span>
<span value="tiger8" class="option">color ink-and-wash drawing</span>
<span value="tiger9" class="option">Chinese ink and wash painting</span>
</span>
<br>
<br>
</p>
</div>
</div>
<div class="row">
<div class="col-6">
<img id="teddy-image" src="paper_images_green_watermark_outputs/figures/cherries/teddy_bears/teddy_1.jpg">
<p id="teddy-prompt" class="selection">
A teddy bear wearing a motorcycle helmet and cape is
<span class="options">
<span value="teddy_0" class="option">standing in front of Loch Awe with Kilchurn Castle behind him</span>
<span value="teddy_1" class="option selected">driving a speed boat near the Golden Gate Bridge</span>
<span value="teddy_2" class="option">car surfing on a taxi cab in New York City</span>
<span value="teddy_3" class="option">riding a motorcycle in Rio de Janeiro with Dois Irmãos in the background</span></span>. dslr photo.
</p>
<br>
<br>
</div>
<div class="col-6">
<img id="water-image" src="paper_images_green_watermark_outputs/figures/cherries/water/water_crocodile.jpg">
<p id="water-prompt" class="selection">
A photo of a
<span class="options">
<span value="water_maple" class="option">maple leaf</span>
<span value="water_palm" class="option">palm tree</span>
<span value="water_clover" class="option">four-leaf clover</span>
<span value="water_lotus" class="option">lotus flower</span>
<span value="water_panda" class="option">panda</span>
<span value="water_teddy_bear" class="option">teddy bear</span>
<span value="water_crocodile" class="option selected">crocodile</span>
<span value="water_dragonfly" class="option">dragonfly</span>
</span>
made of water.
</p>
<br>
<br>
</div>
</div>
<div class="row">
<div class="col-6">
<img id="vase-image" src="paper_images_green_watermark_outputs/figures/cherries/cherry2/pangolin_basketball.jpg">
<p id="vase-prompt" class="selection">
A photo of an Athenian vase with a painting of
<span class="options">
<span value="panda" class="option">pandas</span>
<span value="toucan" class="option">toucans</span>
<span value="pangolin" class="option selected">pangolins</span>
</span>
playing
<span class="options">
<span value="tennis" class="option">tennis</span>
<span value="soccer" class="option">soccer</span>
<span value="basketball" class="option selected">basketball</span>
</span>
in the style of Egyptian hieroglyphics.
</p>
<br>
<br>
</div>
<div class="col-6">
<img id="tornado-image" src="paper_images_green_watermark_outputs/figures/cherries/tornado/tiger_cubism.jpg">
<p id="tornado-prompt" class="selection">
A tornado made of
<span class="options">
<span value="sharks" class="option">sharks</span>
<span value="tiger" class="option selected">tigers</span>
<span value="bees" class="option">bees</span>
</span>
crashing into a skyscraper. Painting in thestyle of
<span class="options">
<span value="hokusai" class="option">Hokusai</span>
<span value="cubism" class="option selected">abstract cubism</span>
<span value="watercolor" class="option">watercolor</span>
</span>
</p>
<br>
<br>
</div>
</div>
</div>
<!-- Limitations, responsibility, and acknowledgments -->
<div class="table">
<div class="row">
<div class="col-12">
<h1>PartiPrompts benchmark</h1>
<p>
PartiPrompts (P2) is a rich set of over 1600 prompts in English that <a href="https://github.com/google-research/parti">we release as part of this work</a>. P2 can be used to measure model capabilities across various categories and challenge aspects.
<br>
<br>
<img style="max-width: 1024px; margin: auto; display: block;" src="assets/bcp.png">
<br>
<br>
P2 prompts can be simple, allowing us to gauge the progress from scaling.
They can also be complex, such as the following 67-word description we created for Vincent van Gogh’s <i>The Starry Night</i> (1889):
<br>
<br>
<p style="font-style: italic; max-width: 1024px;">
Oil-on-canvas painting of a blue night sky with roiling energy. A fuzzy and bright yellow crescent moon shining at the top. Below the exploding yellow stars and radiating swirls of blue, a distant village sits quietly on the right. Connecting earth and sky is a flame-like cypress tree with curling and swaying branches on the left. A church spire rises as a beacon over rolling blue hills.
</p>
</p>
</div>
</div>
<div class="row">
<div class="col-3">
<img src="paper_images_green_watermark_outputs/teaser_images/starry_night/11.jpg">
</div>
<div class="col-3">
<img src="paper_images_green_watermark_outputs/teaser_images/starry_night/20.jpg">
</div>
<div class="col-3">
<img src="paper_images_green_watermark_outputs/teaser_images/starry_night/28.jpg">
</div>
<div class="col-3">
<img src="paper_images_green_watermark_outputs/teaser_images/starry_night/34.jpg">
</div>
</div>
</div>
<div class="table white-on-black">
<div class="row">
<div class="col-12">
<h1>Discussion and limitations</h1>
<p>
Many of the images shown here have been selected, or cherry-picked, from a large set of examples generated during prompt exploration and modification interactions. We call this process “Growing The Cherry Tree'' and provide a detailed example of it in <a href="https://arxiv.org/abs/2206.10789">the paper</a>, where we build a very complex prompt and strategies to produce an image that fully reflects the description.
<br>
<br>
While Parti produces high quality outputs for a broad range of prompts, the model nevertheless has many limitations. In the paper, we discuss these challenges with examples, current failure modes, and opportunities for future work. We provide a sample of some of these failure cases in the interactive visualization below.
<br>
<br>
</p>
</div>
</div>
<div class="row">
<div id="failure-column" class="col-4">
<img id="failure-image" src="paper_images_green_watermark_outputs/figures/limitations/banana_juice/banana_juice_1.jpg">
<p>
<span id="failure-reason">
Failure: improper handling of negation or indication of absence.
</span>
<br>
<br>
</p>
</div>
<div class="col-8">
<p id="failure-prompt" class="selection">
<span class="options">
<span value="basetennisballs/2tennis1mergedball" class="option">
Two baseballs to the left of three tennis balls.
</span>
<br>
<span value="rhinobeetle/rhino2" class="option">
A rhino beetle this size of a tank grapples a real life passenger airplane on the tarmac.
</span>
<br>
<span value="yellow_anubis/yellow_anubis4" class="option">
A portrait of a statue of Anubis with a crown and wearing a yellow t-shirt that has a space shuttle drawn on it. A white brick wall is in the background.
</span>
<br>
<span value="labra_cat/labra_cat4" class="option">
A cream colored labradoodle next to a white cat with black-tipped ears.
</span>
<br>
<span value="banana_juice/banana_juice_1" class="option selected">
A plate that has no bananas on it. there is a glass without orange juice next to it.
</span>
<br>
<span value="robot_airplane/robot_airplane1" class="option">
A robot painted as graffiti on a brick wall. The words "Fly an airplane" are written on the wall. A sidewalk is in front of the wall, and grass is growing out of cracks in the concrete.
</span>
<br>
<span value="impossible/robot_impossible" class="option">
A shiny robot wearing a race car suit and black visor stands proudly in front of an F1 race car. The sun is setting on a cityscape in the background. comic book illustration.
</span>
<br>
</span>
</p>
</div>
</div>
</div>
<div class="table">
<div class="row">
<div class="col-12">
<h1>Responsibility and broader impact</h1>
<p>
As we discuss at greater length in the paper, text-to-image models introduce many opportunities and risks, with potential impact on bias and safety, visual communication, disinformation, and creativity and art. Similar to
<a href="https://imagen.research.google/">Imagen</a>,
we recognize there is a risk that Parti may encode harmful stereotypes and representations. Some potential risks relate to the way in which the models are themselves developed, and this is especially true for the training data. Current models like Parti are trained on large, often noisy, image-text datasets that are known to contain biases regarding people of different backgrounds. This leads such models, including Parti, to produce stereotypical representations of, for example, people described as lawyers, flight attendants, homemakers, and so on, and to reflect Western biases for events such as weddings. This presents particular problems for people whose backgrounds and interests are not well represented in the data and the model, especially if such models are applied to uses such as visual communication, e.g. to help low-literacy social groups. Models which produce photorealistic outputs, especially of people, pose additional risks and concerns around the creation of deepfakes. This creates risks with respect to the possible propagation of visually-oriented misinformation, and for individuals and entities whose likenesses are included or referenced.
<br>
<br>
Text-to-image models open up many new possibilities for people to create unique and aesthetically pleasing images – essentially, acting as a paint brush to enhance human creativity and productivity. However, in assessing design or artistic merit, it is important to have
<a href="https://direct.mit.edu/leon/article-abstract/55/2/130/102695/Who-or-What-Is-an-AI-Artist?redirectedFrom=fulltext">a nuanced understanding of algorithmically based art</a>
over the years, the model itself, the people involved and the broader artistic milieu. Bias also matters here, as the range of outputs from a model is dependent on the training data, and this may have biases toward Western imagery and further prevent models from exhibiting radically new artistic styles – the way human artists can.
<br>
<br>
For these reasons, we have decided not to release our Parti models, code, or data for public use without further safeguards in place. In the meantime, we provide a Parti watermark on all images that we release. We will focus on following this work with further careful model bias measurement and mitigation strategies, such as prompt filtering, output filtering, and model recalibration. We believe it may be possible to use text-to-image generation models to understand biases in large image-text datasets at scale, by explicitly probing them for a suite of known bias types, and potentially uncovering other forms of hidden bias. We also plan to coordinate with artists to adapt high-performing text-to-image generation models’ capabilities to their work. This is especially important given the intense interest among many research groups, and the rapid development of models and data to train them. Ideally, we hope these models will augment human creativity and productivity, not replace it, so that we can all enjoy a world filled with new, varied, and responsible aesthetic visual experiences.
<br>
<!-- Many broader issues should be considered with large-scale models for text-to-image generation, such as:
<ul>
<li>Development process issues like the use of large, mostly uncurated training datasets of images obtained from the web with little oversight.</li>
<li>Conceptual vagueness of constructs in task formulation.</li>
<li>The role of large text-to-image models as “foundation” models that act as a form of infrastructure for a range of system applications and fine-tuning for image generation tasks, shaping our conceptions of what is both possible and desirable.</li>
</ul>
<br>
It is difficult (if not impossible) to predict all possible uses and consequences of infrastructure, so following responsible AI practices that emphasize transparently documenting and sharing information about datasets and models is essential.
Text-to-image generation models could potentially have an impact on creativity and art, visual communication, disinformation (including deepfakes), and bias and safety. Considerations for this potential impact contribute to our decision not to release our models, code, or data at this time.
<br>
<br>
Instead, we will focus on following this work with further careful model bias measurement and mitigation strategies, such as prompt filtering, output filtering, and model recalibration. We believe it may be possible to use text-to-image generation models to understand biases in large image-text datasets at scale, by explicitly probing them for a suite of known bias types, and potentially uncovering other forms of hidden bias.
<br>
<br>
We also plan to coordinate with artists to adapt high-performing text-to-image generation models’ capabilities to their work, for purely creative ends or art-for-hire. This is especially important given the intense interest among many research groups, and the rapid development of models and data to train them.
Ideally, we hope these models will augment human creativity and productivity, not replacing it, so that we can all enjoy a world filled with new, varied, and responsible aesthetic visual experiences.
<br> -->
<br>
<br>
<a href="https://github.com/google-research/parti/tree/main/data_cards" class="button black-on-white">Data card</a>
<!-- <a href="#" class="button black-on-white">Model card</a> -->
<br>
<br>
<br>
</p>
<h1>Acknowledgements</h1>
<p>
Parti is a collaboration that spans authors across multiple <a href="https://research.google/">Google Research</a> teams:
<br>
<br>
Jiahui Yu<sup>*</sup>,
Yuanzhong Xu<sup>†</sup>,
Jing Yu Koh<sup>†</sup>,
Thang Luong<sup>†</sup>,
Gunjan Baid<sup>†</sup>,
Zirui Wang<sup>†</sup>,
Vijay Vasudevan<sup>†</sup>,
Alexander Ku<sup>†</sup>
<br>
Yinfei Yang,
Burcu Karagol Ayan,
Ben Hutchinson,
Wei Han,
Zarana Parekh,
Xin Li,
Han Zhang
<br>
Jason Baldridge<sup>†</sup>,
Yonghui Wu<sup>*</sup>
<br>
<br>
<sup>*</sup>Equal contribution
<sup>†</sup>Core contribution
<br>
<br>
We would like to thank Elizabeth Adkison, Fred Alcober, Tania Bedrax-Weiss, Krishna Bharat, Nicole Brichtova, Yuan Cao, William Chan, Zhifeng Chen, Eli Collins, Claire Cui, Andrew Dai, Jeff Dean, Emily Denton, Toju Duke, Dumitru Erhan, Brian Gabriel, Zoubin Ghahramani, Jonathan Ho, Michael Jones, Sarah Laszlo, Quoc Le, Lala Li, Zhen Li, Sara Mahdavi, Kathy Meier-Hellstern, Kevin Murphy, Paul Natsev, Paul Nicholas, Mohammad Norouzi, Niki Parmar, Ruoming Pang, Fernando Pereira, Slav Petrov, Vinodkumar Prabhakaran, Utsav Prabhu, Evan Rapoport, Keran Rong, Negar Rostamzadeh, Chitwan Saharia, Gia Soles, Austin Tarango, Ashish Vaswani, Tao Wang, Tris Warkentin, Austin Waters, Ben Zevenbergen for helpful discussions and guidance, Peter Anderson, Corinna Cortes, Tom Duerig, Douglas Eck, David Ha, Radu Soricut and Rahul Sukthankar for paper review and feedback, Erica Moreira and Victor Gomes for help with resource coordination, Tom Small for designing the Parti watermark, Google ML Data Operations team for collecting human evaluations on our generated images and others in the Google Brain team and Google Research team for support throughout this project.
<br>
<br>
We would also like to give particular acknowledgments to the Imagen team, especially Mohammad Norouzi, Chitwan Saharia, Jonathan Ho and William Chan, for sharing their near complete results prior to releasing Imagen; their findings on the importance of CF guidance were particularly helpful for the final Parti model. We also thank the Make-a-Scene team, especially Oran Gafni, for helpful discussion on CF-guidance implementation in autoregressive models. We thank the DALL-E 2 authors, especially Aditya Ramesh, for helpful discussion on MS-COCO evaluation. We also thank the DALL-Eval authors, especially Jaemin Cho, for help with reproducing their numbers.
</p>
<!-- <br>
<br>
<pre style="font-family: Courier; line-height: 1.0;">
@article{key ,
author = {Arthur B Cummings and David Eftekhary and Frank G House},
title = {The accurate determination of college students’
coefficients of friction},
journal = {Journal of Sketchy Physics},
volume = {13},
year = {2003},
number = {2},
pages = {46--129}
}
</pre> -->
</div>
</div>
</div>
<!-- Gallery -->
<div id="gallery" class="table white-on-black">
<div class="row">
<div class="col-12">
<h1 style="font-size: 150px; text-align: center;">Gallery</h1>
</div>
</div>
</div>
</body>
Google Frontend