talhaz / simulacra-aesthetic-captions Goto Github PK

Simulacra Aesthetic Captions is a dataset of over 238000 synthetic images generated with AI models such as CompVis latent GLIDE and Stable Diffusion from over forty thousand user submitted prompts. The images are rated on their aesthetic value from 1 to 10 by users to create caption, image, and rating triplets. In addition to this each user agreed to release all of their work with the bot: prompts, outputs, ratings, completely public domain under the CC0 1.0 Universal Public Domain Dedication. The result is a high quality royalty free dataset with over 176000 ratings that can be used for projects such as:

• Filtering Datasets
• Guiding Generative Models
• Training A Prompt Generator
• Extracting vitamin phrases ("trending on artstation", etc)
• Alignment Research

Because earlier versions of SimulacraBot used a separate upscaler to make latent GLIDE outputs palatable to users the dataset also includes limited contrastive data. It consists of which images were chosen to be upscaled from each batch with 12,648 samples, or about 2000 batches. The resulting upscaled images are not included in the dataset.

An early version of Simulacra Aesthetic Captions was used to create LAION Aesthetic, a version of LAION 5b filtered using around 4000 ratings.

Special thanks to Victor Weaver (twitter: @_victorweaver) for his early feedback and suggestions on the bot, as well as writing the initial public domain prompt and rating dataset that allowed us to make LAION-aesthetic as early as we did.

The index of Simulacra Aesthetic Captions is a sqlite database containing all information and metadata besides the images themselves:

curl -OL https://raw.githubusercontent.com/JD-P/simulacra-aesthetic-captions/main/sac_public_2022_06_29.sqlite

The images in version 1.0 of Simulacra Aesthetic Captions can be downloaded from their bucket with the following cURL command:

curl -OL https://s3.us-west-1.wasabisys.com/simulacrabot/sac.tar

If you use Simulacra Aesthetic Captions in your published work we suggest the following citation:

@techreport { pressmancrowson2022 ,
author = { John David Pressman and Katherine Crowson and Simulacra Captions Contributors } ,
year = 2022 ,
title = { Simulacra Aesthetic Captions } ,
institution = { Stability AI } ,
type = {} ,
number = { Version 1.0 } ,
note = {\ url { https://github.com/JD-P/simulacra-aesthetic-captions }}
}

SimulacraBot has used multiple generation methods over its lifetime, and will continue to use new ones in the future. Each generation method is defined as a model inferenced with a specific set of parameters. If the parameters change the method number should be incremented for generations done with that method. For reference here is every generation method used over SimulacraBot's lifetime, even though only 2 and 3 appear in the dataset as currently released.

The initial design of Simulacra Aesthetic Captions called for collecting all ratings into an anonymous pool. However during data collection it was observed that some users would downrate objectively well drawn pictures because they're "scary", or included public figures they didn't like such as the US president. Some users would also ignore rating instructions to downrate watermarks and other marks that encroach on the canvas of the output. Rather than try to force people to go against their intuition, which seemed somewhere between tyrannical and futile, an aesthetic survey was implemented on signup that records users behavior around these confounding factors in aesthetic judgment. The survey currently consists of 20 questions in the form of 1-10 ratings of fixed hand chosen images that each represent either a conjectured component of human aesthetic judgment or a bias that users of the dataset would like to be able to handle during their analysis. This survey data also enhances the possibility of using Simulacra Aesthetic Captions for general research into human aesthetic preferences by e.g. deriving clusters of human aesthetic judgment. The images and their associated captions are shown below:

A unique combination of ratings on the aesthetic survey also serves as a synthetic user ID through the dataset. This was done to balance privacy against being able to get at the aesthetic preferences of coherent agents versus an averaged out mush. Unfortunately due to a bug in earlier version of SimulacraBot, some aesthetic surveys were corrupted with more than 20 entries. It is not known exactly what caused this or what the ratings are supposed to be, so it may be preferable to exclude these surveys from analysis that relies on the survey to make its conclusions.

Simulacra Aesthetic Captions has many biases in its portrait of human aesthetic preferences. The relevant biases for different use cases are summarized in the tables below.

The level of concern for the bias on the use case has been marked with 0-3 X's. With no X being almost no concern and three X's being something the user will probably make a mistake if they don't think carefully about.

Because Simulacra Aesthetic Captions does not intentionally contain NSFW entries, it is missing a key component of human aesthetic values and experience. Unfortunately the challenges of collecting a dataset like this from users at scale were considered too perilous for a first release. It is likely that a NSFW supplement to Simulacra Aesthetic Captions would be collected as a separate dataset that could be added onto the main body of SAC.

The guidance given to users on what constitutes 'NSFW' was as follows:

What is NSFW?

NSFW basically means:

- Excessive gore/blood/etc
- Licentious nudity, nudity that exists to suggest sex including sexualized poses, etc
- Sex. Anything that involves actual sex or sex acts is way over the line, never deliberately submit prompts that involve this.

To get pedantic about nudity, if it's fantasy art depicting bare nipples but otherwise not licentious it's probably fine. If it depicts a penis or vagina it is probably not fine, but exceptions might be made. Avoid deliberately submitting prompts involving nude figures. As a stock example of nudity which is not NSFW, the Rider-Waite depiction of The Star is nude, but it is not lewd or licentious.

Also note that something doesn't need to be nude to be removed as sexually NSFW. A sufficiently trashy prompt/image combo can and will be removed. "Slut looking at you from across the bar", things like that.

The lack of copyrighted content in Simulacra Aesthetic Captions bars models trained with it from a full understanding of pop culture. Because mass culture in Western countries largely consists of specific copyrighted content, models that rely on SAC as their ground truth for any aspect of their function will have that functionality impaired when handling popular culture.

For compelling reasons hateful content is excluded from Simulacra Aesthetic Captions. However hatred of others and reactions to that hatred are undeniably part of human values. Even if we find it so abhorrent as to place it into some other category, the paradox of tolerance involves hatred and not tolerating the intolerant is a major value of contemporary Western society. Because the dataset lacks hate or reactions to hate users who would like to use SAC as part of their moderation suite will probably find it lackluster. While a "Hateful Aesthetic Captions" might be a useful supplement for moderation suites and AGI alignment, it's not clear where a population of users could be found who would cooperate with the intent of such a dataset while producing authentic samples. Furthermore the benefits of releasing such a dataset would have to be weighed against the impact to other people of making large amounts of such content high quality, curated, and freely available even for legitimate academic purposes.

Participants in Simulacra Aesthetic Captions are largely [WEIRD]https://www2.psych.ubc.ca/~henrich/pdfs/WeirdPeople.pdf), or Western, Educated, Industrialized, Rich, and Democratic. This means that the aesthetic preferences recorded are not universal among humanity. While we didn't take a survey of the demographic makeup of SimulacraBot users, it should be assumed they largely reside in the United States and Europe. We welcome replications of Simulacra Aesthetic Captions which sample their results from other locales and demographics.

Further narrowing the scope of aesthetic feedback in Simulacra Aesthetic Captions is the userbase consisting largely of people who are power users and developers of AI art. This means that their aesthetic feedback is going to lean STEM, fantasy, nerdy, esoteric, etc.

Because users write their prompts in the presence of other people, they are going to avoid "guilty pleasures" and other topics on which they might be judged. This means that anything which might only be expressed anonymously or in private is less likely to appear in Simulacra Aesthetic Captions.

Users are naturally going to gravitate towards things the model is good at drawing. Each generative model has a set of biases about what it is capable of drawing well, users pick on these and tailor their prompt style to accomodate it. This means that Simulacra Aesthetic Captions only represents human preferences conditional on the models biases.

Some users of Simulacra Aesthetic Captions are only interested in having the model draw a couple of things. For example one user really likes using the model to draw a 'world tree', or the earth represented as a giant tree. These fixations and narrow interests can become large parts of the dataset relative to their general importance. Before using Simulacra Aesthetic Captions to filter a dataset, it may pay to look closely at repeated prompts and themes to cull their numbers down to sane levels.

During discussions about rating, users repeatedly express a preference to grade models on a curve. They give charity points to models where they know they're lacking so they don't rate everything low. This means that different methods probably have different rating distributions relative to the 'ground truth' aesthetic preference of the image without the context of the model that generated it.

During the period that SimulacraBot offered upscaling, it had a bug when multiple users used the bot that meant the author of a prompt had to 'race' to open their upscale menu first if they wanted to upscale their outputs. This race condition probably had subtle effects on the distribution of images found in the upscaled data.

Like most services, SimulacraBot usage is on a power law where a small minority of the nearly 400 users contribute the vast majority of the content. This means that while in theory the ratings and prompts are contributed by many people, in practice a handful of users aesthetic preferences dominate the dataset.