HooYa! is a Web3 community for commissioning and supporting manga & dōjin artists; it offers leasable storage space for illustrations submitted by artists and network users and provides a folksonomy of descriptive, searchable tags and metadata for each illustration it hosts. HooYa! maintains also a folksonomy for images that are not stored on-chain but that may be accessible through other services such as Pixiv and Danbooru.

It aspires to be rapidly become completely autonomous and self-governing by issuing its own governance tokens for use in image proposal / acceptance and metadata tagging. It also seeks to incentivize the on-chain commission + minting of new artworks by manga & dōjin artists (hereafter "illustrators") via these native utility tokens.

Background

The service HooYa! provides is not novel; it finds inspiration from a handful of Web2 companies (mainly Japanese ones) and from the communities that form around these companies and the illustrators they support. I see HooYa! as a combination of these two communities in particular:

• Pixiv, a Japanese-language community of illustrators who connect with fans by posting original and derivative artworks
• Danbooru, a community dedicated to the archive of illustrations like those posted on pixiv, known mostly for its extensive vocabulary of tags used for organizing and labelling these images

The wealth of information provided by Danbooru has helped many unique projects to create machine-learning models that can both (a) predict appropriate tags for a given illustration (example, src) and (b) generate new illustrations that resemble the Japanese style (example, background).

HooYa! will leverage these pioneering predictive models to suggest tags on images submitted for inclusion on the network.

What is Pixiv?

Pixiv is a community of primarily Japanese illustrators who upload their illustrations, comics and short stories and share them with other pixiv members. This is one of the primary ways that illustrators can connect with their communities and make their artwork accessible to their fans. It is the 10th most visited site in Japan and is most popular among young Japanese illustrators and their fans.

Pixiv offers a Patreon-like subscription program called pixivFANBOX which connect fans with illustrators publishing derivative and original illustrations on a regular basis by providing fans in each subscriber tier with special goods, articles and hi-res renders of illustrations released by the artist.

Because drawing is a popular hobby in Japan (and one which is gaining popularity abroad) the idea of earning a living as an illustrator has attracted many creators, both Japanese and internationals, to the Pixiv platform.

In addition to subscription services pixiv also provides a space for illustrators to sell their illustrations, manga (both short- and long-form) and artbooks on its BOOTH site. These can be digital-only releases but there are also physicals for sale which are produced through its pixivFACTORY service.

Fans can commission personal works from illustrators by a feature on pixiv known as Requests. Requests can be funded entirely by an individual or can be crowd-funded by many fans pooling cash together to commission the piece.

Finally, Pixiv lets fans and illustrators to add a limited set of tags to an image (up to 10) to an image. These Pixiv tags are more akin to hashtags on Twitter than semantic metadata as seen on sites like Danbooru.

Danbooru

Danbooru is a community that tags and archives images from Pixiv, Twitter, DeviantArt and many other platforms. Each image submitted for inclusion on Danbooru goes through a 3-day review period after which it is either accepted or denied inclusion into the Danbooru canon. Danbooru maintains a large vocabulary of tags which are key-value pairs used to describe an image. This makes the set of images easily searchable by specifying (eg) the illustrator, character, franchise or any objects in an image.

As shown above Danbooru lets me search things such as: illustrations of the Love Live! franchise which include product placement of some kind, by issuing the query love_live! product_placement. From left to right, top to bottom these are the Subaru WRX, Starbucks, Coca-Cola, Sprite and the Yamaha Vino. Danbooru limits searches to 2 tags unless the user has purchased a Gold account, in which case the limit is raised to 6. Gold accounts cost a one-time purchase of $20 USD.

Danbooru hosts over 4.9 million images tagged with 162 million tags, of 498k unique tags as of the end of 2021. A single copy of this data consumes 4.5 terrabytes of disk space.

As alluded to earlier, this massive data set has allowed researchers to create predictive models leveraging neural networks to label the objects in a brand new illustration, thus reducing the workload on a single user tagging an image. In addition to predictive models, generative adversarial networks (in particular NVIDIA's StyleGAN) have been trained with the Danbooru set to draw brand new illustrations when prompted with a series of tags.

To the left is Gwern's Danbooru2018 portrait model trained to generate portraits in the Danbooru style (credit Gwern for the model and video). Each image in the interpolation is unique in that it is an original of the GAN, and though the model was seeded from the images collected on Danbooru the portraits are original creations.

HooYa! uses the predictive model and the generative model in a few different ways (see below 3: Prediction and 4: Generation).

HooYa! — A Decentralized Platform for Dōjin & Manga Artists

Now with some understanding of the context in which HooYa! is being developed I can explain the mechanics the HooYa! network.

HooYa! at its core can be broken into three components:

1. Consensus (of metadata, tags, governance, and the image canon)
2. Leased ephemeral storage (of commissioned images and those minted on-chain)
3. Prediction (of tags) + generation (of reference images)

The first two of these four components may be familiar to some; the last one is a novel component to compute on a blockchain and is one thing (among many) which I hope will set HooYa! apart from many other chains, both on Cosmos and beyond.

1 — Consensus

HooYa! is a blockchain built on the Cosmos SDK. It uses Tendermint (like nearly every other Cosmos SDK chain) to maintain network connectivity and consensus across a group of validators. Many chains I have seen begin with a set of 100 validator slots and so this is what HooYa! adopts initially.

Each node has a perfect copy of the canon which refers to the set of:

• Images (uniquely identified by a SHA2-256 multihash)
• Ownership of those images
• Tags describing those images
• Metadata describing those images

... that HooYa! indexes. HooYa!'s canon is a subset of network state and therefore any archival node can retrieve a full copy of the canon at any given chain height. On-chain governance, ongoing subscriptions to illustrators, payment settlement for commissioned artwork is a subset of HooYa! network state, and therefore is a part of data stored on the blockchain, but is not what I mean when I reference the canon, which refers to exclusively those four things above.

The canon is contributed by HooYa! network users, not necessarily the original illustrators themselves. HooYa! users may consult the neural networks proposed in (3) during the image tagging process to more accurately and swiftly tag images. Furthermore because canon includes only metadata about an image but not the image itself there is no risk of infringement on any copyrights by storing canon; the image data is stored in ephemeral storage (2) if it has been leased; this data is released once its lease expires.

[... more to come here]

2 — Leased Ephemeral Storage

In addition to running the HooYa! daemon which maintains consensus (1) all validators are expected to run an IPFS daemon. This daemon connects validators to each other in a private cluster separate from the "global" IPFS instance that we often refer to when talking of IPFS. The private IPFS cluster provides ephemeral storage of the images that HooYa! has reached consensus in (1) should be stored on the network a period of time determined by the image's lease terms.

Unlike for metadata stored as a part of consensus (1) no guarantee is made regarding the retrieval of data whose lease has expired. Validators are not required to store copies of all data ever indexed by HooYa!, only data whose lease is current because that space is reclaimed once the lease for that space expires.

To my knowledge no other Cosmos SDK chain (not even Stargaze which arguably should use this) leverages IPFS like this. I believe this is a wonderful decentralized alternative to things like Pinata for IPFS pinning that is especially attractive for NFT projects on Cosmos. Even if other projects do not use HooYa!'s consensus (1) or neural-net prediction + generation (3) they can lease ephemeral storage (2) to store NFT metadata and images.

[... more to come here]

3 — Prediction + Generation

There are a number of pre-trained models which can predict tags for a given illustration. Validators host these neural networks and provide suggestions for each image uploaded to ephemeral storage in (2). The deterministic nature of these networks dictates that each image has exactly one set of network outputs (ie predicted tags); models are therefore versioned and distributed as part of network upgrades, ensuring each validator has the latest neural network, and those that do not provide bitwise identifical network outputs will see their stake slashed after an amount of time. This behavior is similar to the slashing on Gravity Brdige ( see GRAVSLASH-02 )

For providing one-time GPU compute power for tag prediction validators are compensated with a portion of the data leasing fee.

[... more to come here]