Breeding Season Estrus: FAQ Part 2

Support Breeding Season Estrus on Patreon to gain access to the prototype!

Who even are you?

I'm Hartista Pipebomb (AKA H-Bomb), the creator of the original Breeding Season, and I’m now back to finish what I started an entire decade (!!) ago. I studied computational neuroscience in college and have degrees in biology and psychology. I’ve always been intensely interested in the intersection of cognitive science and data science. Out of college I was a bioinformatician working on Huntington’s disease research, until I stumbled ass-backwards into heading a team making a sex game that was the highest paid project on Patreon for a time. That project had a very sudden and unfortunate ending, but I learned a massive amount in the process. 

Where were you all this time?

Studying the blade.

For a summer I gave kayak lessons on the Chicago river so I could finally be outside for a change, but after that I took on a senior software engineering position at a menswear e-commerce company where I obtained a ludicrous number of neckties, headed up a project to rewrite their entire front end, did some fancy analytics and UX research, and then eventually moved on and joined a start-up focused around using AI to evaluate advertisements for racial discrimination. While there I got my machine learning chops up to snuff with the current state of the art, ran a lot of experiments, processed a lot of data, performed a lot of analysis, and ended up rewriting yet another full front end (I’m looking forward to never writing JavaScript again in my entire life). 

In that time I gained extensive experience in software architecture design, agile best practices for planning out, managing, and tracking a project end-to-end (as well as the really critical stuff you only really get from experience like how to adjust project plans on the fly in response to unforeseen blockers so at the very least something is always getting done, how to code for reusability and repurposability so you throw out as little as possible when plans change and you never have to write the same code twice, how to actually cut things when they need to be cut and manage scope to prevent projects from ever dying a bloated death, etc), I finally found out how to tell the difference between all the important patterns and anti-patterns which has made looking at my old code an incredibly cursed endeavor, as well as all the very basic important stuff I never got to learn before I started the first Breeding Season project like version control and continuous deployment (DevOps in general, really) why you always want to have separate dev, staging, and production branches and how to coordinate with QA testers effectively to make sure that you never let something broken sneak past staging onto production, how to write good unit tests etc etc etc.

If you have been spared from ever working close to software development and you don’t know what any of that means; the point I’m trying to convey here is that when I started the first Breeding Season there was an entire world of things I just didn’t know, and most of them were things I didn’t even know that I didn’t know.

Basically, in the year 2013 in real, production-level software development terms I was a tiny wee little baby who didn’t even know that they were actually a tiny wee little baby and was constantly being caught off-guard by repeated revelations that I might, in fact, actually be a tiny wee little baby. Now that I’ve actually been through it all and made it all the way through shipping multiple full feature-complete applications that have supported traffic from millions of users and are still generating tens of millions of dollars in revenue each year for the companies I built them for; I’m now an incredibly huge, very strong, muscular baby, who can fight.

Why should I donate? 

As you might expect, I’ve spent a massive amount of time thinking about Breeding Season since I started the original project an entire decade ago. I’ve known for a very long time all the things I’d do differently were I to start over again from the beginning, the way I would design the game and the story I would want to tell with it, and most importantly how I would plan out and execute the project to ensure that it was air-tight against any eventuality that could get in the way of its completion. I have already completed considerable design work to that end, and worked out development models that experience has taught me fit perfectly for the kind of project I’m building here. I’ve just been waiting for when I felt it was the right time to finally put all this into action and try again.

One thing that I have been most adamant about is fundamentally designing the project in a scalable manner that can be adjusted very fluidly in response to the amount of resources at our disposal. I’ve set up asset pipelines in such a way that as patron donations grow, I can quickly and easily commission more and more work to incorporate into the project through limited contract engagements with artists, writers, and programmers without the need to have long training periods and without the need to manage a large long-term team in a way that benefits both sides of the relationship. Because I’ve taken the pains to work out how this can all be automated beforehand, it means there will be extremely little struggle in figuring out how to turn any amount of donations I get into more content.

Furthermore, with the use of machine learning models notions that would’ve been completely impossible before; such as massively more monster combinations and variations, hybrid monsters that are mixes between multiple species, multiple different kinds of sex scenes for every pairing, etc, all become infinitely more possible with simply enough processing power churning them out (after experimenting enough to be able to automate them to a high level of quality). Basically, every little bit of funding the project gets can be turned very efficiently into more and higher quality content.

But even more than that…I want to be clear; if I just wanted to make a lot of money for myself, I wouldn’t be doing this. If you don’t believe me; look up what the average salary of a senior machine learning engineer is. That’s what I’m giving up to do this. And the reason is because, as silly as it may sound for what is fundamentally a goofy porn game…this is actually really important to me.

Partly it’s because I want to finally make good on those promises I made so long ago, and I want the people who were so kind to me then to know that in spite of everything I never gave up, and I still haven’t given up now, nor have I forgotten their kindness.

But even more than that, the longer I spent thinking about why I ever even made that janky little flash game in the first place, the more I started to realize that there was actually always something buried deep in there incredibly important to me that I had wanted to say. And seeing how the world has changed since then, and the ways that it hasn’t…it’s still something I need to say. I know this is all a bit needlessly cryptic, but I can’t think of a better way to fully communicate it than through the game itself, and my sincere hope is that when I finally finish it and all is said and done, you’ll understand exactly what I mean.

Isn’t AI art wrong?

I know that AI art is an incredibly controversial topic at the moment, and for understandable reasons. I have some very strong opinions on it myself, and admittedly part of the reason I felt it was a good time to start this project is because I wanted to directly contribute to this discussion.

My goal is to, through broader public discussion and with Patron feedback and input, help build an example for what the ethical application of these kinds of machine learning models should look like, and try to bring a lot more nuance and understanding to a general audience about what can possibly go into (and come out of) these models and how individual artists can utilize them in addition to how these models can be utilized very directly against the interests of artists or in ways that harm people more generally. 

From just my own experiences so far, I think that machine learning can facilitate artists to leverage their own work, ability, and passion to create things that were impossible previously, and that building and manipulating models through creative applications of training and merging techniques could eventually become recognized as a new form of art in itself in the vein of things like song covers and remixing. However, like music remixing, we need to reach understandings on what's acceptable and expected regarding proper crediting of and approval from the artists whose work the model is derived from as well as to make sure that we don't allow corporations to use this tech to try to remove the artists from the art altogether or otherwise apply them in ways that would further exploit an already abused workforce and would be disastrous for both artists and people who enjoy the art they produce.

More importantly, I believe that we have an ethical duty to keep these models and the architectures they're built on as open source as possible and ensure that all state-of-the-art AI development is done completely transparently and out in the open: it is imperative that we allow everyone to have access to the means to understand these models and how they work on a deep level, that we understand and quantify the capabilities, dangers, and limitations of these models in a concrete manner, and that we form open communities to build upon and spread the benefits of these models to people other than the already ludicrously rich and powerful who would consolidate them under corporations and opaque institutions under the false premise of protecting us from them. 

I believe strongly that the only real viable counter to malicious use of machine learning is the protective use of machine learning, and the most critical part of preventing abuses is for the general public to have the opportunity to understand this technology, what it can and can't do, and how it works for themselves in a direct and hands-on manner.

As an aside, I stand firmly with SAG-AFTRA and the WGA, as well as every other union fighting for the right to dictate how workers and artists will be made to interact with this technology and how it will influence their day-to-day lives and their personal livelihoods moving forward. I believe that, if there is any level at which we must successfully challenge the ways that corporations and governments plan to abuse this technology to harm workers and consumers, it must be done at the level of worker organization and through collective bargaining and labor action. Only the people who will be working with these models directly and/or the content creation pipelines that will be based around them on a daily basis can ever truly have the power to ensure that they aren't being abused.

What kind of AI models are you using?

My own custom models used in this project so far are the result of ~6 months of experimentation, training, editing, and block merging with a lot of trial and error involved, focusing on trying to keep up with the massive number of new developments and really understand diffusion models inside and out. 

Like most out there, my current models are all based in Stable Diffusion 1.5 and the open source ecosystem around it. For developing my models, I started with a base of a large number of different open source SD checkpoints (about 20 in all) and successively block merged them until I got something that I felt was both an interesting overall style and also produced output I judged adequately distinct from all the individual models it was merged from to represent something wholly new. I then used a mix of explicitly license-free artwork, renders using simple models I made myself, renders of free 3d assets, and renders of a large number of 3d assets I'd purchased, upscales and photoshop edits of images generated with the models themselves, and occasionally my own doodles (feels funny to be drawing again after all this time) in order to train and refine the models, create LoRAs, and develop a pipeline for producing assets for use in this game.

It remains a work in progress in a rapidly developing field and I'll have to do much further experimentation moving forward, but what I've seen so far has been extremely promising and I have a dozen new ideas I want to test every day. The images you see in the game currently are all works in progress or placeholders; in the long run I am shooting for a much more consistent and coherent visual style, but consistency along with flexibility while automating the work flow as much as possible to create as many unique variations as possible in-game is a very big challenge. My next step is to start building some newer, better models off the base of SDXL using what I've learned, which promises much better output than 1.5 in the long run. It will also offer me the opportunity to rely less on existing 1.5 models and unaccountable chains of training data, and especially if the Patreon funds are good then my plan is to commission artists for unique assets to use for creating training material and compensating them as they deem fair for the explicit use of their work in model production.

Though my focus has been on taking great pains to try to create work I definitely feel is distinct from the prior work of others that went into it, as I've slowly developed my own workflows and models I've still had to rely heavily on open source models trained by others on unknown datasets. I've tried to avoid models that clearly egregiously misuse the work of other artists or people's likenesses, but I also can't account for the training data that went into every model that I've used in producing what I have so far or its chain of providence. If you are an artist and you see something present in the game that appears to derive from a model overfit on a particular work of yours, don't hesitate to contact me and we can work together to address it. One thing I have been experimenting with is training particularly overfit images out of models by subtractive merging.

On the subject of ownership and copyright of the AI-generated assets in this game: it remains to be seen how the law falls on these things. This game is, ultimately, a game I am intending to release to the general public for free to everyone, at least in its release version. I won't make and am not currently making any legally binding statements myself regarding the copyright status of these assets until further decisions are made by the courts and I've consulted thoroughly with my own lawyers etc, but in a purely philosophical rather than in any legal sense: I believe this work belongs to everyone, after all that's the only reason I wanted to make it in the first place anyway; to share it with everyone. Again, I won’t be able to make any binding promises until we’ve worked out the legal situation exhaustively, but if possible I would like to have all the AI-generated assets made for this game be freely available to anyone to reuse for their own projects in any way that they would like, and at the moment I’m looking to pursue getting them covered under a copyleft-style license that could ensure that (that is, if they can even be licensed at all).