This is a very big subject which will be later be expanded in a devlog. But putting it simply:
A worker is computer with usually a mid-range or higher class GPU, who has installed specific software to generate images via Stable Diffusion or KoboldAI locally and it connected to the horde through what we call the "bridge". It is constantly polling the horde for new generations and receiving Kudos in return for creating them.
Generating images for others 24/7 technically costs electricity, we know.
People want to contribute to the horde for many reasons
Technically, yes. While the worker software and the bridge code is not set to allow it, ultimately the software resides at someone else's computer and is open source. As such, anyone with the know-how can modify their own code to not just see all prompts passing through, but even save all results they generate.
However workers do not have any identifying information about individual requestors as they cannot see their ID or IP.
Nevertheless always request generations as if you're posting in a public forum like using a discord bot. While the horde is technically more private than that, it's a good practice anyway.
Yes! We do not require workers to stay always online. We only request that you put your worker in maintenance before you do so to avoid messing with a currently running generation your worker might have picked up.
Yes, but you select your model manually. You can also not select a model and get the first model the next worker that picks up your work provides
Yes! By default, your own requests are always served first by your own worker, regardless of Kudos. But you can also specify specific usernames which will also be prioritized the same way.
Another big subject. This one actually has a devlog about it.
Connect a worker to the horde, that is all! You will generate kudos for each request you fulfil, relevant to its difficulty, and you will also generate kudos every 10 minutes your worker stays online.
The Kudos cost reflects the amount of processing required to generate the image.
As each payload on the horde can have too many variables which affect its speed, we have trained a neural network which takes as input a request payload, and calculates how much kudos it would require, based on how much faster or slower it would generate compared to a baseline of 10 kudos for a 50 step 512x512 image. The baseline costs 10 kudos. So if a payload is expected to take double that time, it will be valued as 20 kudos.
The AI Horde API provides a dry_run payload key. When set to true, it will return the kudos cost for an image, without actually requesting a generation.
On top of that there is what's known as the "horde tax" which represents the extra costs to the infrastructure for each request. These kudos are not received by the worker but are rather "burnt" forever.
We use Kudos to support good behaviour in the community. As such we have ways to receive Kudos outside of generating images for others (although that's the best way)
NO! Kudos is inherently valueless and we do not allow anyone exchanging kudos for money. Bypassing this requirement is an existential threat to the AI Horde. Please do not attempt to do this under the table. If you exchange money for Kudos and we discover it, we might zero out your account and whoever you bought it from!
No! Kudos is completely centralized and involved no blockchain tech whatsoever. The AI Horde is explicitly hostile to blockchain technologies and we will never integrate with any of them. Likewise, there is no way to convert kudos to anything other than favours benefiting the improvement of the AI Horde.
Yes, but you might have a smaller a pool of workers to fulfil your request, which will lead to longer generation times.
Horde-wide, we censor only one type of generation: CSAM Images. We have two mechanisms to achieve this, one is a regex replacement filter during initial API request. The other is the anti-CSAM AI running on each worker. See more detailed answers below.
Other than this one instance, the horde does not censor text generations at all, or images which do not appear to be CSAM.
However each individual worker might have its own censorship guidelines. And each requestor can voluntarily opt-in to accidental NSFW censorship.
Those generations have been NSFW-censored by the worker generating them. If you've specified your request as SFW, individual SFW workers who fulfil it might have the NSFW censorship model active, which will return just this black image. To avoid such images, turn on NSFW, or ensure your prompt is not too close to the edge of SFW/NSFW. If the image mentions censor due to the anti-CSAM filter, this cannot be turned off.
Each individual worker can optionally define a censorlist. If any word inside that list is found, the worker will automatically post-process using a NSFW censorship model. These words are things that should never be combined with NSFW content or would run into legal troubles for the worker if they did.
This means your censored images triggered one such worker's censorlist. You can rerun the prompt and hope you get a generation with a seed that doesn't trigger the NSFW model, or hope to get a new worker, or tweak your prompt.
If you feel a worker is using the censorlist maliciously, or improperly, please contact us with the content of your prompt and the worker name, and we'll address it.
When an image request first comes into the AI Horde, it is passed through a private regex filter looking for combination of two contexts: "Underage" context and "Lewd" Context. An example of an underage context might be "child" and an example of a lewd context might be "without clothes".
If none, or one of these contexts is detected in a prompt, then the prompt is allowed to pass through unaffected. For example "child in the playground" is ok. "Without clothes in the bathroom" is also OK.
If however both of these terms are present in the same prompt, this will trigger the regex protection. This has two modes of operation:
If the replacement_filter is true in your API payload and the prompt is less than 1000 chars, each triggering term will be automatically replaced by an "adult" version of those terms. For example "school" will be replaced by "university". This attempts to point the inference to try and generate safe content.
If however replacement_filter is false or your prompt is above 1000 chars, then the request is automatically blocked instead, and you get an IP timeout for a couple of minutes. This IP timeout increases further, every time you get caught by the regex filter. This is to prevent people deliberately trying to out the filter to reverse engineer loopholes.
We have written a detailed devlog about this
This project started as a way to consolidate resources for the KoboldAI Client. As we needed a name for it, I came up with something thematic for the concept of "Kobolds". "A Horde of Kobolds". When I started doing image generation as well, I kept the "Horde" part.
We have a dedicated wiki page where you can look a lot of the cross-referenced terms commonly mentioned in the context of the AI Horde
No, the horde itself is not storing such details. The prompts and the generations are only stored in-memory transiently and deleted shortly after the generation is delivered or cancelled.
Not everyone has a power GPU in their PC. The horde allows anyone to use fast Stable Diffusion and KoboldAI, not only the ones privileged enough to be able to afford an expensive graphics card.
Furthermore, local clients, even at the best of times, are difficult to setup up and often error prone due to python dependencies. They also need plenty of internet bandwidth to download 4GB of models. The stable horde provides no-install clients, as well as browser clients you can use even on your phone!
Finally if you wanted to provide a service built on image or text generation, you can now use your own PC to power your image generations, and therefore avoid all the complexity and capital costs required with setting up a server infrastructure.
Because the Horde is free! You will never need to pay to use the horde. Sure if the demand is high, your delivery speed might not be great, but that is true with other services like midjourney
Second, the Horde gives you all the benefits of a local installation, such as freedom in prompts, while still allowing a browser interface. and flexibility.
Finally unlike many of these services, Horde also provides a fully fledged REST API you can use to integrate your applications, without worrying about costs.
Because when the service is free, you're the product!
Other services which run on centralized servers have costs. Someone has to pay for the electricity, and the server infrastructure. The horde is explicit of how these costs are crowdsourced and there's no need for us to ever add anything in the future to change our existing model. Other free services tend to be deliberately obscure in how they use your prompts, results, and data, or explicitly say that your data is going to be the product. Such services eventually pivot their usercount to make money through advertisements and data brokering.
If you're fine with that, go ahead and use them.
Finally a lot of these services do not provide free REST APIs, so if you need to integrate with them, you have to use a browser interface, so that you can see the adverts.
Of course! This software is FOSS and you are welcome to use, modify and share, so long as you respect the AGPL3 license.
Yes, with some restrictions. Due to the voluntary nature of the AI Horde, you must give back to the AI horde at least as much as you take out to make a profit. Please see the detailed explanation in this devlog