Stable Horde

What is Stable Horde?

Stable Horde is a free, volunteer-powered network for AI image and text generation, now running under the name AI Horde. People with spare graphics cards run open-source worker software, and anyone can send prompts to them through a shared queue.

The project’s own documentation compares it to Folding@home and SETI@home, the long-running schemes where volunteers donate computing power. When you submit a request, the horde places it in a first-in, first-out queue ordered by kudos, splits multi-image requests into separate jobs and hands each job to the next suitable worker. You can check progress at any time and collect finished images even before the whole batch is done.

It began in September 2022 as KoboldAI Horde, a text service, and the Stable Horde image fork followed soon after. In February 2023 the two merged as AI Horde, since the network no longer served only Stable Diffusion, and the old stablehorde.net address still shows AI Horde pages. The developer known as db0 created it, and Haidra, a volunteer-run non-profit, now operates it with public donations and NLnet NGI0 Core grants. Its public site and documentation were active in early October 2026.

Key Features

Shared volunteer GPUs and a kudos queue drive Stable Horde, letting anyone generate images or text without owning a graphics card or paying a fee.

• Volunteer worker network: Requests are processed by home and hobby machines running the horde-worker-reGen software for images or bridges such as KoboldCpp and Aphrodite Engine for text. Workers poll the horde for jobs and can go offline whenever they like, using a maintenance mode first.
• Kudos priority queue: Kudos are points that decide your place in the queue rather than whether you are served at all. The baseline is 10 kudos for a 50-step 512 by 512 image, and a dry-run option quotes a request’s cost before you spend anything.
• Image generation controls: The image API accepts img2img, inpainting, outpainting and remix requests, plus ControlNet guidance and hires fix. Sizes run from 64 to 3,072 pixels per side in steps of 64, with up to 20 images per request.
• Community models and LoRAs: Users pick from many community Stable Diffusion checkpoints, including SDXL and Stable Cascade. LoRAs and textual inversions can be requested by name or by CivitAI model ID, so niche styles are available without any downloads.
• Special modes and post-processing: Transparent backgrounds come from Layer Diffuse, and a QR code workflow turns codes into artwork. Optional post-processors include RealESRGAN and other upscalers, GFPGAN and CodeFormers face fixes, and a background remover.
• Text generation: The same horde serves open large language models to tools such as KoboldAI Lite and SillyTavern. Text workers run their own backends, so the list of available models changes with whoever is online at the time.

Pros

Zero cost is why most people try Stable Horde: you can generate anonymously with no card or account, and the team states it will never charge for access.

• No payment at any level: Anonymous use, registered accounts and worker accounts are all free, and kudos cannot be bought. Students, hobbyists and small projects therefore get Stable Diffusion and open text models without a subscription.
• Runs on modest hardware: All heavy processing happens on volunteer machines, so a phone or basic laptop with a browser is enough. The FAQ points out that this also saves you from downloading multi-gigabyte model files and fixing Python dependencies.
• Open source and self-hostable: The middleware is published under AGPL-3.0 and runs on Python, PostgreSQL and Redis with Docker support. Organisations can host a private horde, though they must then supply their own workers.
• Free REST API: Developers get a documented REST API without usage fees, which is rare for image services. Game makers, bot authors and plugin builders can add generation to their projects and even power it with their own GPU.
• Fair, transparent economy: Kudos never expire, and registered OAuth2 users keep a minimum balance of 25. Workers earn kudos for every job and for each 10 minutes online, and lower-spec owners can earn by rating images.

Cons

Unpredictable waiting is the catch with Stable Horde, because speed depends entirely on how many volunteer workers are online compared with current demand.

• Variable queue times: A busy evening can mean long waits, especially for large images, rare models or anonymous requests. Anonymous use may also be restricted under heavy load, so registering for a free key is the practical workaround.
• Uneven output quality: Results depend on the model and worker that pick up the job, and settings that suit one checkpoint can fail on another. Expect to rerun prompts more often than with a single polished commercial service.
• Limited privacy: The horde itself keeps prompts only in memory, but a modified worker could log what passes through it. The team’s own advice is to write prompts as if posting in a public forum.
• Surprise censored images: Some workers run their own NSFW filter or word censorlist and return a black placeholder image instead. A horde-wide anti-CSAM filter also blocks certain prompts and can trigger short IP timeouts.
• Community-only support: There is no paid support desk, so help comes from the FAQ, GitHub documentation and a Discord server. Terms such as kudos, workers and alchemists also take a little learning before everything makes sense.

Who is Using Stable Horde?

Hobbyists without a capable graphics card form the largest group of Stable Horde users, alongside developers who need free generation inside their own apps and bots.

• Hobby artists: People with older laptops or phones use web clients such as ArtBot to make concept art and illustrations. They can rate images while waiting to earn extra kudos.
• Developers and game makers: Builders call the REST API from games, websites and chat bots. Some run a worker on their own GPU so their tool always has capacity available.
• Role-play and fiction writers: KoboldAI Lite and SillyTavern users connect to the text horde for open language models. This gives them story and character chat without running a large model at home.
• GPU owners: Volunteers share idle cards to support the community and bank kudos for faster personal requests. Owners can also prioritise their own and friends’ requests on their worker.

Pricing

Stable Horde costs nothing to use, and workers only pay for their own electricity (official site and GitHub documentation, checked October 2026).

Each request also burns a small horde tax of 1 kudos, plus 1 per job, with a 20 per cent surcharge for fast-only workers. Selling or buying kudos is forbidden, and accounts caught doing so can be reset to zero. One 2026 review mentions paid priority tiers, which official pages do not show.

Disclaimer: Kudos costs and earning rates are set by volunteers and may be adjusted at any time to keep the queue working.

What Makes Stable Horde Unique?

Idle cards in ordinary homes do the work, rather than rented cloud GPUs, so the service has no compute bill to pass on. Requesters pay in community points that cannot be traded for money, and the team says it is openly hostile to blockchain schemes.

That model also explains its stance on data. The FAQ argues that free centralised services often turn users into the product, whereas the horde publishes how its costs are shared. Paid services may be built on top, but only if they give back at least as much compute as they use.

Compatibilities and Integrations

Stable Horde connects to almost anything through its public REST API, and the community has built web, desktop, chat and creative-app clients on top of it.

• Web clients: ArtBot is the recommended starting point, with Lucid Creations and other community frontends listed in the project wiki. All work in a mobile browser without installation.
• Creative software: Community plugins bring generation into Krita, GIMP, Photoshop and Blender, and one directory also lists iOS and Android apps.
• Chat and text tools: Discord, Telegram and Reddit bots use the horde, while KoboldAI Lite and SillyTavern connect for text generation.
• Developer access: A REST API and a Python command-line script support custom integrations, and transfers between accounts use a dedicated kudos endpoint.
• Workers: Images run through horde-worker-reGen, and text workers can bridge KoboldAI, KoboldCpp or Aphrodite Engine.

Stable Horde Tutorials

ArtBot plus a free API key gets most people from nothing to a first image in a few minutes.

1. Register on aihorde.net, ideally with a Google, GitHub or Discord login so the key can be recovered.
2. Copy your API key and store it safely, because pseudonymous keys cannot be reset.
3. Open ArtBot or another client and paste the key into its settings.
4. Choose a model, write a prompt and set size, steps and any LoRAs.
5. Submit the request, then rate images while you wait to earn kudos.
6. Run a worker on a spare GPU if you want faster results later.

How We Rated It

Stable Horde earns 3.9 out of 5, with a near-perfect cost score held back by queue speed and the privacy limits of volunteer hardware.

Summary

Stable Horde suits hobbyists, students and developers who want free images or text without buying hardware. Its feature set is broad, covering inpainting, ControlNet, LoRAs and transparent images through one API. The main limitation is speed, as busy periods bring long waits and output quality varies by worker. Keep personal or sensitive details out of prompts, since volunteer machines process them. If you own a decent GPU, running a worker is the quickest route to priority.