Things are progressing very rapidly in this dawn of the AI and likewise for the AI Horde. I thought it would be a good idea to post about all the things that changed and improved in recent days for our service.
More Requests. More statistics.
I’ve deployed endpoints to measure the usage of the AI horde. Now that one month has passed, we can take a look.
- Per day, we are averaging 356,378 images (3.7 terapixelsteps) and 45,248 texts (4 megatokens)
- In the past month, we produced 11,475,183 images, generating a staggering 127.6 terapixelsteps. Text has also picked up significant speed since merging the hordes with 1,241,895 generated texts for a total of 112.8 megatokens!
Top 10 Stable Diffusion models
The AI Horde offers close to 200 models at the same time. Our statistics allows us to see how the popularity of the various models changes day to day and month to month. The below are just the top 10 models being used.
- Deliberate 22.2% (2550591)
- stable_diffusion 15.1% (1730426)
- Anything Diffusion 11.0% (1257688)
- Hentai Diffusion 4.1% (468473)
- Realistic Vision 3.0% (338742)
- Counterfeit 2.7% (310337)
- URPM 2.6% (297853)
- Project Unreal Engine 5 2.5% (289006)
- waifu_diffusion 1.8% (211572)
- Abyss OrangeMix 1.8% (205268)
For the longest time SD 1.5 (stable_diffusion above) was king, but in the past month, Deliberate has confidently taken the lead and has been leading the pack with a staggering 20% of all image requests passing through the AI Horde! This speaks very highly for the popularity of the model
Top 10 Text models
Almost as many text models exist for the AI Horde, but they’re more varied. However last months saw the release of two big milestones, the Pygmalion models for chat-like generation, which happened after the gimping of the Character AI models. The new Llama model was also released, bringing unparalleled miniaturization of the model size, allowing consumer GPUs far more coherence.
- PygmalionAI/pygmalion-6b 52.4% (651566)
- KoboldAI/OPT-13B-Erebus 14.0% (174393)
- KoboldAI/OPT-6.7B-Erebus 6.7% (83249)
- KoboldAI/OPT-6.7B-Nerybus-Mix 3.8% (46747)
- KoboldAI/OPT-13B-Nerybus-Mix 2.8% (35110)
- KoboldAI/OPT-13B-Nerys-v2 2.7% (33667)
- Facebook/LLaMA-13b 1.9% (23367)
- KoboldAI/OPT-6B-nerys-v2 1.9% (23232)
- OPT-6.7B-Nerybus-Mix 1.6% (19268)
- KoboldAI/OPT-2.7B-Erebus 1.0% (12464)
We can see Pygmalion has immediately dominated text generation, with Mr.Seeker’s storytelling models mopping up the rest, but the Llama Ascendancy is just beggining!
Ratings, botting and counter-measures
A few months ago we started collecting ratings for the LAION non-profit to help improve the models existing in the commons, as the success of midjourney has a lot to do with them training their models with the best images their previous generation created.
The initial design was very simple to allow integrators to onboard it fast and giving good kudos rewards for those helping us. Unfortunately people almost immediately started abusing this by creating bots to rate randomly, therefore poisoning our collection’s accuracy.
I always knew this was a possibility but I was hoping I wouldn’t be forced to add countermeasures quite so soon. So I spent quite a few days adding a captcha mechanism (along other things) to block at least the low hanging fruit.
It immediately led to a drop in ratings per day which automatically shows just how much damage botted ratings were doing
We are fortunate enough to have gathered some great collators for the inference aspect of the AI Horde. So I wanted a big shout-out.
- ResidentChief has stepped up strongly to help add new features and squash bugs in the nataili library. As a result the AI Horde now supports inpainting on many more models, a lot more post-processors, such as more upscalers and background removers, controlnet improvements, and so many other stuff too numerous to mention. They’re a beast!
- Jug has been working on improving the AI Horde worker practically non-stop. Giving us a great terminal control, and improving the webui. Plus a lot of bugfixes and improvements in the bridge part of things
- Tazlin who’s been doing a great deal of tech support in the channels as well as helping me detect and figure out malicious ratings. And also sending some code improvements as well!
- Aes Sedai who’s been putting a ton of work on improving the moderation capabilities of the AI Horde with a custom frontend.
And of course all the frontend integrators like rockbandit, aqualxx, sgt.chaos and concedo, who’ve been keeping the frontends up to date, with a lot of features smartly using the capabilities of the AI Horde in ways even I had not expected!
CI/CD and pypi
I finally got around to adding CI/CD pipelines for AI Horde Worker and nataili. Now they will be automatically versioned when the right tag is applied to a PR. The Nataili package has also been republished to pypi and will also automatically receive new versions whenever we publish a new release on GitHub.
The notifications also automatically publish a notification on discord, so people can be aware when something new is up.
Using the new post-processing improvements from ResidentChief, I’ve expanded the interrogation worker so that it can now perform post-processing on images, as well as img2text operations. Unfortunately the previous name didn’t fit so well, so now I’ve renamed it to “Alchemist”, to signify it’s capability to convert images to something else.
Likewise, the official names for image worker is now “Dreamer” and text worker is now “Scribe”. Why not 🙂
The pace of progress in this space is mind-blowing. I can’t wait to see what we achieve together in the coming days!