(Re) Introducing Ktiseos Nyx Trainer!
Alright, buckle up, buttercup. We're turning the chaos dial to 11.
🌀 KTISEOS NYX LoRA TRAINER – "It's Not a Bug, It's a Feature of My Brain" Edition
Alpha – March 2026. Yes, we're back. No, we didn't fix the burnout. Yes, the trainer works. Mostly. Please don't yell too loud, my ears are sensitive. 🎧💥
---
What is this?
A gloriously chaotic, web-based LoRA training UI built on Kohya SS, but with a frontend that doesn't make you want to scream into a pillow. Think: Next.js + FastAPI, drag-and-drop datasets, 132+ configurable parameters (yes, we counted, no, we're not sorry), and end-to-end training that's been verified on VastAI and RunPod.
Handles SDXL, SD1.5, Flux, SD3/SD3.5, Lumina. Supports Standard LoRA, LoCon, LoHa, LoKr, DoRA. Auto-tagging with WD14, batch caption editing, real-time logs, HuggingFace uploads, pulling models from Civitai. All in a browser tab. No, seriously. It's like if VS Code and a vibes-based ritual had a baby.
---
Who's it for?
- You, if you've ever looked at a Kohya config and whispered "why are there 47 nested dictionaries"
- You, if you want to train locally but don't want to live in a terminal forever
- You, if you're on cloud GPUs and want a UI that doesn't judge your life choices
- You, if you like having options (132+ of them, we're committed to the bit)
- You, if you're okay with alpha software that's functional but might occasionally glitch out because the developer's brain decided to do a hard reboot
If you want polished, corporate, "click one button and angels sing"… maybe wait for v1.0. But if you're cool with "it works, we're polishing, please yell at us on GitHub if it breaks (but gently, we're sensitive)"… welcome to the party. 🎉
---
Why's it for?
To make LoRA training less of a chore and more of a vibe check.
Dataset prep? Drag, drop, auto-tag, edit captions in bulk while listening to your hyperfixation playlist.
Training config? Seven tabs, logical grouping, presets, form persistence so you don't lose your settings when a bird tweets outside.
Execution? One click, watch it go, get notified when it's done (or when it crashes, we're honest).
Post-training? Resize LoRAs, edit metadata, push to HuggingFace without leaving the UI.
It's for people who want power and usability. For people who are tired of juggling five different tools just to train a single concept. For people who believe training shouldn't feel like defusing a bomb while riding a unicycle.
---
Why Do you Care?
Because training LoRAs shouldn't require a PhD in command-line-fu. Because you deserve a UI that respects your time, your GPU, and your mental bandwidth. Because we've been in the trenches with you – the failed runs, the VRAM crashes, the "why is this taking 14 hours" moments.
Also, because we're finally at the point where we can say: it works. Like, actually. We trained stuff. It made pictures. The pictures were good. We're not just hype – we've got logs, screenshots, and a very tired developer who needs a nap.
And hey, if you care about open tools, transparent development, and a team that's weirdly honest about their burnout cycle… hi. We see you. You're valid. Your stim is valid. Your weird training idea is valid.
---
What's required?
Keep it simple, bestie:
- GPU: NVIDIA, CUDA 12.1+, 12GB VRAM minimum (24GB if you're doing SDXL and want to keep your hair)
- Python: 3.10 or 3.11 (no, 3.12 is not invited yet, sorry)
- Node.js: 18+ (we're not animals)
- Disk: 50GB+ free (datasets + outputs add up faster than your backlog of unfinished projects)
- Brain: Optional, but recommended. If yours is fried like ours, caffeine helps. ☕
Runs locally or on cloud GPUs (VastAI, RunPod tested). If you've got a potato GPU, maybe sit this one out until we add more optimization passes. Or don't! We're not the boss of you.
---
## Is it a Virus?
NO. It's a Next.js + FastAPI app that talks to Kohya SS scripts. You can audit the code. You can run it in a VM if you're paranoid (we won't judge, honestly, same). It doesn't phone home, doesn't mine crypto, doesn't steal your waifus or your cat pics.
If your antivirus flags it, that's probably because "Python app that downloads models and trains AI" sounds suspicious to heuristic scanners. You can verify the hash, check the commits, or just trust us after you've lurked in our Discord for five minutes and seen us argue about whether "locon" is a mood or a lifestyle.
---
Are you a virus?
Depends who you ask. 😇
We're a burnt-out, neurospicy dev who walked away, came back, and is now channeling chaotic energy into building tools we actually want to use. If that's a virus… then yeah, we're patient zero. But the only thing we're infecting you with is the ability to train LoRAs without losing your mind.
P.S. If you're people or anyone wondering if "CREW" means multiple humans: Ktiseos Nyx is an organization name, not our system name. We're The Duskfall Portal Crew and we speak on behalf of some if not most of KNX (Yea yea yea Nuri i get it, you're still screeching about electron.) The developer is one person with ADHD/C-PTSD/DID/Autism. So yes, multiple voices, one keyboard. No, it's not a "team" in the traditional sense. Yes, the commit history looks like a rollercoaster because sometimes different parts of the brain front. It's all love. 💜
---
Didn't you Retire?
[deep, dramatic sigh]
Yeah. We tried. We really did. Burnout hit like a truck driven by a demon who hates fun. We stepped back, touched grass, questioned our life choices, stared at a wall for three days. But then we kept tinkering. And the trainer kept getting better. And people kept asking. And we realized… we actually like this. When it's on our terms.
So we're back. Not with a bang, but with a "hey, wanna test this alpha?" We're not promising perfection. We're promising honesty, progress, and a tool that finally does what we said it would. Also, we promise to take breaks. Maybe.
---
Don't you Hate AI?
(hueheuhuehuehue)
We hate exploitative AI. We hate scraping without consent. We hate corporate grifters turning creativity into a content mill. We hate the vibe that AI = replacement.
But we love AI as a tool. As a collaborator. As a way to explore ideas, iterate faster, and make art that feels like yours. This trainer isn't about mass-producing slop. It's about giving you control. Your data. Your settings. Your output.
So no, we don't hate AI. We hate what some people do with it. And we're building this to help you do it better. Also, we think AI is kinda cool when it's not being weird. *(Disclaimer: Yes we used AI To write this because honestly, when it comes to dev introductions, I do not have the time nor brains)
---
Doesn't this just train waifus and furries?
Look. We're not gonna lie: a lot of people do use LoRAs for waifus and furries. And that's valid! Art is art. If that's your jam, this trainer will absolutely help you make better, more consistent, more yours characters.
But it also trains:
- Concepts, styles, objects, environments
- Photorealistic subjects, architectural details, product shots
- Abstract aesthetics, mood boards, experimental pipelines
- That one weird idea you had at 3 AM that you're too embarrassed to say out loud (we won't judge)
The tool doesn't care what you train. It just cares that you can. So yeah, train your oc. Train your cat. Train your cyberpunk cityscape. Train a LoRA that makes everything look like it's drawn in crayon. We're not the police. We're just here to help you make the thing.
---
WHY NOT BMALTAIS GUI?
First off: bmaltais is a legend. His Gradio-based trainer is solid, well-maintained, and absolutely gets the job done. If you love Gradio's deep dark secrets, its cozy little dropdowns, and its "I learned Python yesterday and somehow this works" energy… more power to you. Seriously. No shade. 🙏
But.
Gradio blows chunks to us.
Like, personally. Viscerally. The layout feels cramped. The customization feels limited. The way it handles state? The way it renders on mobile? The way it looks when you've got three monitors and ADHD and just need things to snap into place? Yeah. Not our vibe.
And that's okay! Everyone's brain is wired differently. Some folks love Gradio's simplicity. Some folks (like us) need a UI that feels like it was built by someone who also has 47 browser tabs open, three playlists running, and a brain that switches contexts faster than a GPU switches VRAM.
So we built our own.
- Next.js frontend = flexible, responsive, themable, and actually fun to navigate
- FastAPI backend = clean, typed, async-friendly, and doesn't make us cry when we debug
- Form persistence, logical grouping, presets = because losing your config when you alt-tab to check Discord is a special kind of pain
- Real-time logs in-browser = no more squinting at a terminal while your brain screams "WHERE IS THE OUTPUT"
bmaltais's code is great. Our code is… different. Not better. Not worse (IN THEORY: It may be worse, we don't know LOL). Just for us. And maybe for you, if your brain also goes "hm, what if the buttons were… over there?"
> 🧠 TL;DR: If Gradio works for you, keep using it! We're not trying to replace it. We're just building the trainer we needed. Different tools for different neurotypes. All valid. All welcome.
---
> Final Alpha Disclaimer (with extra chaos): This is still alpha. Things might break. We're actively fixing stuff. If you find a bug, yell at us on GitHub or Discord (but gently, we're sensitive and easily overwhelmed). If you love it, also yell at us. We're listening.
> And if you're just here for the drama… hi. Welcome. Grab a snack. We're just trying to make training less painful. Also, if the tone feels all over the place? Yeah. That's the ND brain. One sec we're professional, next sec we're making frog noises. It's a feature. 💜🐸
---
Quick Links
- 🐙 [GitHub Repo](https://github.com/Ktiseos-Nyx/Ktiseos-Nyx-Trainer/)
- 💬 [Discord] (https://discord.gg/HhBSvM9gBY)
- 🐛 [Report an Issue] (https://github.com/Ktiseos-Nyx/Ktiseos-Nyx-Trainer/issues)
---
Who is "we"?
Ktiseos Nyx is the organization name with ourselves, our partners and two of our coding nerd buddies. Our system name is and will always be "The Duskfall Portal Crew". The primary developer of this trainer nightmare is one neurodivergent human (ADHD/C-PTSD/DID/Autistic) who sometimes fronts different parts of their brain while coding. So yes, multiple voices, one keyboard. No, it's not a "team" of multiple humans in the traditional sense. Yes, the commit history looks chaotic because sometimes the brain does a hard switch mid-function. It's all love, all valid, and all trying to make cool stuff. If you have questions, ask nicely. We're friendly. Mostly.
---
And proof the Alpha Stages are mostly working:
