KNX Trainer Updates - June 2026

Introduction

Hey! This is a great start to the northern hemisphere's summer season, an amazing start with some great new things coming towards our training ecosystem! We're aware it's literally just a skinned UI with a lot of extra toys and Fast API, but we're really proud of it. That's why we're extremely eager to share with you what's going on!

What's New

Well for starters, we've added a custom UI frontend in our system for GENERATION! Yes! We are vendoring ComfyUI for remote VastAI & Runpod especially! We would've done submoduling but it's a little difficult for timing currently with the way proxying works since we're not using a docker container yet for the system. Once we're in a more stable position soon with a few changes to main, we'll produce a stable remote container for remote nerds like ourselves! Local peeps: you likely SHOULD be able to just use your own, but DO NOT QUOTE ME ON THAT.

We are also in the midst of adding a Supermerger/dare/lora baking system without requiring A1111 NOR ComfyUI for it. Our plans are usually to rely on ComfyUI for things that are GEN-adjacent, or things that Node/python really can't do without an engine behind it easily. (Or that are too hard basket for my pea brain to figure out.)

What's Changed

We have been vendoring custom SD-scripts code, and so things are ROUGH off and on as we figure out what's different. We could yes, have been smart and gone straight for NATIVE Sd-scripts that's more stable...

But what's the fun in that? (YES FINE THE REAL REASON: Custom Optimi.. nah ok)

No the real reason this is all a little patchy is more just cause since the old JUPYTER edition we used Derrian Distro, and it's not THEIR FAULT: I just vendored bleeding edge forks for a lot of really neat up and coming things.

The TAGGGING SYSTEM got a small overhaul but BLIP and GIT are still struggling right now, and That's on me to do some research.

Main is behind DEV yes, but soon we'll be PR'ing to MAIN to be able to offer a lot more fixes.

Features

Dataset preparation:

Drag-and-drop upload (individual files, zip archives, URLs)
WD14, BLIP, and GIT auto-tagging
Tag editor with bulk add / remove / replace, frequency viewer, and image gallery

Training:

Supported models: SDXL, SD1.5, Flux, SD3/SD3.5, Lumina, Anima, HunyuanImage
LoRA types: Standard, LoCon, LoHa, LoKr, DoRA
40+ community presets (AdamW8bit, CAME, Prodigy, AdaFactor, Compass, Lion8bit, and more)
Live training monitor: step count, loss, learning rate, ETA
Preset system with form persistence

Models:

Civitai browser: search, filter, and download directly into local model folders
HuggingFace upload after training

Image generation (ComfyUI):

Bundled ComfyUI workspace, installed by default (local and cloud) — generate images to test your trained models without leaving the app
LoRA Manager for browsing and applying your LoRAs

Utilities:

LoRA merge, resize, and metadata editing
Checkpoint merge

What We Need From You!

WINDOWS AND NON STANDARD LINUX TESTERS! - Aka: Local users. We KNOW this works 99% of the time on at least A4500 and 4090s. Anything that can run CUDA can run it.

If you're a ROCm or ZLUDA customer: TEST IT.

MAC SILICON USERS: If you know how to get it to work, TEST IT.

Everysense of this is: We NEED MORE TESTERS On all around, the addage of "IT WORKS ON MY MACHINE" happens to only be the "I tested it on my machine but i can't train locally."...

I use Vast/Runpod interchangably to offer more, so i'm also looking for people who can try this for Kaggle, Sagemaker and even see if there's suggestions to get it to work for COLAB.

Where to Find

In a big round room with small corners...

Er -- no the worms didn't come here yet, here's the link: https://github.com/Ktiseos-Nyx/Ktiseos-Nyx-Trainer/

How To Install

Windows:

git clone https://github.com/Ktiseos-Nyx/Ktiseos-Nyx-Trainer.git
cd Ktiseos-Nyx-Trainer
install.bat

Linux:

git clone https://github.com/Ktiseos-Nyx/Ktiseos-Nyx-Trainer.git
cd Ktiseos-Nyx-Trainer
./install.sh

Both scripts will prompt about creating a virtual environment. Pass --venv or --no-venv to skip the prompt; --auto for fully unattended install.

ComfyUI (image generation) installs by default — including on cloud. To skip it, pass --no-comfyui (e.g. install.bat --no-comfyui, ./install.sh --no-comfyui, or python installer.py --no-comfyui on a remote instance).

Cloud: Use the VastAI or RunPod deploy buttons above. Both auto-configure on launch.

How to Run

You got two feet, just walk really fast -- OH YOU MEAN THE APP!

Starting the App

# Default: frontend on 3000, backend on 8000
start_services_local.bat    # Windows
./start_services_local.sh   # Linux

# Custom ports
start_services_local.bat --port 4000 --backend-port 9000
./start_services_local.sh  --port 4000 --backend-port 9000

# Quick restart without reinstall
restart.bat     # Windows
./restart.sh    # Linux

URLs:

Updating

git pull
install.bat       # Windows
./install.sh      # Linux

If reinstalling doesn't fix the problem:

# Preview what will be removed (dry run)
python clean_slate.py --dry-run

# Remove build artifacts, venvs, node_modules, caches (preserves models/datasets/outputs)
python clean_slate.py

# Also remove models, VAEs, datasets, and outputs
python clean_slate.py --nuclear