Ecosystem

---

Python	License	Deploy	Quality

---

Train, tag, and generate — a full LoRA & checkpoint training suite with ComfyUI built in.

A web interface built on Kohya SS. Prepare datasets, auto-tag, train across many model types (SDXL, Flux, SD3, Lumina, Anima…), then generate and test results in a bundled ComfyUI workspace.

---

Table of Contents

• Requirements
• Installation
• Starting the App
• Updating
• Features
• Testing
• Documentation
• Support

---

Requirements

Hardware

• GPU: NVIDIA with CUDA 12.1+
• Disk: 50 GB+ free space

GPU	Status
NVIDIA (CUDA 12.1+)	Fully supported
AMD (ROCm)	Community supported — see AMD ROCm docs
NVIDIA via ZLUDA	Community supported — see ZLUDA
CPU-only	Not recommended for training; tagging/captioning will work

While baseline VRAM requirements can be high, the community has figured out plenty of workarounds to train on constrained hardware.

Software

• Python: 3.10, 3.11, or 3.12
• Node.js: 20.19+ (22.x recommended)
• Git

Full platform requirements are in documentation/installation/INSTALLATION.md.

---

Installation

Windows Users: Install to a path you own (e.g. C:\Users\YourName\Projects\). Restricted system directories (C:\, Program Files, OneDrive folders) will cause permission errors.

Windows:

git clone https://github.com/UselessToys/Ecosystem_WebUI.git
cd Ecosystem_WebUI
install.bat

Linux:

git clone https://github.com/UselessToys/Ecosystem_WebUI.git
cd Ecosystem_WebUI
./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. Both auto-configure on launch.

---

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

# Restart with Reinstall + Restart supervisory process (Remote only)

bash fetch-restart.sh
./fetch-restart.sh 

URLs:

• Frontend: http://localhost:3000
• Backend API: http://localhost:8000
• API docs: http://localhost:8000/docs

---

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

---

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 models directly into your local folders
• Arc En Ciel browser: browse and download from multiple model sources with unified search
• HuggingFace downloads for training base models
• LoRA Manager for batch downloading from Civitai and HuggingFace via Python or Aria2

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
• 24+ checkpoint merge modes via Chattiori (WS, DARE, cosine blending, ReBasin, and more)
• LoRA-to-checkpoint baking (multi-LoRA bake across all architectures)
• HuggingFace upload after training

---

Testing

Tests validate the API and service layer without requiring a GPU, PyTorch, or a training run. The ML stack is mocked at the subprocess boundary.

# Install dev dependencies
pip install -r requirements_dev.txt

# Run all tests (~5 seconds, no GPU needed)
pytest tests/test_api_plumbing.py -v

# Skip tests that require real hardware
pytest tests/ -m "not slow"

Area	Coverage
Path validation	Drive-letter casing; out-of-bounds paths rejected
Training endpoint	Validation failures return success=False; malformed payloads get 422
Subprocess launch	-u flag and PYTHONUNBUFFERED=1 present for log streaming
Config paths	TOML config dir anchored to project root
Job failure logging	Full traceback written to app.log on crash
Log buffer	Subprocess stdout lines reach the buffer the UI polls

New tests go in tests/. Tests requiring a real GPU: mark with @pytest.mark.slow.

---

Documentation

• Installation Guide — full platform setup, manual installation, troubleshooting
• Beta-Planning — Bug Tracking and Beta Planning.
• Integration Strategy — Workflow Planning with Comfyui in mind.
• Attribution - Full Credits and Licenses.
• General Guides — usage documentation
• Training: General
• Training: SDXL
• Training: Flux
• Training: SD3
• Training: Anima
• LyCORIS Algorithms
• Contributing · Security

---

Support

• GitHub Issues — bug reports and feature requests

When reporting a bug, include error messages, logs (logs/app_YYYYMMDD.log), and the output of:

diagnose.bat      # Windows
./diagnose.sh     # Linux

---

License: MIT — see LICENSE