73 lines
1.9 KiB
Markdown
73 lines
1.9 KiB
Markdown
# animepics
|
|
|
|
Anime image generation + LoRA training setup.
|
|
|
|
## Stack
|
|
|
|
- **ComfyUI** — image generation UI
|
|
- **NoobAI-XL** — base anime model (SDXL-based, SFW+NSFW)
|
|
- **kohya_ss** — LoRA training
|
|
|
|
## Requirements
|
|
|
|
- Python 3.10+ installed and in PATH
|
|
- Git installed
|
|
- NVIDIA GPU with CUDA (4070 recommended)
|
|
- ~20GB free disk space for models
|
|
|
|
## Setup
|
|
|
|
Run once to install everything:
|
|
|
|
```powershell
|
|
.\setup.ps1
|
|
```
|
|
|
|
This will:
|
|
1. Clone ComfyUI and install dependencies
|
|
2. Clone kohya_ss and install dependencies
|
|
3. Install ComfyUI custom nodes (AnimateDiff, ControlNet, etc.)
|
|
4. Download the NoobAI-XL base model + VAE
|
|
|
|
## Launch
|
|
|
|
```powershell
|
|
# Start image generation UI (opens browser at localhost:8188)
|
|
.\launch_comfyui.ps1
|
|
|
|
# Start LoRA training UI (opens browser at localhost:7860)
|
|
.\launch_kohya.ps1
|
|
```
|
|
|
|
## LoRA Training
|
|
|
|
1. Put your training images in `training_data/<your_lora_name>/img/10_<trigger_word>/`
|
|
2. Copy `training/example_lora_config.toml` and edit it
|
|
3. Launch kohya and use the GUI, or run `.\train_lora.ps1 <config_file>`
|
|
|
|
## Directory Structure
|
|
|
|
```
|
|
animepics/
|
|
├── comfyui/ # ComfyUI install (gitignored venv)
|
|
├── kohya_ss/ # kohya_ss install (gitignored venv)
|
|
├── models/ # shared model storage (gitignored)
|
|
│ ├── checkpoints/ # base models (.safetensors)
|
|
│ ├── loras/ # trained LoRAs
|
|
│ ├── vae/ # VAE models
|
|
│ ├── embeddings/ # textual inversions
|
|
│ └── controlnet/ # ControlNet models
|
|
├── training_data/ # LoRA training images (gitignored)
|
|
├── output/ # generated images (gitignored)
|
|
├── training/ # LoRA training configs
|
|
└── workflows/ # ComfyUI workflow JSON files
|
|
```
|
|
|
|
## Model Downloads
|
|
|
|
Base model (NoobAI-XL Vpred):
|
|
- https://civitai.com/models/833294
|
|
|
|
Good NSFW VAE:
|
|
- Already baked into NoobAI, but sdxl_vae.safetensors from stabilityai works too
|