mirror of https://github.com/vladmandic/automatic
Page:
Loader
Pages
AMD MIOpen
AMD ROCm
API
Advanced Install
Backend
Benchmark
CHANGELOG
CLI Arguments
CLI Tools
CLiP Skip
Caption
Control HowTo
Control Settings
Control Technical
Debug
Detailer
DirectML
Docker
Docs
Enso
Extensions
FAQ
FLUX
Features
FramePack
Gated
Getting Started
Google GenAI
HiDream
Hints
Home
Hotkeys
HuggingFace
IPAdapter
Installation
Intel ARC
Kanvas
LTX
Launcher
LoRA
Loader
Locale
MacOS Python
Malloc
Model Support
Models Tab
Models
Networks Search
Networks
Notes
NudeNet
Nunchaku
ONNX Runtime
Offload
OpenVINO
Outpaint
Parameters
Performance Tuning
Process
Profiling
Prompt Enhance
Prompting
Python
Quantization
Reprocess
SD Pipeline How it Works
SD Training Methods
SD XL
SD3
SDNQ Quantization
Schedulers
Scripts
Stability Matrix
Stable Cascade
Styles
Theme User
Themes
Troubleshooting
Update
Using LCM
VAE
Video
WSL
Wildcards
XYZ Grid
ZLUDA
_ToDo
index
nVidia
3
Loader
Vladimir Mandic edited this page 2025-04-09 09:03:27 -04:00
Table of Contents
Custom Model Loader
- in Models -> Loader
- Can be used to load any known diffusion model with default or custom model components
Model Type
- Can be any currently supported model type
- Autodetect can be used to attempt loading repo for any model
e.g. there is a new model on huggingface that is not yet added to predefined list - well, try it out ;) - Current can be used to replace selected components without reloading model itself
Model Repo
Valid model repo on huggingface is required to access base model config
- Default model repo is provided for common models
- Model repo can be overriden to any valid repo on huggingface
Loadable Components
[!NOTE] Any model component that is marked as loadable can be overriden
If neither local or remote values are set, default from model repo will be used
Any loadable model component with set value will be loaded from that value
Local vs Remote
- Can use local or remote models
- Local model can be any valid path to safetensors file or folder with hf-style structure
- Remote model can be any valid huggingface repo, with or without subfolder or full url to file hosted on huggingface
Dtype
- Can be used to override torch compute dtype when loading each individual component
Quant
- Can be used to enable quantization when loading each individual component
- Quant settings will be used from settings -> quantization
Receipe
Model receipe can be saved to JSON file for later use
[!WARNING] Load/Save receipe is not yet implemented