Key features
LoRA training
Train Low-Rank Adaptation weights for any supported model. Produces small, portable adapter files in safetensors format.
Fine-tuning and DreamBooth
Native fine-tuning and DreamBooth training for SD 1.x/2.x, SDXL, SD3/SD3.5, FLUX.1, and LUMINA.
Textual Inversion
Learn new token embeddings for SD and SDXL without modifying model weights.
Image generation
Generate images directly from the command line using any supported checkpoint.
Dataset utilities
TOML-based dataset configuration, WD14 automatic tagging, bucket-based resolution grouping, and multi-resolution dataset support.
Model utilities
Convert model formats, merge LoRA weights, resize LoRA ranks, and more.
Supported models
| Model | LoRA | Fine-tuning | DreamBooth | Textual Inversion |
|---|---|---|---|---|
| Stable Diffusion 1.x/2.x | Yes | Yes | Yes | Yes |
| SDXL | Yes | Yes | Yes | Yes |
| SD3 / SD3.5 | Yes | Yes | Yes | No |
| FLUX.1 | Yes | Yes | Yes | No |
| LUMINA | Yes | Yes | Yes | No |
| HunyuanImage-2.1 | Yes | No | No | No |
Training methods
sd-scripts supports several training approaches, each suited to different goals:- LoRA — The most resource-efficient method. Trains a small set of rank-decomposed weight updates that are stored separately from the base model. Recommended for most use cases.
- Fine-tuning (native training) — Updates the full model weights. Requires significantly more VRAM and storage.
- DreamBooth — Fine-tunes the model on a small set of subject images while preserving the base model’s prior knowledge using regularization images.
- Textual Inversion — Learns a new token embedding to represent a concept without changing any model weights. Available for SD 1.x/2.x and SDXL only.
- ControlNet-LLLite — Lightweight ControlNet training for SD/SDXL.
- LECO — Model editing via contrast-based optimization for SD/SDXL.