-
-
Notifications
You must be signed in to change notification settings - Fork 2
Upstream Reference
Summaries of vendored upstream documentation. Full documents are in documentation/guides/ in the main repository.
The training backend is built on Kohya SS sd-scripts, vendored at trainer/derrian_backend/sd_scripts/.
| Model | Summary | Full Doc |
|---|---|---|
| General | Core training script for SD1.5/SD2.1 LoRA training. Covers dataset preparation, configuration, and training options. | documentation/guides/training/general/train_network.md |
| SDXL | SDXL-specific training parameters, resolution handling, and LoRA configuration. | documentation/guides/training/SDXL/sdxl_train_network.md |
| Flux | Flux.1 dev/schnell LoRA training. Supports fp8 base models for reduced VRAM. | documentation/guides/training/flux/flux_train_network.md |
| SD3 | Stable Diffusion 3 / 3.5 training with network support. | documentation/guides/training/SD3/sd3_train_network.md |
| Anima | Anima model training — high VRAM requirements. Uses clip_skip: 1 with specific token handling. |
documentation/guides/training/anima/anima_train_network.md |
| Lumina | Lumina model training pipeline. | documentation/guides/training/lumina/lumina_train_network.md |
| HunyuanImage | Hunyuan image training network. | documentation/guides/training/hunyuan/hunyuan_image_train_network.md |
| Advanced | Advanced network training with detailed parameter documentation. | documentation/guides/training/advanced/train_network_advanced.md |
| Dreambooth | Dreambooth-style training for SD models. | documentation/guides/training/dreambooth/train_db_README-ja.md |
| Fine-Tune | Full fine-tuning guidance. | documentation/guides/training/fine-tune/fine_tune.md |
| Textual Inversion | Textual inversion / embedding training. | documentation/guides/textual-inversions/train_textual_inversion.md |
| Document | Summary |
|---|---|
| Config Guide (English) | Full configuration parameter reference. |
| Training Config (English) | Training-specific configuration options. |
| Document | Summary |
|---|---|
| WD14 Tagger (English) | WD14 tagger configuration, model selection, threshold settings. |
| Document | Summary |
|---|---|
| Generation Guide (English) | Image generation with trained models. Generation parameters and workflows. |
| Document | Summary |
|---|---|
| Masked Loss (English) | Training with attention masking for focused region training. |
LyCORIS (LoRA beyond Conventional) provides advanced LoRA methods, vendored at trainer/derrian_backend/lycoris/.
| Algorithm | Description | Full Doc |
|---|---|---|
| LoRA | Standard Low-Rank Adaptation | documentation/guides/lycoris/Algo-Details.md |
| LoCon | LoRA with convolution layers | Same doc |
| LoHa | Hadamard product-based LoRA | Same doc |
| LoKr | Kronecker product-based LoRA | Same doc |
| DoRA | Weight-Decomposed Low-Rank Adaptation | Same doc |
| Document | Summary |
|---|---|
| Algorithm Details | Comprehensive breakdown of each LoRA variant, mathematical foundations, and when to use each type. |
| Algorithm List | Quick reference of all available algorithms. |
| API Reference | LyCORIS module API documentation. |
| Conversion Scripts | Converting between LoRA types and formats. |
| Demo | Usage examples and demonstrations. |
| Guidelines | Best practices for using LyCORIS modules. |
| Network Arguments | Network module parameter reference. |
| Preset Configuration | LyCORIS preset system documentation. |
| Resources | Additional references and links. |
Vendored at trainer/derrian_backend/custom_scheduler/. Provides optimizer implementations beyond standard PyTorch:
- CAME — Confidence-guided Adaptive Memory Efficient optimization
- Compass — Composite optimizer
- LPFAdamW — Low-pass filtered AdamW
- Lion8bit — Lion optimizer with 8-bit quantization
- Prodigy — Adaptive learning rate
- AdaFactor — Low-memory factorized optimizer
- D-Adaptation — Learning-rate-free optimization
20 upstream bugs identified by CodeRabbit on PR #361 and fixed in the vendored code. Documented in:
docs/upstream-bug-fixes-2026-05-05.md
Fixes span sd-scripts (anima_vae, lora_diffusers, oft, tagger, SDXL/SD3/Hunyuan/Lumina trainers, EDM2 loss, strategy files, cache utilities) and LyCORIS (full, regex_args, oft, oft_flux). Issues include API inconsistencies, variable shadowing, IndexError, AttributeError, device mismatch, and more.
The documentation in documentation/guides/ is copied from upstream sd-scripts and LyCORIS for reference. Not all context applies 100% to this system — sd-scripts is a native CLI tool, vendored here via the Derrian Distro integration. Native documentation within the training system is being added during the beta phase.