mirror of https://github.com/kohya-ss/sd-scripts.git synced 2026-04-16 08:52:45 +00:00

Files

Umisetokikaze 4ea6032c66 feat: SD1.x/2.x と SDXL 向けの LECO 学習スクリプトを追加 (#2285 )

* Add LECO training script and associated tests

- Implemented `sdxl_train_leco.py` for training with LECO prompts, including argument parsing, model setup, training loop, and weight saving functionality.
- Created unit tests for `load_prompt_settings` in `test_leco_train_util.py` to validate loading of prompt configurations in both original and slider formats.
- Added basic syntax tests for `train_leco.py` and `sdxl_train_leco.py` to ensure modules are importable.

* fix: use getattr for safe attribute access in argument verification

* feat: add CUDA device compatibility validation and corresponding tests

* Revert "feat: add CUDA device compatibility validation and corresponding tests"

This reverts commit 6d3e51431b.

* feat: update predict_noise_xl to use vector embedding from add_time_ids

* feat: implement checkpointing in predict_noise and predict_noise_xl functions

* feat: remove unused submodules and update .gitignore to exclude .codex-tmp

---------

Co-authored-by: Kohya S. <52813779+kohya-ss@users.noreply.github.com>

2026-03-28 19:25:16 +09:00

2.1 KiB

Raw Blame History

LECO Training

This repository now includes dedicated LECO training entry points:

train_leco.py for Stable Diffusion 1.x / 2.x
sdxl_train_leco.py for SDXL

These scripts train a LoRA against the model's own noise predictions, so no image dataset is required.

Current scope

U-Net LoRA training only
networks.lora is the default network module
Prompt YAML supports both original LECO prompt pairs and ai-toolkit style slider targets
Full ai-toolkit job YAML is not supported; use a prompt/target YAML file only

Example: SD 1.x / 2.x

accelerate launch train_leco.py ^
  --pretrained_model_name_or_path="model.safetensors" ^
  --output_dir="output" ^
  --output_name="detail_slider" ^
  --prompts_file="prompts.yaml" ^
  --network_dim=8 ^
  --network_alpha=4 ^
  --learning_rate=1e-4 ^
  --max_train_steps=500 ^
  --max_denoising_steps=40 ^
  --mixed_precision=bf16

Example: SDXL

accelerate launch sdxl_train_leco.py ^
  --pretrained_model_name_or_path="sdxl_model.safetensors" ^
  --output_dir="output" ^
  --output_name="detail_slider_xl" ^
  --prompts_file="slider.yaml" ^
  --network_dim=8 ^
  --network_alpha=4 ^
  --learning_rate=1e-4 ^
  --max_train_steps=500 ^
  --max_denoising_steps=40 ^
  --mixed_precision=bf16

Prompt YAML: original LECO format

- target: "van gogh"
  positive: "van gogh"
  unconditional: ""
  neutral: ""
  action: "erase"
  guidance_scale: 1.0
  resolution: 512
  batch_size: 1
  multiplier: 1.0
  weight: 1.0

Prompt YAML: ai-toolkit style slider target

This expands internally into the bidirectional LECO pairs needed for slider-style behavior.

targets:
  - target_class: ""
    positive: "high detail, intricate, high quality"
    negative: "blurry, low detail, low quality"
    multiplier: 1.0
    weight: 1.0

guidance_scale: 1.0
resolution: 512
neutral: ""

You can also provide multiple neutral prompts:

targets:
  - target_class: "person"
    positive: "smiling person"
    negative: "expressionless person"

neutrals:
  - ""
  - "studio photo"
  - "cinematic lighting"

2.1 KiB Raw Blame History