abide1622/ldc-workflow-krea2-lora

Krea 2 LoRA Trainer

A versioned training recipe for Krea 2 LoRAs. It covers the hyperparameters, the ostris/ai-toolkit config template, how a run is sized, and how the result is validated. It contains config only, no weights.

Files

File	Contents
job.json	Hyperparameters, step sizing, preview schedules per base, training adapters, the ai-toolkit config template, and the validation policy.
workflow.json	Version manifest: tag, release note, the run's inputs, and runtime (container image, entrypoint, minimum VRAM).
catalog/	Catalog card and thumbnail.

Versions

Each release is an immutable git tag (v1, v2, …). Never move a tag; any change to job.json or runtime is a new tag. main carries docs only, so pin a tag when you use the recipe.

git clone --branch v1 https://huggingface.co/abide1622/ldc-workflow-krea2-lora

Tag	Trainer	Notes
v1	ostris/aitoolkit:0.13.23	First release. Validated end to end on Raw (3,000 steps on an H100, about 1.34 s/step).

v1 settings

Network	LoRA, rank 16, alpha 16
Optimizer	AdamW 8-bit, lr 1e-4, flowmatch, linear timesteps, bf16
Resolution	1024
Caption dropout	0.05
Model	quantized transformer and text encoder, low_vram, gradient checkpointing
Saves / previews	about 6 per run, every max(50, ⌈steps/300⌉·50) steps
Default steps	50 per image, clamped to 1,000–3,000
GPU	≥ 48 GB VRAM

Base	Previews	Training adapter
krea2-raw	CFG 4.0, 25 steps	none
krea2-turbo	CFG 1.0, 9 steps	ostris/krea2_turbo_training_adapter

LoRAs trained on Raw transfer to Turbo.

Running a job

1. Dataset. Put images in one flat folder, each with an optional same-named .txt caption. Write the trigger word into each caption yourself. The config leaves ai-toolkit's trigger_word unset, so nothing adds it a second time.

2. Render the config. config_template in job.json is an ai-toolkit config with placeholders:

   • {"$job": "<name>"} is replaced by that value, and is required.
   • {"$job?": "<name>"} is removed when unset. An object left empty by removals is removed too.

   Placeholders	Value
   name, output_dir, dataset_dir, checkpoint_path	Job name and paths
   steps, checkpoint_every, sample_prompts	Per job
   arch, rank, learning_rate, optimizer, timestep_type, caption_dropout_rate, resolution	From job.json → arch and hyperparameters
   guidance_scale, sample_steps	From job.json → schedules[<base>]
   assistant_lora_path (optional)	From job.json → training_adapters[<base>]
   text_encoder_path, vae_path (optional)	Local copies; omit them to let ai-toolkit download its own

   Keep the template's key order when writing YAML.

3. Run it in the pinned image with the entrypoint from workflow.json. The entrypoint runs from /app/ai-toolkit, inside the image:

   docker run --gpus all -v "$PWD:/workspace" -w /app/ai-toolkit --entrypoint "" \
     ostris/aitoolkit:0.13.23 python run.py /workspace/job/config.yaml
   

   Wrap the run in a wall-clock limit (timeout). A run that is stopped still leaves its most recent checkpoint.

Validation

job.json → validation defines when a trained LoRA counts as passing:

• Structure: the file has LoRA modules at the configured rank, for the base it was trained on.
• Smoke test: one text-to-image generation on the training base, run twice at a fixed seed (424242), once without the LoRA and once with it.
  • Effect: the mean absolute pixel difference (0–255, at 256×256) must be ≥ 4.0, which proves the LoRA was applied.
  • Sanity: both images must have a standard deviation ≥ 8.0, which proves neither is a flat field.

A pass means the LoRA loads and changes the image without breaking it. It does not measure how well the concept was learned.

Contributing

Open a discussion or pull request against main. A change to the recipe ships as a new tag.