lora-training
lora-training
Train with ostris/ai-toolkit (MIT, https://github.com/ostris/ai-toolkit). It is an external install that the user does on the training machine. This skill never installs or runs it for them. The design follows agentY's lora-training skill (MIT, copied to references/agentY-lora-training.md; its tool names are agentY-specific). Use open-weights base models only.
0. Should this be a LoRA?
Only when an identity, style, product or motion has to come back reliably across many prompts. For a single look, use a reference image with image-edit (Qwen edit) or fix.edit instead. It is cheaper and needs no training.
1. Build the dataset (scripts/dataset_check.py)
python scripts/dataset_check.py <folder> --kind character|product|style|motion [--trigger tok]. It is stdlib only and reads nothing over the network. It checks:
- the item count against the targets (character 15–40, product 15–30, style 25–60, motion 10–30 clips of 2–5 s)
- for each item, a matching
<name>.txtcaption - exact duplicates (by sha256) and empty captions
- tiny images (PNG/JPEG headers are read for the size)
- whether a caption contains the trigger word, which it should not, because the trainer prepends it
It writes dataset.json (items, warnings and a sha256 of the whole set) and prints the warnings.
Captions describe what should stay changeable: setting, pose, clothes, light, camera, motion. They do not describe what the LoRA must learn (the face, the product's shape, the style's look). Write natural sentences for FLUX.2, Qwen and Z-Image. For video, write subject, action and camera move in one paragraph.