Lora Finetune

Dependency Source Review

This skill documents an optional compatibility dependency path for dataset validation, schema migration, and pipeline execution workflows. Treat scripts/_setup_deps.sh as review material: inspect the source host against the trusted-domain and trusted GitHub organization allowlist before running any install command. It is not required for ordinary pipeline, schema, or export requests.

LoRA Fine-Tuning (Apple Silicon)

Train custom LoRA adapters for Stable Diffusion 1.5 on Mac hardware. Tested on M4 24GB — produces 3.1MB weight files in ~15 minutes at 500 steps.

Hardware Requirements

Training Pipeline

1. Prepare dataset: 15-25 images in consistent style, 512×512, with text captions
2. Train LoRA: 500 steps, learning rate 1e-4, rank 4
3. Evaluate: Generate test images, compare base vs LoRA vs reference (Gemini/DALL-E)
4. Score: LLM-as-judge rates each on style consistency, quality, prompt adherence

Quick Start

# Prepare training images in a folder
ls training_data/
# image_001.png  image_001.txt  image_002.png  image_002.txt ...

# Train (see scripts/train_lora.py for full options)
python3 scripts/train_lora.py \
  --data_dir ./training_data \
  --output_dir ./lora_weights \
  --steps 500 \
  --lr 1e-4 \
  --rank 4

Evaluation with LLM-as-Judge

# Compare base model vs LoRA vs commercial (Gemini/DALL-E)
# Pixtral Large scores each image 1-10 on:
# - Style consistency with training data
# - Image quality and coherence
# - Prompt adherence

# Our results: Base 6.8 → LoRA 9.0 → Gemini 9.5
# Lesson: Gemini wins without training, but LoRA closes the gap significantly

Key Lessons

• float32 required on MPS — float16 silently produces NaN on Apple Silicon for SD pipelines
• mflux is faster than PyTorch MPS for FLUX (~105s vs ~90min) but doesn't support LoRA training
• SD 1.5 is the ceiling for 24GB — FLUX LoRA OOMs even with gradient checkpointing
• 15-25 images is the sweet spot — fewer undertrain, more doesn't help proportionally
• Gemini (Imagen 4.0) beats fine-tuned SD 1.5 with zero training — use commercial APIs for production, LoRA for experimentation and offline use

Files

• scripts/train_lora.py — Training script with Apple Silicon MPS support
• scripts/compare_models.py — LLM-as-judge evaluation comparing base vs LoRA vs reference