End-to-end checklist

Why this checklist is what you deliver

If you put the model on a phone, you do not deliver a notebook. You deliver a baked UNet (LoRA merged in), the VAE and text-encoder parts the graph needs, the notes that can rebuild that folder, and a device measurement. If any box below is missing, the work is still a demo, not a finished delivery.

The sibling course, LLM LoRA for your writing style, uses this same bake-then-export rule on text. Complete both if you want both methods.

The steps, one box at a time

Environment setup

Replay the Mac environment before you say you are done. Use Python 3.11+, uv, and MPS or CPU.

curl -LsSf https://astral.sh/uv/install.sh | sh
mkdir -p diffusion-lora-coreai && cd diffusion-lora-coreai
uv init --python 3.11
uv venv
uv add torch torchvision diffusers peft transformers accelerate datasets safetensors coreai-torch
uv run python -c "import torch; print(torch.__version__); print('MPS:', torch.backends.mps.is_available())"
uv run coreai.diffusion.export --help
xcodebuild -version

Before training

• Choose sd-1.5 or sd-2.1 and record the exact revision and license.
• Define subject versus style and write one success sentence.
• Keep local, licensed data, captions, a holdout split, and a privacy plan.
• State in the contract that Core AI has no documented hot-swap LoRA.

During training

• Use MPS if supported, otherwise CPU. Do not add CUDA.
• Pin packages. Save checkpoints, previews, training settings, and device metadata.
• Compare base and personalized behavior on held-out prompts with a fixed seed.

At merge, export, and device

• Merge or fuse the LoRA into the UNet or the recipe's target component. Keep the adapter.
• Export with installed coreai.diffusion.export, or with torch.export plus coreai-torch.
• Inspect the multi-component folder, checksums, tokenizer, and metadata.
• Load the merged resource in Xcode 27. Prepare once. Measure a real generation.
• Complete lesson 07: golden-set grids for pre-merge, merged, and device, open codes, four buckets, and yes-or-no checks.

Failure modes

• Calling the adapter the product. The product is the baked resource plus the notes that rebuild it.
• Missing holdout or missing device time. Either gap makes the claim untested.
• Describing a runtime LoRA swap. That is not a documented Core AI path.
• Losing the adapter after merge. You cannot rebuild without it.

Done when

• Contract, adapter, merged pipeline, export folder, device measurement, and lesson 07 eval artifacts all exist.
• The device run used the baked resource only.
• A new engineer can rebuild from the manifest without guessing flags.
• You can point to the sibling writing-style course as the same method for writing.

Keep learning

PyTorch to Core AI in Xcode · From the metal to the model · Core AI models, typed · Model architectures in plain English · LLM LoRA for your writing style