sheigl/craftroom-ai-payload

Craft Room AI — compiled CoreML device payload

Pre-compiled CoreML (.mlmodelc) artifacts of the Craft Room AI coloring-page pipeline: FLUX.2-klein-base-4B merged with a custom image-to-image coloring LoRA (v2.1, rank 16, 1500 steps), quantized and compiled for on-device inference on Apple silicon (CPU/GPU/ANE via the neural_network MLProgram format).

This repo is the first-launch download for the Craft Room AI iOS test harness. It is not a transformers/diffusers checkpoint set — it is the exact runtime payload an iPad fetches and runs unmodified.

Why the payload lives here

Apple's ITMS-90558 limit caps a thinned app bundle below 4 GB; the int8 DiT alone is 3.88 GB of weight.bin (old bundled .app: 4,078,949,396 B). The ipa ships only the harness + small baked bins (~50 MB); the models come from this repo over HTTPS (public, Range-resumable CDN — no token needed on device).

Contents (device-v1/)

Path	Size	What
dit_int8.mlmodelc/weights/weight.bin	3,698 MiB	DiT, per-channel int8 weights, fp16 activations
vae_decoder.mlmodelc/…	~95 MiB	VAE decoder, fp16 (static 48×48 latent input)
vae_encoder_land_1248x832.mlmodelc/…	~66 MiB	VAE encoder, fp16, static 1248×832 landscape input
manifest.json	small	every file: path, size, sha256, url

The app verifies every file against manifest.json (size before install, sha256 once per freshly downloaded file). Re-publishing a payload version means regenerating the manifest with scripts/publish_device_payload.py in the harness repo — do not hand-edit.

Provenance

• Base: black-forest-labs/FLUX.2-klein-base-4B (Apache-2.0).
• Adapter: custom coloring LoRA v2.1 (rank 16, 1500 steps), merged into the FP16 checkpoint before export (the "Converter v2.1" config approved in the project's evaluation rounds).
• Export: diffusers Flux2KleinPipeline edit path → Core ML via coretool/coremlbuild compile-on-target on a GitHub Actions macos-latest runner (iOS 17 target, arm64). weight.bin is therefore device-compiled and only loadable by Core ML, not re-loadable as a portable checkpoint.
• The companion bake (txt embeds, BN stats, sigmas, position ids) ships as flat little-endian bins inside the ipa, not here.

License / attribution

Apache-2.0, inherited from FLUX.2-klein-base-4B (Black Forest Labs). The quantization, compilation and LoRA are the project's contribution under the same license.