mahmudplx/coreml-absolutereality-6bit

AbsoluteReality — Core ML (6-bit UNet)

Core ML conversion of Lykon/AbsoluteReality for on-device generation with apple/ml-stable-diffusion.

Everything sits under Resources/, the layout StableDiffusionPipeline(resourcesAt:) expects:

File	Precision	Size
Unet.mlmodelc	6-bit palettized	618 MB
TextEncoder.mlmodelc	fp16	235 MB
VAEDecoder.mlmodelc	fp16	95 MB
VAEEncoder.mlmodelc	fp16	65 MB
vocab.json, merges.txt	—	1.4 MB

1.01 GB total, 512×512, SPLIT_EINSUM_V2 attention — built for the Neural Engine, so pair it with MLComputeUnits.cpuAndNeuralEngine. The VAE encoder is included, so image-to-image works.

Only the UNet is palettized: the text encoder holds CLIP's -inf causal mask, which the k-means palettizer cannot cluster. The UNet is ~80% of the weight, so the saving is nearly the same.

Converted with:

python -m python_coreml_stable_diffusion.torch2coreml \
  --model-version <AbsoluteReality diffusers fp16> \
  --convert-unet --convert-text-encoder --convert-vae-decoder --convert-vae-encoder \
  --attention-implementation SPLIT_EINSUM_V2 \
  --latent-h 64 --latent-w 64 \
  --bundle-resources-for-swift-cli -o .
# UNet palettized separately to 6 bits

Licence

CreativeML Open RAIL-M, inherited from the source checkpoint. The licence's use restrictions travel with these weights: if you redistribute them or build on them, pass the same restrictions on to your users.