jivenchy/sd-turbo-coreml-384-6bit

sd-turbo — Core ML, 384×384, 6-bit (Apple Neural Engine)

Powered by Stability AI

A Core ML conversion of stabilityai/sd-turbo, compiled and quantized to run on-device on iPhone via the Apple Neural Engine.

Measured on a real device: ~2.2 s per image on an iPhone 13 mini (A15), no crashes.

⚠️ The Stability AI Materials in this repository have been changed. See NOTICE for the full list of modifications.

---

Why this exists

The published Core ML conversions of sd-turbo are compiled for Mac GPU (ORIGINAL attention), shipped uncompressed (fp16, ~2.5 GB), and locked to resolutions we could not use. Those three properties are baked in at conversion time and cannot be changed at runtime.

This conversion targets the opposite end: a phone. It trades away flexibility — one fixed resolution, no guidance, no img2img — for a bundle that fits in memory on a 4 GB iPhone and runs on the Neural Engine.

---

Conversion parameters

Every value below is compiled in and cannot be changed at runtime.

Parameter	Value	Why
Resolution	384 × 384 (latent 48 × 48)	512 renders subjects too small; 256 crops too tightly
Attention	SPLIT_EINSUM (v1)	Targets the Neural Engine. SPLIT_EINSUM_V2 requires the latent width to be a multiple of 512, so it cannot do 384
Quantization	6-bit palettized (text encoder + U-Net)	VAE decoder left at fp16 — it is small and quality-sensitive
U-Net batch size	1	sd-turbo does not use classifier-free guidance; batch 2 doubles the work for nothing (measured: 2.07× faster at batch 1)
VAE encoder	not included	Text-to-image only
Compute units	CPU_AND_NE

Produced with apple/ml-stable-diffusion:

python -m python_coreml_stable_diffusion.torch2coreml \
  --model-version stabilityai/sd-turbo \
  --convert-unet --convert-text-encoder --convert-vae-decoder \
  --attention-implementation SPLIT_EINSUM \
  --quantize-nbits 6 \
  --latent-h 48 --latent-w 48 \
  --unet-batch-one \
  --bundle-resources-for-swift-cli \
  --compute-unit CPU_AND_NE

Runtime settings

These are not compiled in, but the bundle only behaves correctly with them:

Setting	Value
Scheduler	LCM (see the caveat below)
Steps	4
Guidance scale	0.0
Compute units	cpuAndNeuralEngine
reduceMemory	true on iPhone

⚠️ Upstream apple/ml-stable-diffusion has no LCM scheduler

LCM support has been open and unmerged since 2024 (apple/ml-stable-diffusion#319). To use this model you need LCMScheduler.swift, which exists in keijiro's fork. Cherry-picking the single commit bd0cb9e onto current apple/main is enough — it is ~180 lines, and the three conflicts are each one extra case in a switch.

---

Contents

17 files, 1,072,200,142 bytes (~1.07 GB) total.

├── manifest.json                     ← integrity + version metadata
├── TextEncoder.mlmodelc/             304 MB
├── Unet.mlmodelc/                    620 MB
├── VAEDecoder.mlmodelc/               94 MB
├── vocab.json
└── merges.txt

.mlmodelc entries are directories, not files. They are stored here unarchived so that a client can stream them file-by-file and reconstruct the tree, rather than downloading a ~1 GB archive and needing ~2 GB of free space to unpack it.

manifest.json

{
  "schemaVersion": 1,
  "modelVersion": "1.0.0",
  "totalBytes": 1072200142,
  "files": [
    { "path": "Unet.mlmodelc/weights/weight.bin", "bytes": 650117120, "sha256": "..." }
  ]
}

Fetch this small file first to get the download size up front, verify each file with its SHA-256 as it lands, and detect when a newer modelVersion is published. Files are served at:

https://huggingface.co/jivenchy/sd-turbo-coreml-384-6bit/resolve/main/<path>

---

License

This model is licensed under the Stability AI Community License. Copyright © Stability AI Ltd. All Rights Reserved.

• Full text: LICENSE.md
• Attribution and the complete list of modifications: NOTICE

Your own use of this model is governed by that license. Two provisions are easy to miss:

• Commercial use requires registration with Stability AI at https://stability.ai/community-license.
• The license terminates if you or your affiliates generate more than USD $1,000,000 in annual revenue; an enterprise license is then required (https://stability.ai/enterprise).

Use is also subject to the Stability AI Acceptable Use Policy.

This is an unofficial community conversion. It is not affiliated with, endorsed by, or supported by Stability AI or Apple.