README.md

---
license: openrail++
base_model: ByteDance/SDXL-Lightning
pipeline_tag: text-to-image
tags:
  - core-ml
  - coreml
  - stable-diffusion
  - apple-silicon
  - ios
  - quantized
---

# ByteDance SDXL-Lightning — Core ML (8-bit)

![demo](demo.png)

<sub>**Generated on-device from this exact Core ML build** (4 steps, guidance 0, trailing timestep spacing, 1024x1024, seed 42 — 38 s on an M3 Ultra).</sub>

**Not a style fine-tune** — a few-step *distillation of SDXL base*, so it inherits base SDXL's look. Pick it for speed.

Core ML conversion for Apple silicon (iOS / iPadOS / macOS, Neural Engine), built with Apple's
[ml-stable-diffusion](https://github.com/apple/ml-stable-diffusion) for
[mindfire-image](https://github.com/Gatcha-man/mindfire-image).

## Original model

Converted from **[ByteDance/SDXL-Lightning](https://huggingface.co/ByteDance/SDXL-Lightning)** — go there for the
original weights, full model card and licence.

## Demo prompt

The prompt and settings used for this model's demo image (also the reference example shipped
in mindfire-image):

**Prompt**
```
A girl smiling
```

| Setting | Value |
|---|---|
| Steps | 4 |
| Guidance (CFG) | 0.0 |
| Size | 1024x1024 |

Guidance **must be 0** (CFG disabled) and the scheduler needs **trailing** timestep spacing. 2-8 steps.

## Modifications from the base model

Converted from PyTorch/diffusers to Core ML (`.mlmodelc`) and **quantized to 8-bit palettized**
weights. No fine-tuning — behaviour tracks the base model, though quantization can shift outputs
slightly.

## Licence

Inherited from the base model: `openrail++`. This carries the OpenRAIL **use-based restrictions
(Attachment A)**, which bind you as a downstream user of this conversion exactly as they do for the
base model. Read the base model's licence before use or redistribution.