---
license: apache-2.0
language: en
pipeline_tag: text-to-image
tags:
  - text-to-image
  - diffusion
  - chroma
  - flux
  - gguf
  - quantized
  - on-device
  - ios
  - mobile
  - apple-silicon
base_model: lodestones/Chroma1-HD
---

# Chroma1-HD — iOS bundle

<p align="center">
  <a href="https://github.com/haplollc/Mirage"><img alt="Mirage" src="https://img.shields.io/badge/Runs%20on-Mirage-orange" /></a>
  <a href="https://huggingface.co/lodestones/Chroma1-HD"><img alt="Upstream" src="https://img.shields.io/badge/Upstream-lodestones%2FChroma1--HD-blue" /></a>
  <img alt="License" src="https://img.shields.io/badge/license-Apache--2.0-lightgrey" />
  <img alt="Params" src="https://img.shields.io/badge/params-8.9B-purple" />
</p>

A mobile-friendly bundle of **Chroma1-HD** (8.9B Flux-derived DiT) packaged with T5-XXL + Flux VAE, sized for on-device inference via [**Mirage**](https://github.com/haplollc/Mirage).

Chroma1-HD is the high-resolution flagship variant from [lodestones/Chroma1-HD](https://huggingface.co/lodestones/Chroma1-HD) — fully open weights built on FLUX.1's architecture with significant retraining for general-purpose generation. Excellent at photorealism, illustration, and concept work.

## What's inside

| File | Role | Size |
|---|---|---|
| [`Chroma1-HD-Q4_K_S.gguf`](./Chroma1-HD-Q4_K_S.gguf) | Diffusion transformer — Flux-arch, 8.9B params, Q4_K_S | 5.1 GB |
| [`t5xxl_fp16.safetensors`](./t5xxl_fp16.safetensors) | Text encoder (T5-XXL, fp16) | 9.1 GB |
| [`ae.safetensors`](./ae.safetensors) | VAE (from FLUX.1) | 320 MB |

Total bundle: **~14.5 GB**. Total GPU residency: ~15-16 GB. **iPhone 17 Pro / Air (12 GB) or M-series Mac required.** Older iPhones cannot hold this model.

## Quick start (Mirage)

```swift
import Mirage

let docs = FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0]

let engine = try Engine(models: ModelFiles(
    diffusionModel: docs.appendingPathComponent("Chroma1-HD-Q4_K_S.gguf"),
    vae:            docs.appendingPathComponent("ae.safetensors"),
    textEncoder:    docs.appendingPathComponent("t5xxl_fp16.safetensors")
))

let image = try await engine.generate(.init(
    prompt: "a Hopper-style oil painting of an empty diner at dawn, soft warm light",
    width: 1024, height: 1024,
    steps: 28,
    cfgScale: 4.0    // Chroma's recommended CFG
))
```

## Provenance

| Component | Upstream | License |
|---|---|---|
| Diffusion transformer | [lodestones/Chroma1-HD](https://huggingface.co/lodestones/Chroma1-HD) | Apache 2.0 |
| GGUF conversion | [silveroxides/Chroma1-HD-GGUF](https://huggingface.co/silveroxides/Chroma1-HD-GGUF) | Apache 2.0 |
| Text encoder | [comfyanonymous/flux_text_encoders](https://huggingface.co/comfyanonymous/flux_text_encoders) (`t5xxl_fp16.safetensors`) | Apache 2.0 |
| VAE | [ffxvs/vae-flux](https://huggingface.co/ffxvs/vae-flux) (re-host of FLUX.1's `ae.safetensors`) | FLUX-1-dev-non-commercial |

## Sizing

- **iPhone 17 Pro / Air (12 GB RAM)** — works but tight; expect ~8 min per 1024² image.
- **iPhone 16 Pro (8 GB RAM)** — does NOT fit. The text encoder alone is 9 GB. Use [`Z-Image-Turbo-iOS`](https://huggingface.co/jc-builds/Z-Image-Turbo-iOS) instead.
- **M2 / M3 Mac** — comfortable, ~4-6 min per 1024² image at 28 steps.

## Built by

[Haplo](https://haplo.app) · [Mirage on GitHub](https://github.com/haplollc/Mirage)