| --- |
| license: mit |
| base_model: HiDream-ai/HiDream-O1-Image-Dev |
| tags: |
| - mlx |
| - mlx-vlm |
| - hidream |
| - text-to-image |
| - apple-silicon |
| - quantized |
| - q6 |
| language: |
| - en |
| pipeline_tag: text-to-image |
| library_name: mlx |
| inference: false |
| authors: |
| - Mrbizarro |
| --- |
| |
| # HiDream-O1-Image-Dev — MLX Q6 (Apple Silicon) |
|
|
| > Ported by **[Mrbizarro](https://huggingface.co/Mrbizarro)** · MIT licensed · published to mlx-community |
|
|
| ## 🎛️ Run it one-click in **[Phosphene](https://github.com/mrbizarro/phosphene)** |
|
|
| Phosphene is a free local generative-video panel for Apple Silicon. HiDream is wired into its Image Studio. **[Install Pinokio](https://pinokio.computer)**, then in Pinokio install [Phosphene](https://github.com/mrbizarro/phosphene). Note: Phosphene's HiDream integration uses BF16 by default since **edit requires BF16** — this Q6 repo is for text-to-image-only workflows on RAM-constrained machines. |
|
|
| --- |
|
|
| A 6-bit quantized MLX port of [HiDream-ai/HiDream-O1-Image-Dev](https://huggingface.co/HiDream-ai/HiDream-O1-Image-Dev). |
|
|
| **⚠ Q6 does NOT support edit / multi-ref.** Per-group dequantization noise compounds against reference-image features in attention and produces degenerate output. For edit / multi-reference workflows use the [BF16 sibling repo](https://huggingface.co/mlx-community/HiDream-O1-Image-Dev-mlx-bf16) instead. |
|
|
| ## Sibling repos |
|
|
| - 🟢 **[BF16 (full precision)](https://huggingface.co/mlx-community/HiDream-O1-Image-Dev-mlx-bf16)** — 17.5 GB, ~16 GB RAM, clean across all dimensions. **Use this when in doubt.** |
| - 🟡 **Q6 (this repo)** — 8 GB, ~8.5 GB RAM, fast. **Best at square 2048×2048 or 1024×1024.** Visible 32-pixel patch grid in flat regions at non-square dims. |
| - 🟡 [Q8](https://huggingface.co/mlx-community/HiDream-O1-Image-Dev-mlx-q8) — 10 GB, ~11.5 GB RAM, same artifact behaviour as Q6 at non-square dims. |
|
|
| ## When to use Q6 |
|
|
| - ✅ **Square 1024×1024 or 2048×2048** — clean output, half the time of BF16 |
| - ✅ **RAM-constrained** — fits 16 GB Macs alongside other apps |
| - ❌ **Non-square dims (1440×2560, 3104×1312, etc)** — visible 32-pixel patch grid in skies, walls, water → use BF16 |
|
|
| ## What's in this repo |
|
|
| - `model.safetensors` — Q6 quantized backbone (8 GB) |
| - `extras/custom_heads.safetensors` — diffusion-side heads (75 MB, BF16) |
| - `config.json` (with `quantization: {bits: 6, group_size: 64}` so mlx-vlm wraps `Linear → QuantizedLinear` correctly) |
| - Tokenizer + processor configs |
|
|
| ## Code |
|
|
| The inference scripts are **not** in this repo — they live in the [BF16 sibling repo](https://huggingface.co/mlx-community/HiDream-O1-Image-Dev-mlx-bf16) under `scripts/hidream_o1/`. Clone that for code, this for weights only. |
|
|
| ## Quick start |
|
|
| ```bash |
| # Get the code |
| hf download mlx-community/HiDream-O1-Image-Dev-mlx-bf16 --local-dir hidream-o1-mlx \ |
| --include "scripts/*" --include "*.md" --include "*.txt" --include "*.gitattributes" |
| cd hidream-o1-mlx |
| uv venv --python 3.11 && uv pip install -r requirements.txt |
| |
| # Get the Q6 weights |
| hf download mlx-community/HiDream-O1-Image-Dev-mlx-q6 --local-dir mlx_models/hidream-o1-dev-q6 |
| |
| # Run (square dims only for clean output) |
| .venv/bin/python scripts/hidream_o1/generate_hidream_o1_mlx.py \ |
| --model-path mlx_models/hidream-o1-dev-q6 \ |
| --prompt "your prompt here" \ |
| --width 2048 --height 2048 \ |
| --output out.png |
| ``` |
|
|
| ## Performance |
|
|
| | Resolution | Per step | Total (28 steps) | Peak RAM | Quality | |
| |---|---|---|---|---| |
| | 1024×1024 | 1.30 s | 36 s | 8.5 GB | ✅ clean | |
| | 2048×2048 | 5.51 s | 154 s | 9 GB | ✅ clean | |
| | 1440×2560 (non-square) | 4.50 s | 127 s | 8.5 GB | ⚠ patch grid visible | |
|
|
| ## License |
|
|
| MIT — see the [BF16 repo](https://huggingface.co/mlx-community/HiDream-O1-Image-Dev-mlx-bf16) for the full LICENSE file and acknowledgements. |
|
|
