> For the complete documentation index, see [llms.txt](https://unsloth.ai/docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://unsloth.ai/docs/models/minimax-h3-image.md).

# MiniMax-H3-Image - How to run locally

Run MiniMax-H3-Image locally with Unsloth FP8 and GGUF quants for text-to-image and image editing in 6 steps.

**MiniMax-H3-Image** is MiniMax's open 33B model for text-to-image and editing with up to 16 reference pictures. It likes long, complete prompts. It is distilled to **6 steps, no CFG**, on a **2048 × 2048** canvas: about **5 s** per image with the FP8 repo and **18 s** with UD-Q4\_K\_XL on a 40 GB A100.

{% columns %}
{% column width="50%" %}
The quants use [Unsloth Dynamic](https://unsloth.ai/docs/basics/dynamic-3.0-ggufs), which keeps important layers in higher precision. Run them in [Unsloth](https://unsloth.ai/docs/desktop.md), stable-diffusion.cpp and more.

* MiniMax-H3-Image: [GGUF](https://huggingface.co/unsloth/MiniMax-H3-Image-GGUF) • [FP8](https://huggingface.co/unsloth/MiniMax-H3-Image-FP8)
* Original model: [MiniMaxAI/MiniMax-H3-Image](https://huggingface.co/MiniMaxAI/MiniMax-H3-Image)

Tested with a development build of Unsloth (newer than 2026.11.4) on NVIDIA A100 (40 GB) and RTX PRO 6000 (96 GB).
{% endcolumn %}

{% column width="50%" %}

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FN17t739HDhs4Bs9VVAEL%2Fminimax-h3-image-out-bookshop.webp?alt=media" alt="Rainy bookshop window with an UNSLOTH BOOKS chalkboard and a sleeping cat"><figcaption><p>FP8 repo (INT8), 2048 × 2048, about 5 s on an A100</p></figcaption></figure>
{% endcolumn %}
{% endcolumns %}

<div><figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FdhaLehHOWudDk9TmHJfW%2Fminimax-h3-image-out-valley.webp?alt=media" alt="Misty autumn valley with a wooden footbridge"><figcaption><p>FP8 repo (INT8)</p></figcaption></figure> <figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FE5iZZZi5oMaVw4Jo3ViS%2Fminimax-h3-image-out-island.webp?alt=media" alt="Isometric floating island city"><figcaption><p>FP8 repo (INT8)</p></figcaption></figure> <figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FQ8V3wLPL9YacqR4YD5LK%2Fminimax-h3-image-out-ramen.webp?alt=media" alt="Overhead photo of ramen"><figcaption><p>GGUF UD-Q4_K_XL</p></figcaption></figure></div>

### Memory requirements <a href="#memory-requirements" id="memory-requirements"></a>

With a 32B Qwen3-VL text encoder (from MiniMax-H3), it needs a large GPU or plenty of system RAM.

| Recipe                         | Image model             | Text encoder                   | VAE    | Total download  |
| ------------------------------ | ----------------------- | ------------------------------ | ------ | --------------- |
| **FP8 repo** (Unsloth default) | INT8-ConvRot, 20.3 GB   | INT8 Qwen3-VL 32B, 27.1 GB     | 2.3 GB | **about 50 GB** |
| **GGUF UD-Q4\_K\_XL**          | 13.6 GB                 | Q4\_K\_M Qwen3-VL 32B, 18.2 GB | 2.3 GB | **about 34 GB** |
| **GGUF Q8\_0**                 | 22.0 GB                 | Q4\_K\_M Qwen3-VL 32B, 18.2 GB | 2.3 GB | **about 43 GB** |
| Prompt enhancer (optional)     | Qwen3.5-9B UD-Q4\_K\_XL |                                |        | **6.9 GB**      |

| Your hardware                          | Start with                                                              |
| -------------------------------------- | ----------------------------------------------------------------------- |
| **NVIDIA GPU with 80 GB or more**      | FP8 repo, fully on the GPU. Tested on a 96 GB RTX PRO 6000: 54 GB peak. |
| **NVIDIA GPU with 40 GB**              | FP8 repo. Tested on a 40 GB A100 with 83 GB RAM: parts stream from RAM. |
| **NVIDIA GPU with 24 GB**              | GGUF **UD-Q4\_K\_XL**, or the FP8 repo with 48 GB+ system RAM           |
| **GPU with 16 GB**                     | GGUF **UD-Q4\_K\_XL** or smaller, with 48 GB+ system RAM                |
| **Mac or CPU only, 64 GB RAM or more** | GGUF **UD-Q4\_K\_XL** (slow)                                            |

Only the 96 GB and 40 GB rows were tested; the rest are estimates. On 24 GB, UD-Q4\_K\_XL beats Q8\_0 (which spills to RAM). Macs and CPUs get the GGUF only.

#### FP8, INT8 and GGUF <a href="#fp8-int8-and-gguf" id="fp8-int8-and-gguf"></a>

The FP8 repo holds two 20.3 GB files, INT8-ConvRot and FP8. **Unsloth loads INT8-ConvRot by default** (with the INT8 text encoder) because it is closer to bf16 (LPIPS 0.140 vs 0.255 for FP8 at 2048 × 2048). The picker still calls it **Fast FP8**. It is about 3x faster than the GGUF on a GPU with enough memory.

GGUF quants: **Q8\_0** 22.03 GB (closest to bf16), **UD-Q6\_K\_XL** 17.58 GB, **UD-Q4\_K\_XL** 13.61 GB (**recommended**), **UD-Q2\_K\_XL** 8.68 GB (tight memory).

#### Hardware notes <a href="#hardware-notes" id="hardware-notes"></a>

These need Unsloth 2026.11.5 or newer.

* **T4 and other GPUs without bf16:** the FP8 repo (INT8 image model and text encoder) renders and edits in fp16. Tested on a 15 GB Colab T4 with 51 GB of RAM: blocks stream from system RAM (about 43 to 47 GB used), up to about 14 GB of GPU memory.

  | Task, 6 steps              | Time per image |
  | -------------------------- | -------------- |
  | Text to image, 1024 x 1024 | about 28 s     |
  | Text to image, 2048 x 2048 | about 83 s     |
  | Edit, 1 reference, 1024    | 2.5 to 3 min   |

  LoRAs load and render, but the load streams the full 66 GB model (about 12 min on the T4). **Sequential** placement needs about 52.4 GB of free system RAM; with less, Unsloth refuses and shows both figures.
* **AMD on Linux** (Radeon 8060S, gfx1151): LoRA training with a **bf16** base, reloading, rendering with the LoRA and export all work.

### Quickstart <a href="#quickstart" id="quickstart"></a>

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FUySaKhMrEM4BEdibJvbp%2Fminimax-h3-image-flow.webp?alt=media" alt="Animation: pick MiniMax-H3-Image (Fast FP8), type a prompt, click Generate and get an alpine valley at sunrise" width="375"><figcaption><p>The Unsloth flow, sped up</p></figcaption></figure>

{% stepper %}
{% step %}

#### Install Unsloth <a href="#install-unsloth" id="install-unsloth"></a>

Get the [Unsloth Desktop app](https://unsloth.ai/docs/desktop.md) for [macOS](https://unsloth.ai/docs/get-started/install/mac.md), [Windows](https://unsloth.ai/docs/get-started/install/windows-installation.md) or [Linux](https://unsloth.ai/docs/get-started/install/linux.md).

<a href="https://unsloth.ai/download" class="button primary" data-icon="down-to-bracket">Download Unsloth</a>

Or from the terminal. MacOS, Linux, WSL:

{% code collapsedlinecount="10" %}

```bash
curl -fsSL https://unsloth.ai/install.sh | sh
```

{% endcode %}

Windows PowerShell:

{% code collapsedlinecount="10" %}

```powershell
irm https://unsloth.ai/install.ps1 | iex
```

{% endcode %}

Then launch Unsloth and open `http://localhost:8888`:

{% code collapsedlinecount="10" %}

```bash
unsloth studio -H 0.0.0.0 -p 8888
```

{% endcode %}
{% endstep %}

{% step %}

#### Open Images <a href="#open-images" id="open-images"></a>

Select **Images** in the sidebar. **Create** opens by default.

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FlQtHN8moCQiBVRksvUtk%2Fminimax-h3-image-images-page.png?alt=media" alt="The Unsloth Images page with the Create workflow open and no model loaded"><figcaption><p>The Images page</p></figcaption></figure>
{% endstep %}

{% step %}

#### Pick MiniMax-H3-Image <a href="#pick-minimax-h3-image" id="pick-minimax-h3-image"></a>

Click **Select image model**, search `MiniMax-H3-Image` and pick:

* **MiniMax-H3-Image (Fast FP8)**: the FP8 repo. NVIDIA GPUs with 24 GB or more.
* **MiniMax-H3-Image-GGUF (Slow)**: pick a quant. **UD-Q4\_K\_XL** is recommended.

<div><figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FHhCLEoV6z1L4Db8ily2v%2Fminimax-h3-image-picker-search.png?alt=media" alt="Model picker search results for MiniMax-H3-Image with the Fast FP8 and GGUF entries"><figcaption><p>Search results</p></figcaption></figure> <figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2F3ZuWCPvHggmFB2WbdphS%2Fminimax-h3-image-gguf-quants.png?alt=media" alt="The GGUF entry expanded to show its quants, with UD-Q4_K_XL marked recommended"><figcaption><p>GGUF quants</p></figcaption></figure></div>

The first load downloads the model, text encoder and VAE.

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FqfYd6Ju8MCpA2vWphkeV%2Fminimax-h3-image-download.png?alt=media" alt="Download panel with the 27 GB INT8 text encoder" width="400"><figcaption><p>First download of the FP8 repo</p></figcaption></figure>
{% endstep %}

{% step %}

#### Write your prompt <a href="#write-your-prompt" id="write-your-prompt"></a>

Describe the whole image. Unsloth sets **6 steps** (locked: the model is distilled for exactly 6), **guidance 1** and **2048 × 2048**.

**Enhance prompt** (on by default, MiniMax's reference workflow) has a Qwen3.5-9B model rewrite your prompt. It is a separate **6.9 GB** download and slow: about **3.5 minutes** extra per image on an A100. Turn it off for quick tries.

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FBJhHYOSYImp8McKcvGjM%2Fminimax-h3-image-enhancer.png?alt=media" alt="Enhance prompt switch with its 6.9 GB download button" width="350"><figcaption><p>Enhance prompt before its download</p></figcaption></figure>
{% endstep %}

{% step %}

#### Generate <a href="#generate" id="generate"></a>

Click **Generate**. **Recipe** shows every setting, the precision and the enhanced prompt.

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2Fmqa9vZ88KIaOFN27Gl5n%2Fminimax-h3-image-result-lossless.webp?alt=media" alt="Ginger cat asleep on books in a rainy bookshop window"><figcaption><p>FP8 repo (INT8), short prompt, enhancer on, seed 42</p></figcaption></figure>

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FwwUefNK5OGmbIuwl2tfE%2Fminimax-h3-image-recipe-lossless.webp?alt=media" alt="Recipe card with the enhanced prompt, INT8 precision, 6 steps and seed 42"><figcaption><p>Recipe card</p></figcaption></figure>
{% endstep %}
{% endstepper %}

### Image editing <a href="#image-editing" id="image-editing"></a>

Select **Edit**, upload **Picture 1**, and **Add picture** for more (up to 16). Name them as `<Picture 1>`, `<Picture 2>` and say what to keep and change. Output follows Picture 1's shape on the 2048 canvas; these examples use **Output size** **Custom** `1024 × 1024`, seed 7, enhancer off: 3 to 4 s each on an RTX PRO 6000 (FP8 repo). With the enhancer, an edit took about 9 minutes on an A100.

{% hint style="info" %}
Do not edit a picture with the same seed and size that generated it. The edit then starts from that picture's own noise and comes out over-sharpened, ignoring the instruction. Change the seed (or leave it random).
{% endhint %}

**One picture:** "Edit \<Picture 1>. Turn the scene into a snowy winter morning. Keep the bridge, the river and the mountains the same."

<div><figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FdYqwblQlu0tXyHWZR02d%2Fminimax-h3-image-edit1.png?alt=media" alt="Edit with the autumn valley as Picture 1 and the snowy result"><figcaption><p>One picture</p></figcaption></figure> <figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FCG9GfRXXtmBhCMLt7ko6%2Fminimax-h3-image-edit-winter.webp?alt=media" alt="The same valley, bridge and river on a snowy winter morning"><figcaption><p>Result</p></figcaption></figure></div>

**Two pictures:** "Put the hot-air balloon from \<Picture 2> in the sky above the valley in \<Picture 1>. Keep the bridge, the trees, the river and the light the same."

<div><figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FJ0iVCIJ1aDg9ibKSM9nY%2Fminimax-h3-image-edit2.png?alt=media" alt="Edit with the valley and a balloon as Pictures 1 and 2"><figcaption><p>Two pictures</p></figcaption></figure> <figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FnyY7SUigpABhfJK6q40y%2Fminimax-h3-image-edit-balloon.webp?alt=media" alt="The valley from Picture 1 with the hot-air balloon from Picture 2 in the sky"><figcaption><p>Result</p></figcaption></figure></div>

### Train a LoRA <a href="#train-a-lora" id="train-a-lora"></a>

See [Fine-tune MiniMax-H3-Image](https://unsloth.ai/docs/models/minimax-h3-image/fine-tune-minimax-h3-image.md)

#### Before and after <a href="#before-and-after" id="before-and-after"></a>

Seed 42, 1024 × 1024, enhancer off: `a trtcrd of a cat sitting on a throne holding a golden cup, "queen of cats"` and `a trtcrd of an astronaut standing on the moon holding a lantern, "the explorer"`.

<div><figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2Fkg53StPeOhBWosYeBEkn%2Fminimax-h3-image-lora-cat-before.webp?alt=media" alt="Without the LoRA: photo-like cat on a throne"><figcaption><p>No LoRA</p></figcaption></figure> <figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2F5idj4QIy1DMxo4QWxs88%2Fminimax-h3-image-lora-cat-w06.webp?alt=media" alt="LoRA weight 0.6: illustrated cat"><figcaption><p>Weight 0.6</p></figcaption></figure> <figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FoD4Irr8nLFeWdVOvHUMT%2Fminimax-h3-image-lora-cat-after.webp?alt=media" alt="LoRA weight 1.0: tarot card style cat"><figcaption><p>Weight 1.0</p></figcaption></figure></div>

<div><figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FjEaC0aZLy24V1JEzLzdH%2Fminimax-h3-image-lora-astronaut-before.webp?alt=media" alt="Without the LoRA: astronaut photo"><figcaption><p>No LoRA</p></figcaption></figure> <figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2F7KTnboVty4ipiJY5X4OH%2Fminimax-h3-image-lora-astronaut-after.webp?alt=media" alt="LoRA weight 1.0: tarot card style astronaut"><figcaption><p>Weight 1.0</p></figcaption></figure></div>

* Weight **1.0** gives the full style; **0.6 to 0.8** is subtler.
* The LoRA slightly changes prompts without the trigger word too. Lower the weight or load without it for other work.
* The style is strongest at 1024 × 1024 (the training size) and weaker at 2048 × 2048.

[Continue in the fine-tuning guide](https://unsloth.ai/docs/models/minimax-h3-image/fine-tune-minimax-h3-image.md#before-and-after)

### Run from code <a href="#run-from-code" id="run-from-code"></a>

#### Unsloth API <a href="#unsloth-api" id="unsloth-api"></a>

Load the model on the Images page, then call the OpenAI-compatible `POST /v1/images/generations` endpoint (to load by `model` instead, turn on **Switch image and video model by request** in **Settings → API**). Create an API key there ([Unsloth API](https://unsloth.ai/docs/basics/api.md)); the examples read `UNSLOTH_API_KEY`.

With `openai` (`pip install openai`):

{% code collapsedlinecount="10" %}

```python
import base64
import os

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8888/v1", api_key=os.environ["UNSLOTH_API_KEY"])

result = client.images.generate(
    model="unsloth/MiniMax-H3-Image-FP8",
    prompt="A misty mountain valley at sunrise with a wooden footbridge over a calm river, autumn trees in red, orange and gold on both banks, and soft golden light breaking through the mist. Detailed landscape photograph.",
    size="2048x2048",
    response_format="b64_json",
    extra_body={"steps": 6, "guidance": 1.0, "seed": 42},
)

with open("valley.png", "wb") as f:
    f.write(base64.b64decode(result.data[0].b64_json))
```

{% endcode %}

With `curl`:

{% code collapsedlinecount="10" %}

```bash
curl -s http://localhost:8888/v1/images/generations \
  -H "Authorization: Bearer $UNSLOTH_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "unsloth/MiniMax-H3-Image-FP8",
    "prompt": "Isometric 3D illustration of a tiny floating island city with pastel houses and terracotta roofs, a waterfall pouring off the edge into soft clouds, and small hot-air balloons drifting around it. Clean lines, soft shadows, calm pastel palette.",
    "size": "2048x2048",
    "response_format": "b64_json",
    "seed": 7
  }' \
  | python3 -c "import sys, json, base64; open('island.png', 'wb').write(base64.b64decode(json.load(sys.stdin)['data'][0]['b64_json']))"
```

{% endcode %}

`size` is `WIDTHxHEIGHT` in multiples of 16; `steps` and `guidance` default to 6 and 1. The API does not run the enhancer. About 3 s per image on an RTX PRO 6000, 5 s on an A100.

**LoRAs over the API** (Unsloth 2026.11.5 or newer): requests use the LoRA merged in at load, for example after **Deploy to Create**. A request with a different `loras` list is refused; reload to change it.

#### GGUF with stable-diffusion.cpp <a href="#gguf-with-stable-diffusion.cpp" id="gguf-with-stable-diffusion.cpp"></a>

Run the GGUF with `sd-cli` from [stable-diffusion.cpp](https://github.com/unslothai/stable-diffusion.cpp) (Unsloth's pinned release).

**1. Get `sd-cli`** (Linux, CUDA 13). The [release page](https://github.com/unslothai/stable-diffusion.cpp/releases/tag/master-982-0518b34-u27b9ddc) also has CUDA 12, Vulkan, CPU, macOS and Windows builds.

{% code collapsedlinecount="10" %}

```bash
curl -L -o sd.zip https://github.com/unslothai/stable-diffusion.cpp/releases/download/master-982-0518b34-u27b9ddc/sd-master-982-0518b34-u27b9ddc-bin-Linux-Ubuntu-22.04-x86_64-cuda13.zip
unzip -o sd.zip
mv sd-master-982-0518b34-u27b9ddc-bin-Linux-Ubuntu-22.04-x86_64-cuda13 sd
./sd/sd-cli --version
```

{% endcode %}

**2. Download the model, text encoder and VAE** (about 34 GB):

{% code collapsedlinecount="10" %}

```bash
pip install -U huggingface_hub
hf download unsloth/MiniMax-H3-Image-GGUF minimax_h3_image_pruned-UD-Q4_K_XL.gguf --local-dir models
hf download unsloth/MiniMax-H3-GGUF qwen3vl_32b_minimax_h3-Q4_K_M.gguf --local-dir models
hf download MiniMaxAI/MiniMax-H3-Image vae/diffusion_pytorch_model.safetensors --local-dir models
```

{% endcode %}

**3. Create an image.** `--offload-to-cpu` keeps weights in RAM and moves each part to the GPU when needed (peak 19.4 GB VRAM, 39 GB RAM).

{% code collapsedlinecount="10" %}

```bash
./sd/sd-cli \
  --diffusion-model models/minimax_h3_image_pruned-UD-Q4_K_XL.gguf \
  --llm models/qwen3vl_32b_minimax_h3-Q4_K_M.gguf \
  --vae models/vae/diffusion_pytorch_model.safetensors \
  -p "A misty mountain valley at sunrise with a wooden footbridge over a calm river, autumn trees in red, orange and gold on both banks, and soft golden light breaking through the mist. Detailed landscape photograph." \
  --steps 6 --cfg-scale 1.0 --sampling-method euler --sigmas minimax-h3-image --rng cpu \
  --offload-to-cpu \
  -W 2048 -H 2048 --seed 42 \
  -o valley-sdcpp.png
```

{% endcode %}

103 s on a 40 GB A100, 12 s on an RTX PRO 6000.

**4. Edit an image.** Each `-r` is the next `<Picture N>`; `sd-cli` crops them to `-W` × `-H`, so match Picture 1's shape. No vision file needed. 16 s and 20 s on an RTX PRO 6000.

{% code collapsedlinecount="10" %}

```bash
./sd/sd-cli \
  --diffusion-model models/minimax_h3_image_pruned-UD-Q4_K_XL.gguf \
  --llm models/qwen3vl_32b_minimax_h3-Q4_K_M.gguf \
  --vae models/vae/diffusion_pytorch_model.safetensors \
  -r valley-sdcpp.png \
  -p "Edit <Picture 1>. Turn the scene into a snowy winter morning. Keep the bridge, the river and the mountains the same." \
  --steps 6 --cfg-scale 1.0 --sampling-method euler --sigmas minimax-h3-image --rng cpu \
  --offload-to-cpu \
  -W 1024 -H 1024 --seed 42 \
  -o valley-sdcpp-winter.png
```

{% endcode %}

Two pictures (`balloon.png` is a balloon photo):

{% code collapsedlinecount="10" %}

```bash
./sd/sd-cli \
  --diffusion-model models/minimax_h3_image_pruned-UD-Q4_K_XL.gguf \
  --llm models/qwen3vl_32b_minimax_h3-Q4_K_M.gguf \
  --vae models/vae/diffusion_pytorch_model.safetensors \
  -r valley-sdcpp.png -r balloon.png \
  -p "Put the hot-air balloon from <Picture 2> in the sky above the valley in <Picture 1>. Keep the bridge, the trees, the river and the light the same." \
  --steps 6 --cfg-scale 1.0 --sampling-method euler --sigmas minimax-h3-image --rng cpu \
  --offload-to-cpu \
  -W 1024 -H 1024 --seed 42 \
  -o valley-sdcpp-balloon.png
```

{% endcode %}

On a Mac or CPU, use the GGUF in the app or the `Darwin-macOS-arm64` / CPU `sd-cli` builds, with RAM for the image model plus the 18.2 GB text encoder (not tested by us).

### Recommended settings <a href="#recommended-settings" id="recommended-settings"></a>

| Setting                | Unsloth (FP8 or GGUF)                                    | stable-diffusion.cpp                                          |
| ---------------------- | -------------------------------------------------------- | ------------------------------------------------------------- |
| **Steps**              | **6** (locked)                                           | `--steps 6`                                                   |
| **Guidance / CFG**     | **1**                                                    | `--cfg-scale 1.0`                                             |
| **Sampler / schedule** | Set by Unsloth                                           | `--sampling-method euler --sigmas minimax-h3-image --rng cpu` |
| **Negative prompt**    | Leave empty                                              | Leave empty                                                   |
| **Size**               | `2048×2048` to start (the **Model sizes** list has more) | `-W 2048 -H 2048`                                             |
| **Seed**               | Any number to repeat a result                            | `--seed 42`                                                   |

Best at about 4 MP and up. **Model sizes** lists native sizes such as `2368×1760` (4:3) and `2720×1536` (16:9), plus 16 MP sizes like `4096×4096` (about 4x memory and time).

### Troubleshooting <a href="#troubleshooting" id="troubleshooting"></a>

<details>

<summary>MiniMax-H3-Image is not in the model picker</summary>

[Update Unsloth](https://unsloth.ai/docs/get-started/install/updating) and restart. If a model is loaded, the picker opens on **On Device**: click **Recommended**. Macs and CPUs only list the GGUF.

</details>

<details>

<summary>Generation takes minutes</summary>

Turn **Enhance prompt** off. The GGUF is also about 3x slower than the FP8 repo on NVIDIA.

</details>

<details>

<summary>The API ignores the model I asked for</summary>

Load MiniMax-H3-Image on the Images page, or turn on **Switch image and video model by request** in **Settings → API**.

</details>

<details>

<summary>sd-cli says libcudart.so.12 or libcudart.so.13 is missing</summary>

Use the `cuda13` zip with CUDA 13, `cuda12` with CUDA 12, or the Vulkan build.

</details>

See also [How to Run Image Diffusion Models Locally](https://unsloth.ai/docs/basics/diffusion-image.md).


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation by asking a question.

Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters:

```
GET https://unsloth.ai/docs/models/minimax-h3-image.md?ask=<question>&goal=<user_goal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.
