| --- |
| license: other |
| license_name: minimax-h3-image-community-license-agreement |
| license_link: LICENSE |
| library_name: mlx |
| pipeline_tag: text-to-image |
| base_model: MiniMaxAI/MiniMax-H3-Image |
| base_model_relation: quantized |
| language: [en] |
| tags: [mlx, apple-silicon, mac, macos, text-to-image, image-to-image, image-editing, diffusion, bf16, minimax, minimax-h3, minimax-h3-image] |
| extra_gated_heading: "MiniMax-H3-Image Community License" |
| extra_gated_description: >- |
| These weights are a modified distribution of MiniMax-H3-Image, licensed under the MiniMax-H3-Image Community |
| License Agreement and its Acceptable Use Policy. The licence excludes the European Union, the United Kingdom, the |
| Republic of Korea and the United States of America from its territory; use there requires separate authorization |
| from MiniMax (https://platform.minimax.io/license). Access is approved automatically. |
| extra_gated_prompt: >- |
| By requesting access you agree to the MiniMax-H3-Image Community License Agreement (LICENSE in this repository) |
| and its Acceptable Use Policy, and you confirm that you will not use the weights or their outputs outside the |
| licence's Applicable Territory unless MiniMax has authorized you to. |
| extra_gated_fields: |
| I agree to the MiniMax-H3-Image Community License and Acceptable Use Policy: checkbox |
| extra_gated_button_content: "Agree and access" |
| --- |
| |
| # MiniMax-H3-Image for MLX (bf16) |
|
|
|  |
|
|
| [MiniMax-H3-Image](https://huggingface.co/MiniMaxAI/MiniMax-H3-Image) for Apple Silicon. It does text-to-image and image editing with a 33B model, at up |
| to ~16 MP, in 6 steps. Runner: [felix-ab/minimax-h3-image-mlx](https://github.com/felix-ab/minimax-h3-image-mlx). Unofficial and not affiliated with MiniMax. |
| Powered by MiniMax-H3-Image. |
|
|
| | | | |
| |---|---| |
| | Download | 94 GB | |
| | Memory | 5 GB at 2048×2048, 15 GB at 4736×3552 | |
| | Time (M4 Pro Mac mini, 48 GB) | 3 min at 2048×2048, 17 min at 4736×3552 | |
|
|
| These are the original bf16 weights, not quantized, in MLX format. They are read from disk one transformer block at a time. |
|
|
| ## Usage |
|
|
| Accept the licence on this page, then: |
|
|
| ```bash |
| pip install git+https://github.com/felix-ab/minimax-h3-image-mlx |
| hf auth login |
| minimax-h3-mlx @prompt.txt --size 4736x3552 --seed 42 -o out.png |
| minimax-h3-mlx @edit.txt --image ref.png --size 4096x4096 --seed 1234 -o edit.png |
| ``` |
|
|
| The second command edits `ref.png`; the prompt refers to it as `<Picture 1>`. The model expects long, detailed |
| English prompts ([examples](https://github.com/felix-ab/minimax-h3-image-mlx/tree/main/examples/prompts)). Sizes under 4 MP are generated at 4 MP and scaled |
| down, as in MiniMax's pipeline. |
|
|
| ## Accuracy |
|
|
| - **float32:** output matches the PyTorch reference (diffusers / transformers) to within 3e-5 after 6 steps. |
| - **bf16:** output differs from the reference about as much as PyTorch's own bf16 backends differ from each other. |
| - **Seeds:** they match MiniMax's pipeline. Their published seed-42 example gives the same image (left: MiniMax, |
| right: this port). |
|
|
|  |
|
|
| Full measurements: [VALIDATION.md](https://github.com/felix-ab/minimax-h3-image-mlx/blob/main/docs/VALIDATION.md). |
|
|
| ## Why there is no 8-bit version |
|
|
| On an M4 Pro it gave the same image quality but ran slower: 218 s against 193 s at 2048×2048. Unless streamed, it |
| also needed 27 GB of memory. Its only advantage would be a smaller download. For quantized weights, see |
| [mlx-vlm](https://github.com/Neywa-Labs/mlx-vlm), which supports MiniMax-H3-Image at |
| [8-bit](https://huggingface.co/nativ-community/MiniMax-H3-Image-8bit) and |
| [4-bit](https://huggingface.co/nativ-community/MiniMax-H3-Image-4bit). |
|
|
| ## Examples |
|
|
| Generated at ~16 MP with this repository and shown downscaled. The prompts are in [examples](https://github.com/felix-ab/minimax-h3-image-mlx/tree/main/examples/prompts). |
|
|
| | | | |
| |---|---| |
| |  |  | |
| |  |  | |
| |  |  | |
|
|
| ## Changes from the original files |
|
|
| - The transformer and text encoder are split into one file per block or layer. |
| - Each block's AdaLN projection is replaced by its precomputed output for the 7 timesteps the sampler uses. The |
| results are identical. |
| - The text encoder ends after layer 49; the model only reads `hidden_states[50]`. |
| - The VAE and tokenizer are unchanged. |
|
|
| Source: [`MiniMaxAI/MiniMax-H3-Image@492fd8c`](https://huggingface.co/MiniMaxAI/MiniMax-H3-Image/tree/492fd8c4eca666f1085f57e30738c1e0d86b3439). |
|
|
| ## Licence |
|
|
| [MiniMax-H3-Image Community License Agreement](LICENSE), including its Acceptable Use Policy and territorial |
| restrictions. The example images are AI-generated. |
|
|
