| --- |
| base_model: briaai/FIBO |
| pipeline_tag: text-to-image |
| library_name: mlx-gen |
| license: other |
| license_name: bria-fibo |
| license_link: https://huggingface.co/briaai/FIBO |
| tags: |
| - mlx |
| - mlx-gen |
| - mflux |
| - apple-silicon |
| - 8-bit |
| - mixed-q8-bf16 |
| - fibo |
| - text-to-image |
| --- |
| # FIBO 8-bit for MLX-Gen |
|
|
| This repository contains MLX-Gen saved weights for [`briaai/FIBO`](https://huggingface.co/briaai/FIBO). |
| It is designed for local Apple Silicon inference with [`mlx-gen`](https://github.com/lpalbou/mlx-gen). |
|
|
| This is an MLX-Gen/mflux saved-weight folder, not a Diffusers or Transformers |
| `from_pretrained()` checkpoint. |
|
|
| ## License and access |
|
|
| This checkpoint is a quantized derivative of the gated `briaai/FIBO` source model and follows the |
| source model's `bria-fibo` non-commercial license terms. Access to this derivative should require |
| users to accept the upstream FIBO terms before downloading the files. |
|
|
| ## Capability |
|
|
| This package is for base FIBO text-to-image generation only. |
|
|
| FIBO Edit is a separate upstream model family and is not included in this package. Current MLX-Gen |
| public capability discovery intentionally does not expose FIBO Edit through unified `mlxgen generate` |
| until source-model and prepared-package edit validation pass. |
|
|
| ## Quantization |
|
|
| This is a mixed q8/BF16 FIBO package: |
|
|
| - q8 for quantizable FIBO transformer and text-encoder linears. |
| - BF16 for the VAE. |
| - BF16 for FIBO conditioning, timestep, caption-projection, normalization, and output-projection |
| paths that are precision-sensitive in MLX-Gen. |
|
|
| ## Validation |
|
|
| Measured on an Apple M5 Max with 128 GB unified memory using `mlx-gen 0.18.11`. |
|
|
| Prompt file: [`validation/fibo_seed103_prompt.json`](validation/fibo_seed103_prompt.json) |
|
|
| Settings: `width=512`, `height=512`, `steps=8`, `guidance=5`, `seed=103`. |
|
|
| | Package | Disk | Time | Max RSS | MAE vs source | |
| | --- | ---: | ---: | ---: | ---: | |
| | Source `briaai/FIBO` | 23.8 GiB cache | 15.13 s | 24.13 GB | baseline | |
| | This q8/BF16 package | 14.5 GiB | 16.45 s | 15.89 GB | 8.81 | |
| | q4/BF16 package | 10.2 GiB | 15.24 s | 11.39 GB | 17.56 | |
|
|
|  |
|
|
| Detailed validation notes: [`validation/fibo_base_q8_q4_seed103_metrics.md`](validation/fibo_base_q8_q4_seed103_metrics.md) |
|
|
| ## Usage |
|
|
| ```bash |
| python -m pip install -U mlx-gen |
| |
| mlxgen download --model AbstractFramework/fibo-8bit |
| |
| mlxgen generate \ |
| --model AbstractFramework/fibo-8bit \ |
| --prompt-file validation/fibo_seed103_prompt.json \ |
| --width 512 \ |
| --height 512 \ |
| --steps 8 \ |
| --guidance 5 \ |
| --seed 103 \ |
| --output fibo-q8.png |
| ``` |
|
|
| ## Attribution |
|
|
| FIBO is published by Bria.ai. MLX-Gen is based on |
| [mflux](https://github.com/filipstrand/mflux) by Filip Strand and the original mflux contributors. |
|
|
| Quantized and contributed by [@lpalbou](https://huggingface.co/lpalbou). |
|
|
