| --- |
| license: other |
| license_name: bria-fibo |
| license_link: https://creativecommons.org/licenses/by-nc/4.0/deed.en |
| base_model: |
| - briaai/Fibo-2.0-dev |
| - briaai/Fibo-2.0-edit-turbo |
| tags: |
| - wangp |
| - text-to-image |
| - image-to-image |
| - inpainting |
| - int8 |
| - convrot |
| --- |
| |
| # Fibo 2.0 for WanGP |
|
|
| Prepared BF16 and INT8 ConvRot checkpoints for [WanGP](https://github.com/deepbeepmeep/Wan2GP). |
| Credit to **Bria AI**, creators of [Fibo 2.0 Dev](https://huggingface.co/briaai/Fibo-2.0-dev) |
| and [Fibo 2.0 Edit Turbo](https://huggingface.co/briaai/Fibo-2.0-edit-turbo). |
|
|
| | Model | Workflows | Starting settings | |
| |---|---|---| |
| | Dev | Text-to-image, reference editing and masked editing | 50 steps, guidance 5 | |
| | Edit Turbo | Reference and masked editing; source image required | 4 steps, guidance 1 | |
|
|
| Both accept up to four ordered references and optional structured captions. In WanGP, |
| Main Image Ref derives the output dimensions from the first reference; the alternate |
| reference mode keeps the selected dimensions. Masked editing uses one image and mask. |
| Outside-mask pixels and exact requested object counts are not guaranteed. |
|
|
| Transformer checkpoints are at the repository root. The shared Qwen3-VL-4B-Instruct |
| conditioning encoder and tokenizer are in `Qwen3-VL-4B-Instruct/`; the FP16 Flux2 VAE |
| and optional text-rendering fonts are in `fibo2/`. These are WanGP checkpoints, not a |
| complete Diffusers pipeline directory. Place files in the same relative hierarchy |
| under WanGP's checkpoint directory, or let WanGP download the selected format. |
|
|
| BF16 transformers retain the original tensor values. INT8 uses ConvRot with BF16 |
| unquantized parameters and FP32 quantization scales. The VAE has FP16 parameters |
| and preserves its FP32 normalization statistics. Both formats were validated with |
| real generation/editing through WanGP and MMGP; enhancement remains opt-in. |
|
|
| ## Sources and licensing |
|
|
| These are repackaged/quantized derivatives, not new base models. Upstream revisions: |
|
|
| - Dev: `530604492bf2bd1364ae43ca6b2bf2285ef0f2ba` |
| - Edit Turbo: `63316ff6d5c4c4ba37729a8a6510bced7e8f585b` |
|
|
| Bria's model license and restrictions continue to apply. The upstream model cards |
| link to [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/deed.en) and state |
| that commercial use requires a separate agreement with Bria; see |
| [Bria licensing](https://platform.bria.ai/licensing/pricing?model=fibo-2). |
| Original model cards are retained under `upstream/` for attribution and terms. |
| Qwen conditioning and the Flux2 autoencoder retain their respective upstream terms: |
| [Qwen3-VL](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct) and |
| [FLUX.2](https://huggingface.co/black-forest-labs/FLUX.2-dev). |
| Aileron and DejaVu retain the license files shipped beside the fonts. |
|
|
