File size: 2,760 Bytes
72bf0a8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 | ---
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.
|