Hugging Face's logo

AcademiaSD
/
Fibo-2.0-NF4-for-LoRA-Training

Text-to-Image
Diffusers
Safetensors
English
BriaFibo2Pipeline
fibo-2
fibo
bria
nf4
bitsandbytes
lora-training
image-editing
academiasd
Model card Files
xet
 Community
How to use from the
Use from the
Diffusers library
pip install -U diffusers transformers accelerate
import torch
from diffusers import DiffusionPipeline

# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("AcademiaSD/Fibo-2.0-NF4-for-LoRA-Training", dtype=torch.bfloat16, device_map="cuda")

prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]

Fibo 2.0 NF4 for LoRA Training

Not an official BRIA release, and not made for inference.

This repository holds an NF4 conversion of Fibo 2.0 Dev, made so that the Fibo 2.0 trainer of AcademiaSD LoRAlab Trainer Studio can download it and train LoRAs on consumer NVIDIA GPUs (from 8 GB of VRAM).

The weights are the property of BRIA AI and are redistributed here only to make LoRA training possible on a single consumer GPU. The transformer is stored in the trainer's own pre-quantized cache format, which diffusers and ComfyUI do not load: this repository is for training, not for generating images.

Academic and personal use only. Fibo 2.0 is published under CC BY-NC 4.0: non-commercial use only. You may not use the model, the LoRAs trained with it, or any derivative for commercial purposes, for a product or service, or for anything that brings you money. Commercial use of Fibo 2.0 requires a separate agreement with BRIA (see the model page). Read the terms of the original model before you use it: they are the ones that apply here too.

You do not need to download this repository by hand: the trainer downloads only the parts it uses on first use.

Contents

Folder Content Size
transformer/ Fibo 2.0 Dev transformer (8.1 B): the attention, MLP and cross attention of the 44 blocks and of the 2 refiner blocks in NF4 (bitsandbytes, pre-quantized), 342 layers; the perceptor that turns the caption into gist tokens, the adaLN modulation, the embedders and the output layer in BF16, 73 layers 5.2 GB
text_encoder_NF4/ Qwen3-VL-4B text encoder (the model's own) in NF4 2.9 GB
text_encoder_INT8/ The same text encoder in INT8 (bitsandbytes LLM.int8) 4.8 GB
text_encoder_BF16/ The same text encoder in BF16, unchanged 8.9 GB
vae/ AutoencoderKLFlux2 (the FLUX.2 VAE, 32 latent channels) 0.3 GB
processor/, scheduler/, fonts/ Unchanged from the original model; the fonts are the ones Fibo 2.0 was trained with, to draw the text of the captions 0.01 GB
Total with the three text encoders 22.1 GB (original: 24.3 GB)

With only the NF4 text encoder (the default of the trainer) the download is 8.4 GB. The pre-cache pulls just the variant it is going to use and skips the other two.

Format

  • transformer/ is not a standard diffusers checkpoint. It is a pre-quantized cache: one .safetensors file per layer in weights/ (the NF4 weight plus its packed bitsandbytes QuantState, or BF16 for the excluded layers), others.safetensors with the norms, the modulations and the perceptor, and index.json. The trainer builds an empty BriaFibo2Transformer2DModel from config.json and fills it from this cache, so the BF16 transformer is never downloaded or loaded. metadata.json lists the layers kept in BF16.
  • text_encoder_NF4/, text_encoder_INT8/ and text_encoder_BF16/ load with transformers: Qwen3VLForConditionalGeneration.from_pretrained(".../text_encoder_NF4").
  • vae/, processor/ and scheduler/ are unchanged from the original model.
  • The converter that produced all of this is tools/fibo2/5_conversor_Fibo2_NF4.py in the Trainer Studio repository.

NF4 is meant for training. The quantization of the transformer costs a little accuracy in the weights, and a pre-quantized cache cannot be used to generate images at all. For inference, use BRIA's original model.

LoRAs trained with it

The trainer exports LoRAs in the format ComfyUI's loader takes (diffusion_model.<layer>.lora_A / lora_B plus alpha).

ComfyUI runs its own repack of Fibo 2.0, where the attention is a single qkv matrix, the MLP a single w13 and the cross attention a to_kv, so the layers that fuse are rewritten as one LoRA of several times the rank: the change that ComfyUI applies is exactly the same as the one of the diffusers LoRA.

Measured on an RTX 5080 at 512×512, rank 8: ~8.7 GB of VRAM and ~2 s per step.

License and ownership

The weights belong to BRIA AI. This repository is a redistribution of briaai/Fibo-2.0-dev in a different storage precision, for LoRA training only.

  • The model is published by BRIA AI under CC BY-NC 4.0 (licence): non-commercial use only.
  • Academic research and personal use only. No commercial use, no products, no services, no paid work, and nothing derived from it that is used commercially.
  • The LoRAs you train with it inherit those terms: they are derivatives of BRIA's weights.
  • For commercial use, contact BRIA: https://platform.bria.ai/licensing/pricing?model=fibo-2.0.
  • The original model's page, its gated access form and its terms are the ones that apply: https://huggingface.co/briaai/Fibo-2.0-dev.

The code of the trainer is MIT licensed; the weights are not.

Links

Downloads last month
11
Inference Providers NEW
Text-to-Image
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for AcademiaSD/Fibo-2.0-NF4-for-LoRA-Training

Base model

briaai/Fibo-2.0-dev
Quantized
(6)
this model