akhaliq/Fibo-2.0-edit-turbo-fp8

Fibo-2.0-edit-turbo — fp8

fp8 (float8_e4m3fn) storage quant of the transformer in briaai/Fibo-2.0-edit-turbo — Bria's 4-step, guidance-free image-editing model, released 2026-10-20. All other components (SmolLM3 text encoder, VAE, scheduler, tokenizer) are untouched bf16 copies from the original repo.

• ~16.3 GB → ~8.2 GB transformer weights: half the download, and half the resident VRAM in runtimes that keep fp8 in place (ComfyUI upcasts during ops on pre-Hopper GPUs, so quality is preserved there too).
• Same pipeline class (BriaFibo2Pipeline), same prompts, same 4-step settings.

Usage

With the ComfyUI nodes: akhaliq/Fibo-2.0-ComfyUI — point the model_id widget at this repo or set FIBO_EDIT_MODEL=akhaliq/Fibo-2.0-edit-turbo-fp8.

With diffusers (BriaFibo2Pipeline, currently in diffusers PR #14991):

import torch
from diffusers import BriaFibo2Pipeline

pipe = BriaFibo2Pipeline.from_pretrained(
    "akhaliq/Fibo-2.0-edit-turbo-fp8", dtype=torch.bfloat16
)
pipe.enable_model_cpu_offload()

image = pipe(
    prompt="make the sky dramatic and stormy",
    image=source_pil_image,
    num_inference_steps=4,
    guidance_scale=1.0,
).images[0]

Loading with dtype=torch.bfloat16 upcasts the fp8 transformer on load (storage savings only); runtimes that load fp8 natively get the VRAM savings.

License

CC-BY-NC 4.0, inherited from the source model — non-commercial use, attribution to Bria. The source repo is gated; accepting briaai/Fibo-2.0-edit-turbo's license is recommended before downloading.

Reproduce

The conversion (bf16 → e4m3fn cast of every transformer weight) plus a bf16 smoke edit runs from quantize_fp8.py in this repo:

hf jobs uv run --flavor a100-large --timeout 2h --secrets HF_TOKEN \
  https://huggingface.co/akhaliq/Fibo-2.0-edit-turbo-fp8/resolve/main/quantize_fp8.py

(HF_TOKEN must be exported locally with access to the gated source repo.)