Run Fibo 2 Turbo image editing on your own hardware with a self-hosted license. Fibo 2 Turbo edits images only: give it 1 to 4 images and an instruction, and it returns one edited image. This guide installs the model, downloads its weights from Bria, and runs a first edit.
Fibo 2 Turbo is Fibo 2's edit model, distilled to run in 4 steps without guidance. An edit of one image takes about 6 seconds on an NVIDIA L40S (four images: about 18 seconds). It does not generate an image from a prompt alone. To generate, use Fibo 2.
Each model is licensed on its own: a Fibo 2 license does not cover Fibo 2 Turbo, and a Fibo 2 Turbo license does not cover Fibo 2. You do not need a Fibo 2 license to run Fibo 2 Turbo, although its package installs the fibo-2 package.
| Fibo 2 Turbo | Fibo 2 | |
|---|---|---|
| What it does | Edits 1 to 4 images | Generates and edits images |
| GPU memory | 32 GB or more | 40 GB |
| Python | 3.11 or 3.12 | 3.11 or 3.12 |
| LLM setting for free-form instructions | FIBO_2_TURBO_VGL_LLM | FIBO_2_VGL_LLM |
- An active Fibo 2 Turbo self-hosted license (pricing)
- One of your organization's API keys
- A machine with an NVIDIA GPU with 32 GB of memory or more, Python 3.11 or 3.12, and about 26 GB of disk for the weights
- For free-form instructions only: an LLM that turns them into a structured VGL prompt, on your own provider account. A VGL that carries an
edit_instructionneeds no LLM (see LLM for free-form instructions) - For the manual install:
curlandjq, and network access to the Bria Engine, AWS CodeArtifact and Bria's weights CDN (no Hugging Face account) - On H100 and H200 GPUs only: network access to
huggingface.cowhen the model starts, for its FlashAttention-3 kernel (no account needed)
Buy a Fibo 2 Turbo license on the pricing page. You manage it, and see the installation steps, on the Licenses tab in the Console.
Bria Studio installs the package, downloads the weights, and runs a test for you. It needs Node.js 18 or newer, uv and an NVIDIA GPU.
npx @bria-ai/studioChoose Set up a Bria model, then pick FIBO 2 Turbo (edit) when Studio lists your licensed models. It installs fibo-2-turbo into a Python 3.12 virtual environment at ~/.bria/venvs/fibo-2-turbo, downloads the weights, and edits a sample photo to confirm the setup. The test edit uses a VGL that carries the edit instruction, so it needs no LLM.
Fibo 2 Turbo needs Studio 0.6.0 or newer. If Studio does not list it, run npx @bria-ai/studio@latest.
A VGL that carries an edit_instruction needs no LLM. A free-form instruction (prompt) goes to an LLM on your own account, with the images, and the LLM writes the VGL of the edit. Set FIBO_2_TURBO_VGL_LLM to any model name litellm supports, and set that provider's credentials. If FIBO_2_TURBO_VGL_LLM is not set, Fibo 2 Turbo uses Gemini 3.5 Flash-Lite: on Google AI Studio when GEMINI_API_KEY or GOOGLE_API_KEY is set, otherwise on Vertex AI with your Vertex AI credentials.
# For example, OpenAI:
export FIBO_2_TURBO_VGL_LLM="openai/gpt-5.4"
export OPENAI_API_KEY="<your-openai-api-key>"Fibo 2 Turbo does not read Fibo 2's FIBO_2_VGL_LLM, so a Fibo 2 install on the same machine does not change it.
setup() checks the LLM setting before it downloads the weights. If FIBO_2_TURBO_VGL_LLM is set, but litellm does not know it or its provider's credentials are missing, setup() stops with an error. The error names the missing variables, never their values. If no LLM is set up at all, setup() continues and logs a warning: a request with a vgl works, and a request with a prompt stops with an error.
- The LLM must read images. Every request is an edit, so the LLM gets the images.
- Quality. Bria tuned the VGL prompts on Gemini 3.5 Flash-Lite. Other LLMs write different VGLs, so the edits can be different.
- Models that do not work yet. GPT-4o fails on every request with a
prompt: use GPT-4.1 or a newer OpenAI model. Gemini through a proxy (for exampleopenrouter/google/…) can fail on edits: leaveFIBO_2_TURBO_VGL_LLMunset, or usegemini/…orvertex_ai/…. - Cost. You pay your provider for the LLM calls. The LLM does not change the usage that Bria counts.
Run from the environment the model is installed in. If you used Studio, activate the venv it created:
source ~/.bria/venvs/fibo-2-turbo/bin/activateThen load the model with your API key and edit an image. On first load setup() checks your license and downloads the weights (25.5 GB) to ~/.bria/models; later runs reuse them. On an AWS server with an NVIDIA L40S, the download took about 3.5 minutes. Your time depends on your network.
import os
from io import BytesIO
from urllib.request import urlopen
from PIL import Image
from fibo_2_turbo import Fibo2Turbo, Fibo2TurboConfig, Fibo2TurboInput
# Reads BRIA_API_TOKEN; checks the license and downloads ~26 GB of weights on first load.
pipe = Fibo2Turbo(config=Fibo2TurboConfig(api_token=os.environ["BRIA_API_TOKEN"]))
pipe.setup()
# Fibo 2 Turbo edits images: one to four of them, loaded with PIL
url = "https://bria-test-images.s3.us-east-1.amazonaws.com/strawberry.jpg"
photo = Image.open(BytesIO(urlopen(url).read()))
# A structured prompt (VGL) that carries the edit instruction needs no LLM
vgl = '{"edit_instruction": "Make the strawberry blue"}'
result = pipe.execute(Fibo2TurboInput(images=[photo], vgl=vgl, seed=42))
result.image.save("edited.png")
# A free-form instruction needs an LLM (see "LLM for free-form instructions")
# result = pipe.execute(Fibo2TurboInput(images=[photo], prompt="Make the strawberry blue", seed=42))
pipe.cleanup() # free the GPU when you are doneFibo2TurboInput takes:
images: 1 to 4 PIL images. The instruction can refer to them as<image_1>,<image_2>and so on, in this order. Any image mode works (RGBA, palette, CMYK); the alpha channel is not used.- The instruction: a free-form
promptthat your LLM turns into VGL, or avglthat carries anedit_instruction. Avglsent with apromptis the first image's complete VGL, kept as the baseline of the edit. - Optional
seed, and the size:widthandheight, oraspect_ratiofor an edit of two or more images. Without a size, the result is about one megapixel, in the first image's shape. - Optional
mask: the region to edit, white on black. One image only, at that image's size. A masked edit keeps the image's size and its pixels outside the mask.
Fibo 2's generation fields, such as style and negative_prompt, are refused. The result carries image, the edited image, and vgl, the VGL the model read: send it back to reproduce or adjust the edit.
Inputs and outputs stay on your machine — only a small attribution vector is sent to Bria.
No license for FIBO 2 Turbo
Your account holds no active Fibo 2 Turbo license, so the install stops before anything downloads. A Fibo 2 license does not cover Fibo 2 Turbo. Buy or renew it on the Licenses tab.
LicenseError when the model starts
Your account holds no active Fibo 2 Turbo license, or the license ended. Open the Licenses tab — the error links straight to it. A card retry (past_due) stays active.
The model won't load and asks for an API token
No BRIA_API_TOKEN is set. Set it (or FIBO_2_TURBO_API_TOKEN), or pass Fibo2TurboConfig(api_token="...").
FIBO 2 Turbo edits images: send 1 to 4 'images'
The request has no image. Fibo 2 Turbo edits images only. To generate an image from a prompt, use Fibo 2, which has its own license.
A free-form instruction stops and asks for an LLM
No LLM is set up for the instruction. Set FIBO_2_TURBO_VGL_LLM and its provider's credentials, or GEMINI_API_KEY, or your Vertex AI credentials (see LLM for free-form instructions). FIBO_2_VGL_LLM is not read. Or send a vgl that carries an edit_instruction, which needs no LLM.
torch.cuda.is_available() is False during setup()
The installed PyTorch doesn't match your NVIDIA driver. pip's default PyTorch needs driver 580 or newer. Install the CUDA 12.8 build first, as in step 2.
CUDA out of memory during setup() or the first edit
Fibo 2 Turbo needs a GPU with 32 GB of memory or more. On an NVIDIA L40S, the process used up to 30,901 MiB.
WeightsError during download
The download was interrupted. Run again — it resumes where it stopped.
Each edit is counted against your Fibo 2 Turbo license, not a Fibo 2 license, and shown in Usage & billing in the Console, tagged as self-hosted. See Self-hosted models for how licensing and billing work across models.