VGL — Visual Generation Language

VGL turns a text prompt, or images plus a prompt, into a structured JSON description of a scene: objects, background, lighting, aesthetics, camera, and text. The result is a typed Pydantic model that an image model like FIBO takes as its prompt.

Install

pip install bria-vgl

The default LLM is Gemini through Google AI Studio. For other providers, install the extra:

pip install 'bria-vgl[vertex]'     # Gemini on Vertex AI
pip install 'bria-vgl[bedrock]'    # Claude on AWS Bedrock

Authenticate

llm takes any litellm model id, but the prompts are tuned for Gemini.

Quickstart

Create a scene from a prompt

import bria_vgl
from bria_vgl.dialects.thales import Generate

scene = bria_vgl.create(Generate(prompt="a red fox in snow at dusk"))
print(scene.short_description)
print(scene.model_dump_json(indent=2, exclude_none=True))

Edit from images

from PIL import Image
from bria_vgl.dialects.thales import EditImages

room = Image.open("room.jpg")
lamp = Image.open("lamp.jpg")

scene = bria_vgl.modify(EditImages(prompt="make it night", images=[room]))
scene = bria_vgl.modify(EditImages(prompt="put the lamp from <image_2> in <image_1>", images=[room, lamp]))
print(scene.edit_instruction)

Up to 4 images. The prompt refers to them as <image_1> to <image_4>, in list order.

Edit a masked area

from bria_vgl.dialects.thales import EditImageWithMask

mask = Image.open("mask.png")  # white is the area to change; same size as the image
scene = bria_vgl.modify(EditImageWithMask(prompt="replace it with a cat", image=room, mask=mask))

Edit with the image's VGL

If you have the VGL that an image was generated from, pass it as input_image_vgl. The edit then keeps everything the prompt doesn't change, rather than re-describing the image from scratch. vgl doesn't check that it matches the image.

scene = bria_vgl.create(Generate(prompt="a cosy living room"))
image = ...  # generate the image from `scene` with FIBO
night = bria_vgl.modify(EditImages(prompt="make it night", images=[image], input_image_vgl=scene))

Runnable versions are in examples/.

Options

bria_vgl.create(payload, llm="vertex_ai/gemini-3.5-flash-lite", seed=7, style_preset_prompt="Muted film colours, soft grain.")
bria_vgl.modify(payload, llm=..., seed=...)

• llm: the litellm model id; see Authenticate.
• seed: makes a call repeatable, on models that accept one.
• style_preset_prompt: style guidelines added to the system prompt. Applies to create only; modify ignores it.

create also takes overrides for the payload's defaults. Each is optional, and leaving it out keeps the default:

• system_prompt: replaces the system prompt. {llm_input_schema} and {style_preset_block} in it are still filled.
• input_schema: the field guide filled into {llm_input_schema}.
• response_schema: the JSON schema the model must follow, in Gemini's format. Extra keys in the reply are dropped from the VGL.
• thinking_level: a bria_vgl.ThinkingLevel or its string, "minimal", "low", "medium" or "high", on Gemini 3+. Anything else, or a level on an older Gemini, raises ValueError before the call. By default thinking is as low as the model allows.
• temperature and max_output_tokens: thinking tokens count against max_output_tokens, so raise it with a high thinking level.

bria-vgl logs under the bria_vgl logger hierarchy (bria_vgl.gateway, bria_vgl.prompting, …). Configure it the standard way — e.g. logging.getLogger("bria_vgl").setLevel(logging.INFO) and attach your own handler; add a logging.Filter if you need to inject contextual fields.

A Gemini call that keeps failing (3 attempts) or hits a rate limit (429, not retried) moves to gemini-3.1-flash-lite on the same provider.

Errors

Everything bria-vgl raises inherits from bria_vgl.exceptions.VGLError, except a bad payload, which raises pydantic.ValidationError before any LLM call. The classes below all live in bria_vgl.exceptions.

Dialects

A dialect is a small variant of VGL, with its own model, payloads, and prompts. The payload you import decides the dialect. vgl ships the Thales dialect: no aesthetic scores, and every edit states its change in edit_instruction.

More

• Best practices: how images are prepared, masks, and getting consistent results.
• Contributing