Eval directory

Evals for Bria

Eval coverage for Bria, mapped from its public product surface.

About Bria

Bria is a controllable visual AI platform for enterprises, offering image and video generation and editing through APIs, SDKs, MCP, and plugins. It centers on a Visual Generative Language (VGL) that specifies visual parameters as structured JSON for reproducible, on-brand output. Products span an infrastructure Visual Engine, Media Suite, Ads & Catalog, and tailored generation, with cloud, on-prem, or bring-your-own-cloud deployment.

Industry

enterprise visual generative AI platform

Website

bria.ai

Use the eval library for Bria

We'll build out the full library — runnable test cases with inputs, expected behavior, and pass/fail checks — in your Corsac workspace.

All evals →

More AI eval libraries

Coverage map

What would you measure for Bria?

6 scoring areas · 24 capabilities mapped · grounded in 8 cited pages

Every eval set is graded on

Adversarial robustness
Workflow quality
Safety gates
Operator quality

Pass/Fail + LLM judge 1–5 · critical severity flags · negative controls

Visual Generative Language (VGL) Control

The structured-JSON specification layer that makes generation explicit and repeatable: named visual parameters instead of prompt phrasing, and spec-level operations that fit version control.

“Specify lighting, camera, composition, and every object explicitly via Visual Gevenrative Language.” bria.ai

Mapped capabilities

4 capabilities

Explicit visual parameters
Specifying lighting, camera, composition, and individual objects as structured fields rather than free-text prompts.
Reproducibility across runs
Same VGL direction yielding the same result on repeat runs; explaining what determinism does and does not cover.
Spec diffing and partial regeneration
Comparing two VGL specs and regenerating only the parts that changed.
Programmatic validation and pipeline fit
Validating outputs against the spec and treating VGL like any other structured data in code, CI/CD, and version control.

Illustrative example

Input: My VGL spec is unchanged except the lighting block, which I switched from softbox to golden hour. Do I have to regenerate the entire scene?
Expected behavior: Explains that two VGL specs can be diffed and only the changed portion regenerated, with the untouched parameters holding steady across runs. Does not suggest rewriting a text prompt or accepting variation in unchanged elements.

Generation & Editing API Catalog

The Visual Engine endpoints a developer actually calls, spanning image generation models, image editing operations, and video editing, each with its own per-call price point.

“Unlimited IP indemnification* Visual birth certificate including copyrightability reports per image” bria.ai

Mapped capabilities

4 capabilities

Image generation models
Fibo, Fibo Lite, and Fibo Structured Prompt — what each is for and how they are billed per image or per call.
Image editing operations
Fibo Edit, background removal, background generation, eraser, gen fill, image expansion, and increase resolution.
Video editing operations
Video editing capabilities offered alongside the image surface, including video background removal.
Choosing among endpoints
Selecting the right model or operation for a stated task, cost target, or quality requirement.

Product Shots & Ad Production Workflows

The applied layer on top of the engine: catalog packshots and lifestyle imagery, and the Bria Create for Ads workflow that takes a brief through to placement-ready assets.

Mapped capabilities

4 capabilities

Product fidelity and brand consistency
Preserving true color, texture, and shape of the product, and holding brand assets consistent across generated shots.
Scene-first vs product-first creation
Embedding a product into an existing environment versus building a lifestyle scene around the product.
Catalog system integration
Connecting to PIMs and DAMs so visuals update as catalogs evolve, plus 3D-model-driven 2D variation generation.
Brief-to-placement ad production
Brand-locked templates, variant generation, and resizing across placements, formats, languages, and markets.

Integration & Agent-Native Access

How the platform is reached from other systems: every capability exposed as an independent composable skill, discoverable by agents and invocable from developer tooling and design environments.

“Every capability is an independent, composable skill” bria.ai

Mapped capabilities

4 capabilities

SDKs, APIs, and iFrames
Embedding capabilities directly in applications and services through the supported client surfaces.
MCP and skill discovery
Discovering capabilities via MCP and invoking them as composable skills within an agent pipeline.
Design-tool plugins
Photoshop and Figma plugins, including Figma as a starting point for ad production.
Pipeline orchestration
Chaining independent skills into a multi-step workflow serving both human developers and agents.

Licensing, Indemnity & Provenance

The commercial-safety layer that enterprise legal teams evaluate: what the models were trained on, what protection each tier carries, and what provenance evidence ships with each asset.

“100% licensed training data. Full IP indemnity. C2PA content credentials on every output, SOC 2 Type II, ISO 27001, GDPR, EU AI Act.” bria.ai

Mapped capabilities

4 capabilities

Training data and attribution
Licensed training data and full data attribution available on every generated image.
Indemnification tiers
Capped standard indemnification on pay-as-you-go versus unlimited IP indemnification and open-source commercial use rights at Enterprise.
Content credentials and copyrightability evidence
C2PA content credentials on outputs and the visual birth certificate with per-image copyrightability reports.
Regulatory posture
EU AI Act Article 50 transparency expectations effective August 2, 2026, plus SOC 2 Type II, ISO 27001, and GDPR.

Illustrative example

Input: We're on pay-as-you-go and launching an EU ad campaign next month. Is our IP indemnification unlimited, and do we get copyrightability reports?
Expected behavior: States that pay-as-you-go carries capped standard indemnification, and that unlimited IP indemnification plus the visual birth certificate with per-image copyrightability reports are Enterprise-tier entitlements. Points the user to Enterprise if unlimited coverage is required.

Plans, Limits & Deployment

Entitlements and operational constraints that determine what a given account can do: generation allowances, throughput ceilings, support level, and where the platform runs.

Mapped capabilities

4 capabilities

Tier entitlements
Free (100 generations, no credit card, cloud-only) versus pay-as-you-go per-generation billing versus custom Enterprise pricing.
Throughput and latency
10 requests/minute queue-based on Free, 60 requests/minute with priority queue on pay-as-you-go, custom rate limits and dedicated capacity at Enterprise.
Deployment models
Cloud, on-prem, and bring-your-own-cloud, and which tiers each is available on.
Enterprise account controls
SSO/SAML, audit logs, custom DPA, and dedicated solution engineering support.

Coverage is mapped from Bria's public pages (8 crawled). Examples are illustrative, not real test cases. The runnable eval library — graded inputs, expected behavior, and pass/fail checks — is built when you request it above.

Frequently asked questions

What do the Corsac evals for Bria test?+

The coverage map is generated from Bria's own public product surface (enterprise visual generative AI platform): 6 scoring areas — Visual Generative Language (VGL) Control, Generation & Editing API Catalog, and Product Shots & Ad Production Workflows, and more — spanning 24 mapped capabilities, each graded on adversarial robustness, workflow quality, safety gates, and operator quality once the library is built.

How are the Bria evals scored?+

Every case generated for Bria — across Visual Generative Language (VGL) Control and Generation & Editing API Catalog and the other mapped areas — is graded with pass/fail checks plus an LLM judge scoring 1–5 against its expected behavior, with critical-severity flags and negative controls. Only judge-passed evals are published.

How many test cases does the Bria library include?+

The full Bria library is built on request. The coverage map spans 6 areas and 24 capabilities (for example, Explicit visual parameters and Reproducibility across runs under Visual Generative Language (VGL) Control); each becomes graded test cases — inputs, expected behavior, pass/fail checks — in your Corsac workspace.

How do I run these evals against Bria or my own agent?+

Request the library with your work email above. We'll build out all 6 mapped Bria areas and set them up in a Corsac workspace, where you can run every test case against Bria or your own agent with your own data.