Candidate AIdentity hidden
Choose better models,
with evidence.
Run controlled, blind evaluations across generative media. Generate on fal or bring finished outputs, then turn team judgment into a decision you can defend.
- Image
- Video
- Audio
- Text
- Multimodal
Material fidelity evaluation
Product preview · Blind pairwiseCandidate A
Candidate BOrder varies by reviewer
Prefer ATiePrefer B
01 Controlled inputs
02 Media-native review
03 Decision-ready evidence
One evaluation system
Everything around the vote matters.
GenMedia keeps the question, generation contract, blind presentation, reviewer context, and statistical evidence in one traceable workflow.
Release candidate evaluationEvidence building
Current evidence
Candidate B leads
Directional signal across motion, temporal consistency, and prompt adherence.
Candidate A38%
Candidate B62%
24 prompts18 reviewers92% coverage
Controlled generation
The endpoint schema is part of the Evaluation Task
GenMedia resolves current fal OpenAPI contracts, maps each prompt or reference input explicitly, and preserves the exact run lineage behind every output.
Media-native review
Judging tools match the artifact
Compare images, video, audio, text, and 3D-oriented outputs with blind pairwise or rating workflows. Video pairs can move into DeltaFrame for frame-level inspection.
Statistical context
A ranking is not automatically a conclusion
Weighted rankings sit beside uncertainty intervals, pair coverage, multiplicity-aware tests, position checks, and clear language when the evidence is incomplete.
Traceable evaluations
Prompts and decisions stay connected
Prompt packs, projects, taxonomy tags, immutable runs, and privacy-safe exports preserve the path from an evaluation to a release decision.
One trail from Evaluation Task design to release decision.
A deliberate operating loop
From a question to a confident release.
- 01
Define the decision
Choose the comparison intent, media modality, held-constant inputs, audience, and success criteria before any candidate is reviewed.
- 02
Bring or generate media
Import finished outputs from any source or generate directly from schema-verified fal endpoints with explicit prompt and input bindings.
- 03
Review without the reveal
Randomized candidate order, completion gates, and media-specific controls keep model identity out of the judgment loop.
- 04
Act on measured evidence
Read the fast summary first, then inspect uncertainty, pair coverage, prompt effects, bias checks, and the next best sampling action.
DeltaFrame
See the difference,
frame by frame.
Step through synchronized video, compare split and difference views, and jump to the exact moments where motion or detail breaks down.
Explore DeltaFrame
A · 00:08.42
B · 00:08.42Evidence that respects people
Good evaluation protects the reviewer, too.
Blind protocols reduce bias. Uncertainty stays visible. Reviewer reliability is handled conservatively and kept out of ordinary result views.
Blind by default
Evaluator-facing candidate order is deterministic and shuffled. Model identities and aggregate results remain hidden until the task completion gate is satisfied.
Ambiguity is protected
Reviewer reliability uses independent peer pools and treats strong cross-pool disagreement as uncertain evidence, not as an automatic reviewer failure.
History is append-only
Assigned inputs and past outputs are not silently rewritten. New prompts or endpoints create a new run while earlier lineage remains inspectable.
Ready when the question is
Build the eval before
the conclusion.
Create a controlled Evaluation Task in the workspace, or use the documentation to align your team on protocols, roles, and evidence.