Research methodology

A practical standard for claims we can reproduce.

Lewdora’s methodology separates product facts from measured results. A benchmark is published only when the prompt set, settings, sample count, evaluation criteria, and limitations are clear enough for another person to understand what was tested.

Start with free credits View pricing

Short answer

What this means in practice.

Lewdora evaluates AI image workflows by fixing the inputs, changing one variable at a time, running multiple samples, recording the active model and settings, and publishing both evaluation criteria and limitations. No result is generalized beyond the tested setup.

Fixed prompts and disclosed generation settings
Multiple outputs instead of a single cherry-picked image
Separate criteria for identity, pose, anatomy, and prompt fit
Visible limitations and dates for every published result

The details that affect the result.

Define the claim before running the test

A useful test starts with a narrow question such as whether character traits remain recognizable across scene changes or whether a pose reference preserves the intended structure. The prompt set, model, dimensions, settings, references, and sample count are recorded before interpreting outputs.

Change one meaningful variable

When comparing models or controls, all other available inputs stay fixed. Each condition receives multiple runs because generative systems vary. Outputs are assessed against declared criteria such as identity traits, prompt adherence, composition, anatomical defects, or reference influence.

Publish evidence with limitations

A result should include the test date, active product version, model labels shown to users, prompt set, settings, reference rights, sample handling, and known limitations. Lewdora does not convert an internal impression into a benchmark claim or present a small test as universal performance.

Workflow

A useful way to start.

State a narrow question
Define what success means before generating any images.
Freeze the test inputs
Record prompts, settings, references, models, and sample count.
Run and score every sample
Keep failed and imperfect outputs in the evaluated set.
Publish the limits
Explain what the test can and cannot support, with a clear date.