Research standards
Editorial Research Methodology
How Remix Camera documents prompt research and same-prompt image-model comparisons, including inputs, outcomes, metrics, and reproducibility limits.
Last updated: July 13, 2026
How to read our results
Our comparison pages are first-party, dated observations of specific runs. They can show how the named models behaved on the documented test set, but they cannot establish a permanent “best” model. The method written on an individual article controls when it is more specific than this page.
1. Define the question and test set
A comparison starts with a bounded question, such as how a set of image models handles the same group of adult portrait prompts or how often those prompts complete without a recorded block.
The article should state the number and type of cases, identify reused or newly added cases when that distinction matters, and avoid implying that a small portrait-focused set represents every image task.
2. Hold comparison inputs constant
When an article calls a test “same prompt,” each model column should receive the same documented prompt text for that case. If the workflow also uses a reference image, the comparison should identify that fact and keep the reference consistent where the article claims a direct head-to-head test. Provider-specific defaults or unavailable controls are limitations, not hidden equalizers.
3. Record outcomes before interpreting them
The exact legend on a comparison article is the source of truth. Our pages may use the following labels:
When a page reports a completion or acceptance rate, the default calculation is accepted outcomes divided by tested outcomes. Any different denominator or grouping should be defined on that page.
4. Show evidence and limits together
- Show output images when they are available and keep missing, blocked, or failed cells visible.
- Expose the prompt or enough prompt detail to explain what was tested.
- Name the models as they were labeled during the recorded run.
- State sample-size, task, reference-image, and timing limits near the interpretation.
- Do not infer image quality from completion rate alone.
5. Prompt-guide selection
Prompt collections may be curated from public Remix Camera packs around a specific reader intent, such as mirror selfies, dating photos, or professional portraits. The article should make the intended use clear and link to the relevant pack when it is the example source.
Inclusion in a guide means the example fits that guide's stated purpose; it is not a scientific score, safety certification, or guarantee that every model will reproduce the displayed result.
6. Reproducibility limits
Image generation is nondeterministic. A rerun can differ because of sampling, model or safety-system updates, provider routing, reference-image processing, or service errors. For that reason, we preserve the recorded result as a dated snapshot and favor a new documented run when a major model change makes an old result misleading.
See the method in practice
These published comparisons show their case sets, prompts, output grids, and recorded result labels.