Image Model Comparisons: A Fair Test Before You Choose

This URL previously described an older image-model comparison. Its uncited accuracy percentages and definitive winner claims have been removed. It is now a reusable comparison procedure. Check the tool directory for current product availability.

Choose one fictional portrait or an owned product photo. Freeze the brief, aspect ratio and allowed edits. Run the same task on each available platform. Record the actual date, account tier, settings and generation count. Do not compare a selected best output from one platform with the first output from another.

Use a rubric for subject consistency, composition, text legibility, editability and suitability for your intended use. Label each score as an observed judgment, not a provider benchmark. Keep failures alongside successful examples.

If accurate lettering is essential, test a specific label. If identity preservation is essential, inspect the reference/result pair. A platform can suit one brief and fail another. Until you run the test, there is no measured winner on this page.

Start with the reference-photo lesson and save your brief. The current illustration is a cover, not comparative evidence.

• Image generation and editing guidance

This historical URL mentions legacy models. DALL·E 3 is a retired API model; this page should not be used as a current purchase recommendation. Select an image model actually supported in your account and record that selection. The exercise below is a comparison method, not an observed head-to-head result.

Use an authorized product photo or a fictional subject. For a text-only trial, give both platforms the same brief:

A plain matte blue ceramic mug with one C-shaped handle, no lettering,
on a pale neutral tabletop. Soft window light from the left.
Keep the whole mug visible in a square crop. No people or extra objects.

For a reference-image trial, keep the same authorized reference and distinguish supported reference controls from text-only input. Those are different experiments; combining their results into one score conceals an input advantage.

Before generating, write down these five checks: one handle, correct blue color, no unwanted lettering, whole object inside the crop, and light direction consistent with the brief. Then assess artistic appeal separately. A visually attractive two-handle mug fails the product specification.

Keep every attempted output, not only the best image. Record the prompt, model, settings, number of attempts, credits used where visible and any manual editing. Do not imply camera authenticity from camera-style prompt words. Do not claim that an image is ready for a listing until the final exported file has been inspected.

If repeated outputs alter a critical product feature, use a photo-editing workflow or conventional retouching. A campaign requires identity accuracy as well as atmosphere. Review current usage and privacy terms before uploading confidential material.

Current image workflow comparison · Reference subject/style lesson

OpenAI API deprecations records the DALL·E 3 API shutdown. Check the current supported model list before following a historical model-specific workflow.