Common AI Image Mistakes

For identity drift, first inspect the reference, subject count, face size, camera angle, and competing style instructions. Reduce generic beauty or celebrity-like language. Keep stable identity traits separate from pose and expression, and avoid demanding an exact copy of a source angle in a different scene.

For a product, replace vague quality terms with geometry, material, edge, proportion, and branding constraints. Verify that the prompt is not asking for two incompatible versions of the same object.

Simplify the action. Specify the visible hand, what it touches, and the camera crop. Overlapping fingers, crossed arms, small full-body subjects, mirrors, and multi-person contact are harder than a simple pose. Establish the scene with a neutral pose before reintroducing the interaction.

If a hand is not important, compose it naturally outside the focal area rather than filling the prompt with hand-related quality terms. If it is important, make it large enough in frame to evaluate.

Check whether framing, camera distance, lens language, aspect ratio, and required objects agree. “Close portrait” conflicts with “full body and visible shoes.” Rank the priority or split the request into separate images.

For collages, define panel count, order, shot type, and consistency rules. Generate the panels separately if the model cannot maintain identity and wardrobe across a complex single-pass layout.

Name the light source, direction, softness, and relationship to the environment. A realistic window scene needs believable shadow direction and falloff; a night scene needs motivated practical or ambient light. Avoid stacking unrelated styles such as soft natural documentary light, hard studio flash, and volumetric sunset in the same frame unless the scene explains them.

Remove finishing terms until the basic light works. Add grain, color grading, bloom, sharpness, and resolution language after the scene has believable geometry and exposure.

If unwanted text appears, explicitly exclude typography, watermarks, signatures, and logos. If exact text is essential, generate the visual without text and add typography in a design tool unless the selected model reliably supports text rendering.

When a prompt becomes unstable, remove duplicated constraints and group synonyms. More tokens are not automatically more control. Preserve the decisive facts, run a baseline, and add back one block at a time.

• Name one visible failure.
• Locate the conflicting or missing instruction.
• Make one controlled revision.
• Compare against the baseline at the same framing.
• Keep the change only if it improves the intended criterion.