Skip to main content

Common Image Mistakes & Fixes

Most image-edit failures are prompt problems, not model problems. Here are the mistakes that cause the most frustration, and how to fix them.

Classic before & after fixes

  • ❌ "Transform the person into a Viking" → ✅ "Change the clothes to a Viking warrior while preserving facial features"

  • ❌ "Put him on a beach" → ✅ "Change the background to a beach, keeping the person in the exact same position, scale, and pose"

  • ❌ "Make it a sketch" → ✅ "Convert to pencil sketch with natural graphite lines, cross-hatching, and visible paper texture"

The most common failure modes

  • Identity drift — the face, body, or product quietly changes even though you only asked for one edit. Fix: name the invariant first ("preserve the same facial features, expression, and pose") and change only one thing per pass.

  • Overloading one prompt — asking for a background swap, lighting change, and outfit change at once makes the model compromise on all of them. Fix: chain single, sequential edits instead of one ambitious prompt.

  • Vague verbs — "make it better," "improve," or "transform" give the model too much freedom. Fix: use direct verbs like change, add, remove, replace, and describe the exact result.

  • Pronouns — "she," "it," or "this" force the model to guess the subject. Fix: name the subject directly ("the woman with short black hair," "the red car").

  • Negative framing — describing what not to do (e.g., "don't make the arm too long") can confuse the model. Fix: use positive and literal instructions like "normal-length left arm, hand at hip, head-to-body ratio 1:7."

If the result drifts, don't re-roll

When an edit drifts, re-running the same prompt rarely helps. Instead, name the specific error in a follow-up prompt and add it to the preserve list for the next iteration — for example, "the face changed; keep the original facial features unchanged." Converging over 2–4 small passes is more reliable than re-rolling. Additionally, using strong reference images via the multi-edit feature can help preserve character identity across edits by providing the system with clear visual guidance. If the platform repeatedly returns the same render, perform a hard refresh of the page and start a new edit from the original source image to break the loop.

Did this answer your question?