Editing a photo

Hi, I'm Mel. You already have the photograph. The lion is right; the enclosure is not. That sounds like image generation, but it is a different job. The test is whether any part of the original has to survive. If the answer is yes, edit the photograph. The useful instruction says not only what should change, but what the tool is not allowed to change.

This is the source. The lion's face, mane, body, pose, crop and camera angle are the parts I want to keep. The fence, wall and ground need to go. If I gave a generator only a description, it could make another convincing lion. That would not be this photograph. Before choosing a tool, draw the boundary between the material you own and the material you want invented.

Two different methods sit under the word edit. A selection-based tool cuts out or adjusts chosen pixels. A generative editor uses the photograph as a reference and rebuilds material around it. The first gives stronger preservation. The second can make a completely new setting feel natural. It also asks the model to interpret where one region ends and the next begins. That distinction explains most unexpected results.

In a design tool such as Canva, the common jobs appear as separate controls: remove a background, erase an object, generate a new background, extend the frame, blur or adjust the light. You choose the operation, then refine the area. That is efficient when the photograph is heading straight into a poster, slide or social design. Check the crown or upgrade label before building a workflow around any one control; access depends on the plan.

Background removal turns the lion into a transparent cutout. The face, body and pose remain the source pixels, which is strong preservation. It is not a finished replacement. Put that cutout on a new scene and it can still look pasted on if the edge, ground, shadow and colour temperature do not agree. A clean cutout protects the subject. The join still needs judgement.

In Gemini, the wording needed one correction. Replacing the enclosure did not clearly include the ground beneath the lion, so the model could preserve foreground that belonged to the zoo. The revised instruction named both: replace the enclosure and the ground under the lion with open grassland. Then it listed what to keep: the face, mane, body, pose, lighting, crop and camera angle. The rebuilt foreground and background now read as one environment.

The revised result is much better, but it did not obey every part of the instruction. The light became warmer, the overall colour moved, and fine detail may have been regenerated rather than copied. The prompt said to keep the lighting unchanged. It did not. A preservation list gives the model boundaries; it does not freeze the pixels. Compare the result with the source beside it, not with your memory of the source.

Write an editing instruction in two halves. First, name what changes, including the whole region around it. Second, name what stays: subject, pose, identity, product shape, framing, light or text. Make one bounded change at a time. Then check the boundary, the subject and the unintended changes. When a real person, product or place has to remain recognisable, the uploaded source is evidence. A written description is not.

Choose the route by the control you need. Use a general assistant such as ChatGPT or Gemini when describing the edit in conversation is enough. Use Canva when the edited image belongs inside a layout. Use Adobe Firefly for a dedicated AI editing workspace with prompts and selected regions. Use Photoshop when repeated commercial work needs layers, masks and versions. Whichever route you choose, editing a stock image does not make the source yours; its licence still travels with it.

Keep the original beside the result. Decide what is allowed to change, name what must stay, and inspect what came back. If the source must survive pixel-for-pixel, use a selection-based edit. If the setting has to be invented, generative editing can help. Treat every generated change as a proposal until you verify it.