How to Replace Chairs in a Photo with AI

This is the version I wish I had before I started

I started with a real photo of my dining area and thought the AI would “understand” what I meant. That was mistake number one. I gave it a slightly dark phone photo, grabbed a chair image from a product page, typed something vague like replace these chairs with these, and expected a clean preview. What I got was technically impressive and completely useless.

That was the first useful lesson. AI can do this well, but it does not fix bad input. It amplifies it. Once I accepted that, the whole workflow got simpler.

good result = clear room photo
            + clear chair reference
            + boring, specific prompt

bad result  = messy photo
            + random product image
            + "you know what I mean"

The room photo mistake I made first

I thought the chair reference was the important part. It was not. The room photo was the bigger problem. Mine was slightly cropped, one chair was hidden under the table, and the light was uneven. I could still see the room, so I assumed the AI could too. That was me giving it way too much credit.

When I tried again with a brighter photo where the table, chair backs, and chair legs were actually visible, the edit immediately got better. Not perfect. Just much less stupid.

Practical takeaway: before touching the prompt, fix the room photo. If the AI cannot clearly read the chairs you want removed, the swap starts broken.

Then I made the reference image mistake

My next bad idea was using a nice-looking chair photo that was actually a terrible reference. The chair was shot from a dramatic side angle in a styled showroom. It looked good to me as a shopper. It looked confusing to the AI.

Once I used a cleaner chair image, with the chair taking up most of the frame and the shape easy to read, the replacement got way more consistent. That was the moment it clicked: the model is not picking “the vibe.” It is trying to match visible form.

Practical takeaway: a clean reference beats a fancy reference. Boring wins here.

The prompt that finally worked

I wasted time trying long prompts because I assumed more words meant more control. They mostly made things worse. The prompt that worked was simple, direct, and a little bossy.

Replace the existing dining chairs with the chairs from the reference image.
Keep the table, floor, walls, lighting, rug, decor, and camera angle unchanged.
Match the new chairs to the room perspective and realistic scale.

That was enough. In Uniify, I uploaded the room photo, added the chair reference into the AI chat, and used that exact structure. If the first result drifted, I pasted the reference again and repeated the same instruction more clearly instead of inventing a whole new prompt.

That repetition felt dumb the first time I did it. It also helped.

What I expected versus what actually happened

I expected the first output to be final. It usually was not. It was more like a useful draft. One version got the shape right but made the chairs too small. Another nailed the size but changed the wood tone so the whole setup looked colder. Another one was strangely decent except for one backrest that bent like rubber.

That is normal. The first result tells you what the model understood and what it missed. So instead of judging it like a finished render, I started treating it like feedback.

The weird failures that kept repeating

The most annoying problem for me was when the AI replaced the chairs but also quietly redesigned other things. Sometimes the table edge changed. Sometimes the rug pattern softened. Once the room got brighter for no reason, like it had moved to another apartment.

That usually meant I had not said clearly enough what had to stay untouched. Another repeat mistake was assuming the AI would infer the chair count correctly. If six chairs are visible, say six chairs. If two should stay out of frame, say that too.

I also made one very human mistake: I kept changing five things at once. New photo, new prompt, new reference, different crop, different wording. Then I had no clue which change actually helped. Once I started changing one variable at a time, the process stopped feeling random.

What actually matters in practice

The main insight never arrived as one grand revelation. It came in pieces. First, the room photo matters more than I expected. Then the reference image mattered more than the prompt. Then I realized the prompt still mattered, but only because it stopped the AI from wandering off.

So if I had to boil this down into a real-world workflow, it would be this:

Floating conclusion, because that is honestly how this feels: AI chair replacement is not hard, but it is picky. The better you show the room, the less you have to explain. The better you show the chair, the less the AI invents. Everything after that is mostly cleanup.