The Honest Workflow: AI Compositing in a Working Product Photography Studio

Last week, we had a Glow Recipe avocado cleanser shot we were proud of and a Tata Harper reference image we wanted to match for a portfolio composite. Strong product photography, clean retouching, a clear visual direction. The plan was straightforward: use AI to generate an empty editorial set, drop the product in, composite cleanly.

Four Midjourney attempts later, the AI still refused to generate an empty set. Every output included a hero product the model had invented — wrong shape, wrong label, completely unusable. We had to stop and reconsider the approach entirely.

That moment is where this article actually starts.

AI compositing is being talked about everywhere right now, mostly by people demonstrating perfect outputs in controlled examples. What you don't see as often is what happens when you try to use these tools on real client work. Where they break. Where they bias. Where you have to push back on what the model wants to give you and where you have to accept its limitations.

This is a working studio's account of three composites — what worked, what failed, and what we learned about integrating AI into commercial product photography. Not a tutorial. Not a hype piece. The honest version.

Composite One: Avocado Ceramide Cleanser

The Glow Recipe Avocado Cleanser was already shot and retouched — clean studio lighting, proper color, isolated against the original surface. The job was to move it into a more editorial environment without reshooting. We picked a Tata Harper avocado-and-cucumber set as the visual reference: warm cream-to-sage gradient backdrop, halved avocado prop, soft directional light. Thoughtful styling, on-brand for the cleanser.

After Midjourney refused to generate an empty plate (the model kept inserting AI-rendered hero products into the scene), we switched to Photoshop's Generative Fill with Nano Banana. Different tool, different result — Nano Banana handled the empty-set request cleanly and gave us a usable plate on the second generation.

The cutout work happened in the source TIFF. A flattened composite layer feeding Select Subject, refined in Select and Mask, exported as a Smart Object. Standard professional cutout workflow — nothing AI-driven about this step. The product needed to be the real product, not an AI interpretation of it.

The interesting moment came at the integration step. We applied Adobe's Harmonize feature to match the product's color and tone to the new scene. Harmonize did exactly what it was designed to do — it matched our product's color to the AI plate's color palette. The problem: it also flattened the saturation that made the product visually pop. We got better shadows, but a less attractive bottle.

The solution was a selective compositing technique — keeping the harmonized layer's shadow work, masking out everything else, and revealing the original saturated product through the mask. Best of both worlds. The original Glow Recipe color, the harmonized contact shadow.

The takeaway from this shot: AI tools rarely give you the perfect answer in one click. The skill is knowing which parts of an AI output to keep and which to override. Harmonize gave us 50% of the answer. Selective masking gave us the other 50%. Neither tool nor technique alone would have produced the final image.

Composite Two: Watermelon Glow Regimen

The second composite was a flat lay of two Glow Recipe Watermelon Glow products. Different challenge, different lessons.

The original studio shot was workable but uninspired — the products were on a two-tone paper backdrop that fought the products visually. Lighting was clean. The retouching was already done. We wanted to keep the products and replace the surface with a more editorial watermelon-themed scene.

Flat lay compositing is harder than 3/4 angle compositing for one specific reason: every shadow is exposed in plain view. In a hero product shot, the surface horizon hides much of the shadow geometry. In flat lay, the products throw shadows across the surface where the viewer can see them clearly. If those shadows don't match between the products and the AI-generated set, the composite breaks visibly.

This is where the friction started.

We generated the watermelon plate in Nano Banana with explicit lighting direction in the prompt — light coming from upper-right, shadows falling lower-left to match the products' existing lighting. Generation one came back with light from upper-left. Generation two, same problem. Generation three, same. The model had a strong training bias toward upper-left lighting in flat lay imagery — likely because the dominant flat lay aesthetic on the platforms it was trained on uses that direction.

Three failed generations is the limit before you stop fighting the tool. We picked the strongest composition from the failures and flipped the canvas horizontally — one click in Photoshop reversed the entire scene's lighting direction to match the products. The watermelon and mint props mirrored, but watermelon is symmetric enough that no viewer would notice.

The cutouts were trickier than the avocado shoot. Two products instead of one, both with semi-translucent plastic bodies that introduced color contamination at the edges. We worked through them sequentially, used Decontaminate Colors at 50% in Select and Mask, tested the cutouts against magenta backgrounds before committing — magenta is the worst-case background for spotting fringing, so if the edges look clean against magenta, they'll look clean anywhere.

The takeaway from this shot: AI tools have biases baked into their training. Working with them productively means recognizing when to push back and when to accept the constraint. We pushed back three times. The fourth time we accepted the bias and worked around it. That's the right ratio — persistent enough to get the right answer when the tool can give it, pragmatic enough to stop fighting when the tool can't.

Composite Three: Strawberry BHA Blur Drops

The third composite was a more focused job. The product photo was already strong — a vivid red Strawberry BHA Blur Drops bottle on a clean white background. The problem was the prop. The halved strawberry next to the bottle had dried out during the shoot and looked tired in the final retouch. Muted color, wilted leaves, slightly cracked surface. It pulled the whole image down.

We didn't need to rebuild a scene. We just needed to replace the strawberry. The product is well-photographed and retouched. The strawberry, less so.

The first instinct was to use the same tool that worked for the previous composites — Photoshop's Generative Fill with Nano Banana. We made a generous selection around the strawberry, included enough surrounding white surface for the model to read the lighting context, and prompted for a fresh halved strawberry matching the scene.

Nano Banana failed at this task. Multiple generations came back with color casts, edges that didn't blend, and a strawberry that looked pasted into the scene rather than photographed in it.

The fix was a different model. Adobe's Firefly handled the same selection and similar prompt cleanly on the first attempt. The strawberry it generated had appropriate lighting direction, matching shadow density, color temperature consistent with the rest of the scene, and edges that blended invisibly into the surrounding white surface.

This wasn't a one-off. After running both tools across multiple tasks, a clear pattern emerged: different generative models have different strengths. Firefly is consistently better at integrated edits within existing photographs. Nano Banana is consistently better at generating whole new scenes from scratch.

The takeaway from this shot: knowing which tool to reach for saves regeneration cycles. Generative AI in production work isn't about finding the one best model. It's about building a working knowledge of where each tool wins.

The Bigger Picture

After three composites and several days of working with these tools on real client-grade material, a few things became clear.

AI compositing is not a replacement for product photography. It's a replacement for set building. The product still has to be photographed properly — clean lighting, consistent direction, accurate color, proper retouching. Every weakness in the source product photograph carries through to the final composite, often amplified. AI can build a beautiful set around a strong product. It cannot rescue a weak one.

AI compositing also doesn't reduce the skill requirements of commercial production. If anything, it raises them. Every step that used to be "shoot it on set" now becomes "shoot the product, generate the set, integrate the two believably." The integration step is where craft happens — selection quality, shadow logic, color matching, edge handling. These are the same skills senior retouchers have always relied on. The tools are new. The judgment isn't.

What AI compositing genuinely changes is delivery speed and variation flexibility. A product shot in a studio with consistent house lighting can be composited into multiple scenes for multiple campaigns without re-shooting. A client briefing five social variations no longer requires five physical set builds. This is a real efficiency gain, and it's already changing how brand campaigns get produced at the volume end of the market.

For studios that integrate these tools thoughtfully, the math improves on every dimension that matters — turnaround, variation count, cost per variation, client iteration speed. For studios that don't, the same brand work will increasingly go to studios that do.

The framing isn't "AI versus photographer." It's "photographer using AI as a tool." The studios that figure this out will deliver better work, faster, for clients who need it. The craft hasn't changed. The toolkit has.

At Estudio Forja, AI compositing is now part of how we work. Not because it's trendy. Because it lets us deliver work that competes globally, from a studio in Querétaro, for brands anywhere.