Photoroom will cut your product out of a cluttered desk shot, drop it onto a brand-colored studio background, and render a shadow underneath it in about four seconds. The composite that comes back looks ninety percent finished, and the ninety percent is real and genuinely useful. The trouble is the last ten percent, and the last ten percent is almost always the shadow. AI compositing has gotten extraordinary at the cutout and the background and weirdly, persistently bad at making light behave. This lesson takes one product hero from a phone snapshot to a publishable launch image, shows you exactly where Photoroom's AI shadowing breaks the laws of physics, teaches you to fix it by hand, and leaves you with the finished image plus a hand-finishing log that documents every correction so the next composite is faster and your provenance is clean.

The Composite That Was Almost Right

Here is the scene. Marketing needs a hero image for a launch page by end of day: a single product, currently photographed on a cluttered studio table with cables and a coffee cup in frame, sitting cleanly on the brand's signature deep-blue background, lit like a studio shot. You open Photoroom, upload the snapshot, and watch it cut the product out with startling precision, including the tricky transparent edge and the thin cable that loops behind it. You pick the deep-blue background, click the AI shadow option, and four seconds later you have a hero image that looks like it cost a photographer a day.

Then you actually look at it. The product is lit from the upper left, you can see it in the highlights along its top edge, but the shadow Photoroom generated falls straight down and slightly to the left, as if the light were directly overhead. The shadow is also uniformly soft from the contact point outward, with no darkening where the product meets the surface, so the product appears to hover a few millimeters above the background like a sticker. And the shadow's opacity is the same at the far edge as at the base, which never happens with real light. None of this is obvious at a glance. All of it registers, subliminally, as "this image is fake," which is the exact opposite of what a launch hero is supposed to do.

This is the pattern. The cutout is excellent, the background is fine, and the shadow is a small, fixable lie that undermines the whole image. Your job is to find the lie and correct it, and that is a craft skill no AI shadow has yet replaced.

Why AI Shadowing Fails the Way It Does

Understanding why the shadow breaks tells you exactly what to look for. A real shadow is a physical consequence of a specific light source hitting a specific three-dimensional object and being occluded onto a specific surface. It encodes the light's direction, distance, and softness, and the object's geometry and contact with the ground. The AI does not model any of that. It has learned, from millions of images, what shadows tend to look like, and it produces a plausible average shadow that is not derived from the actual light in your scene because it does not know the actual light in your scene.

This is the same generation-versus-understanding gap that runs through this whole program, applied to photons. The model generates what a shadow looks like; it does not understand where the light is. So it gets the easy, high-frequency parts right (there is a soft dark shape below the object) and the specific, physically-determined parts wrong (the direction that matches this object's highlights, the contact darkening, the falloff). The failures are not random. They cluster in four predictable places, and once you know the four, you can audit any AI shadow in fifteen seconds.

The Four Places the Shadow Lies

1. Direction mismatch. The shadow points the wrong way relative to the object's own highlights. If the product is lit from the upper left, the shadow must fall to the lower right; an AI shadow frequently falls straight down or in a direction inconsistent with the lighting baked into the product itself. This is the most common and most damaging error because the eye reads light direction unconsciously and flags the contradiction instantly.
2. Missing contact shadow. Where an object touches a surface, there is a small, dark, sharp region of near-total occlusion - the contact shadow - that anchors the object to the ground. AI shadows routinely omit it or render it too soft, which is what produces the "floating sticker" effect where the product hovers above the background.
3. Wrong softness. Shadow edges are sharp near the contact point and soften with distance from the object, because the light source has a finite size and the penumbra grows with distance. AI shadows are often uniformly soft or uniformly sharp, ignoring this gradient, which reads as unreal even when you cannot name why.
4. Flat opacity and falloff. A real shadow is darkest at the base and fades with distance; AI shadows frequently have constant opacity across their whole length, so the shadow looks painted on rather than cast.

AI compositing has solved the hard-looking problem, the cutout, and left the easy-looking problem, the shadow, mostly unsolved. The shadow is where every composite betrays itself, and fixing it by hand is the difference between a launch hero and a sticker.

The Compositing Workflow, End to End

Now the workflow. The point is not to avoid Photoroom; the point is to let it do the ninety percent it is great at and to own the ten percent it is not. Run it in five stages.

Stage one, the cutout. Upload the snapshot and let Photoroom cut the product out. This is genuinely the part to trust. Check the edges at full zoom anyway, especially transparent or fine elements (cables, glass, hair), and refine the mask if the cutout grabbed background or dropped a thin element. Usually it is close to perfect.

Stage two, the background. Place the cutout on the brand background. Keep the product at the scale and position the composition needs, and leave room below and to the side for the shadow you are about to fix, because a product jammed into the corner has nowhere for a correct shadow to fall.

Stage three, establish the light. Before you touch the shadow, decide where the light is, and read it off the product itself. Look at the highlights and existing shading on the cutout and determine the light direction that is already baked into the photograph. This is the ground truth the AI ignored, and everything in the shadow must obey it. Write the light direction down (for example, "key light upper-left, roughly 45 degrees"); it becomes the rule the rest of the work follows.

Stage four, fix the shadow. Generate or accept the AI shadow as a starting layer, then correct it against the four failures and the light direction you established. Rotate or skew the shadow so it falls in the direction the light demands. Add a contact shadow: a small, darker, sharper region right where the product meets the surface, which is the single highest-impact fix and the one that kills the floating effect. Introduce the softness gradient, sharper at contact and softer with distance. And taper the opacity so the shadow is darkest at the base and fades out. Each of these maps directly to one of the four failures.

Stage five, the final integration pass. Step back and check the whole image. Does the product's lighting match the implied studio environment? Does the shadow now read as cast by the same light that lights the product? Add any final grade so the product and background share one color temperature, because a warm product on a cool background is its own tell. When the shadow obeys the light and the contact anchors the product, and the whole frame shares one color temperature, the composite stops looking composited and starts looking photographed.

Beyond the Shadow: The Other Physical Tells

The shadow is the most common and most damaging tell, but it is not the only place a composite betrays its assembly, and on certain products the others matter just as much. The deeper principle is the one running through this whole lesson: the AI cannot model the specific scene the subject is now in, so any physical cue that depends on that scene will be generically plausible and specifically wrong. Two cues besides the shadow deserve a place in your audit when the product warrants it.

The first is reflection. A reflective product - chrome, glass, glossy plastic - reflects its environment, and after compositing the reflections in its surface still show the original cluttered desk, not the new brand background. The AI does not repaint the reflections because it does not know what the new scene should reflect. On a matte product this is a non-issue; on a polished one it is as loud a tell as the shadow, and it requires hand-painting or masking the reflections to match the new environment and implied light. For highly reflective products the reflection work can exceed the shadow work, and the composite is genuinely harder than it first appeared.

The second is color temperature, which the final integration pass already addresses but which deserves naming as a physical cue in its own right. A subject photographed under warm tungsten light composited onto a cool studio background carries the wrong white balance, and the eye reads the mismatch as two scenes spliced together even when shadow and reflection are correct. The fix is a grade that brings the subject and background to one color temperature, completing the single-scene illusion. Shadow, reflection, color temperature: three cues, one underlying cause, and the same correction strategy each time - establish the real scene, then make the cue obey it.

Consistency Across a Set of Composites

One product hero is a single problem; a product line is a harder one, and it is worth flagging because the failure mode is the cousin of brand drift. If you composite ten products independently, each one finished correctly in isolation, the ten can still disagree about where the light is. Product one is lit from the left, product four from above, product nine from the right, each internally consistent, and laid out together on the catalog page they read as ten separate photo shoots rather than one coherent line.

The fix is to establish one light model for the whole set before you finish any of it: a single key-light direction, a single softness, a single contact-shadow density, and a single color temperature, written down as a set-level specification. Then every composite in the line is finished to that one specification, regardless of who works on it or in what order. The per-image shadow audit guarantees each composite is individually correct; the set-level light specification guarantees they are correct together. A line that shares one implied studio reads as a brand; a line of individually-perfect-but-mutually-inconsistent composites reads as a collection, which is exactly the trap the brand-drift lesson warns about, here rendered in light rather than palette.

The Hand-Finishing Log, the Artifact That Compounds

The named artifact of this lesson is not just the publishable image; it is the hand-finishing log that travels with it. The log is a short record of every correction you made on top of the AI output: what the AI got wrong, what you changed, and why. For this composite it might read: cutout accepted, refined cable edge at top-right; light established as upper-left 45 degrees from product highlights; shadow direction rotated 60 degrees to match; contact shadow added at base to remove float; softness gradient applied; opacity tapered from 70 percent at base to 0 at tip; final warm grade matched product to background.

The log does two jobs. First, it is provenance: it documents exactly where the AI ended and your hand began, which matters for client disclosure, for legal review, and for any later question about what was generated versus what was crafted. Second, it is a checklist that compounds. After three or four composites your log entries start to rhyme, the same four shadow fixes recur, and the log becomes a repeatable finishing procedure that turns the next product hero from a twenty-minute hunt into a five-minute pass. The log is how a one-off fix becomes a workflow.

Why the Shadow Is Where the Craft Lives Now

It is worth sitting with the irony. For years the hard, slow part of compositing was the cutout: masking hair, fine edges, transparency, the stuff that took a skilled retoucher hours. AI solved that almost completely, and it solved it first because cutouts are a high-frequency, well-defined pattern with abundant training data. The shadow, which a competent retoucher could fix in two minutes, turns out to be the part the AI cannot do, because a correct shadow requires modeling the physics of a specific scene, and the model only knows what shadows generally look like.

So the craft did not disappear; it relocated. The valuable skill is no longer the patience to mask a cable. It is the eye to read light off a photograph, the knowledge of how shadows actually behave, the awareness that reflective products betray themselves in a second physical cue, and the discipline to audit every AI composite against the failures before it ships. These are exactly the things that resist automation, because they require modeling the physics of a specific scene rather than reproducing what scenes generally look like, and a model that only knows the general cannot supply the specific. The designer who can do this turns Photoroom into a genuine accelerator, getting the ninety percent for free and supplying the ten percent that makes the difference, while the designer who cannot is at the mercy of whatever the model happened to render. The designer who ships the raw AI composite ships floating stickers and wonders why the launch page looks slightly off. The shadow is small. It is also the whole job.

Putting It to Work This Week

On your next product composite, do three things. Run Photoroom for the cutout and background and consciously trust it for that part; do not waste time re-masking what it nailed. Then, before accepting the shadow, read the light off the product and write the direction down, because that single act forces you to notice when the AI shadow contradicts it. Finally, run the four-failure audit on the shadow - direction, contact, softness, falloff - fix each one, and write the corrections into a hand-finishing log. Keep the log with the file. The image you ship will read as photographed rather than assembled, and the log you keep will make every composite after this one faster and your provenance defensible.

Key Takeaways

• Photoroom nails the cutout and background - the part that used to be hard and slow - and reliably fails on the shadow, the part a retoucher could fix in two minutes. The composite looks ninety percent done; the last ten percent is almost always the shadow.
• AI shadowing fails because the model generates what a shadow looks like without understanding where the light is. It produces a plausible average shadow not derived from the actual light baked into your photograph.
• The shadow lies in four predictable places: direction mismatch with the product's own highlights, missing or too-soft contact shadow (the floating-sticker effect), wrong softness (no sharp-to-soft gradient), and flat opacity with no falloff. Knowing the four lets you audit any composite in fifteen seconds.
• The workflow is five stages: trust the cutout, place on background with room for the shadow, establish the light by reading it off the product, fix the shadow against the four failures, and do a final integration and color-temperature pass.
• Establishing the light direction from the product's own highlights is the pivotal step; it is the ground truth the AI ignored, and everything in the shadow must obey it. The contact shadow is the single highest-impact fix for killing the floating effect.
• The hand-finishing log is the compounding artifact: it serves as provenance (where the AI ended and your hand began) and as a checklist that turns the recurring four-fix procedure into a five-minute pass on the next composite.