Why Does Your Falling Coffee Bean Ad Look Fake?

Roasted arabica coffee beans suspended in vertical freefall along an explicit downward gravity axis, studio product photography with controlled motion separation. Foreground plane: three beans at maximum sharpness, visible oil sheen from medium-dark roast, central crease showing authentic expansion with micro-pore texture catching rim light, individual chaff fragments visible. Mid-ground plane: six beans with directional motion blur trailing vertically, slight rotation, form still recognizable, crease still readable. Background plane: soft bokeh orbs with warm amber falloff, no defined silhouettes. Lighting: 120cm octagonal softbox overhead at 45° providing wraparound fill at 80% intensity, bare reflector dish from camera-right 2 o'clock creating hard specular highlights on oil sheen, gold card bounce from below at 20% intensity adding warmth to shadow areas. Seamless backdrop: vertical gradient from #C4A77D (warm tan) at bottom to #F5E6D3 (pale cream) at top. Scattered chaff particles and fine coffee dust suspended across three depth planes, atmospheric haze increasing with distance, gradient still readable. 8K texture, Phase One IQ4 150MP, Schneider Kreuznach 120mm macro at f/5.6, focus stacking aesthetic, commercial advertising, zero chromatic aberration --ar 9:16 --style raw --s 250 --q 2

Adding motion blur to sell speed is the fastest way to make a falling coffee bean ad look fabricated. The instinct is logical: a subject in freefall should streak. Diffusion models, however, do not resolve blur as velocity. They resolve it as lost focus, and lost focus inside a product frame reads as a failed capture rather than a fast subject. The image stops looking impossible and starts looking incompetent.

The repair begins with what "vertical freefall" actually specifies. The phrase performs three jobs at once: it fixes a gravity axis, removes every supporting surface, and declares a plane of maximum sharpness. Miss any one of those and the model falls back on its strongest prior, which for coffee beans is a horizontal arrangement on a table or inside a sack. Naming the axis explicitly is what forces the vertical composition.

The central crease is the authenticity anchor. Roasted beans in the training data are defined by that fissure, the geometry that separates an intact bean from a broken one. When the prompt asks for "authentic expansion" and "micro-pore texture," it is not decorating the bean; it is handing the model a high-frequency feature to render at full resolution. Remove it and the surface becomes a smooth brown ellipsoid, which is exactly what a synthetic render looks like.

Oil sheen is a specular event, not a material property. Medium-dark roast produces surface oils, but the model only draws them where a light source has a defined position and a defined hardness. Without a positioned highlight, sheen spreads into a uniform gloss that flattens the bean's curvature. That is why the lighting is described as a rig with intensity percentages rather than as a mood.

What the Model Actually Does With Motion Blur

Diffusion models have absorbed two distinct distributions. Sharp, high-frequency texture clusters with studio product photography. Soft, low-frequency surfaces cluster with defocus, compression, and amateur snapshots. Blur therefore does not signal motion to the model; it signals a lower-quality image. The only blur that survives this bias is directional blur that leaves the underlying form intact, so the bean must still be recognizably a bean.

That constraint sets the real threshold. Mid-ground beans can trail vertically, rotate slightly, and lose edge definition, but the moment the crease dissolves into a smear the frame loses its reference object. Viewers cannot measure speed against a shape they cannot identify. Motion blur is only legible when anchored to something sharp, which is why the foreground plane has to stay at maximum sharpness.

The same principle governs texture-stacking work elsewhere in commercial rendering. The hyper-realistic collage breakdown shows how layered high-frequency detail survives compositing only when one plane stays fully resolved. Freefall product work is that same problem with gravity attached. Blur is a layering decision, not a speed setting.

The Lighting Stack and Why Percentages Matter

A 120cm octagonal softbox overhead at 45 degrees and 80% intensity does one job: it wraps. The source is large relative to the bean, so falloff across the bean's curve is gradual and the shadow terminator stays soft. That gradual falloff is what makes a small object read as three-dimensional instead of as a flat cutout pasted onto a gradient.

The bare reflector dish from camera-right at two o'clock does the opposite job. It is small and hard, so it produces a tight specular highlight with a fast edge. The interaction between an 80% soft source and a hard point highlight is the entire oil-sheen read. Soften both and the bean looks waxy; harden both and the crease fills with shadow and the texture vanishes.

The gold card from below at 20% is the correction. It lifts shadow warmth without raising overall exposure, which preserves the gradient backdrop. Push it to 40% and the underside of every bean turns amber, the roast reads lighter than it is, and the color story collapses into a single tone.

Depth Planes, Haze, and the Milky Failure

Three planes are specified, and the background instruction is the one most often softened by mistake. "Soft bokeh beans" is not the same as "bokeh orbs with no definition." The first asks the model to render bean shapes at low contrast, which reintroduces recognizable edges at depth and destroys the sense of distance. The second asks for light, and light has no silhouette.

Atmospheric haze is a depth cue, not a filter. When haze, dust, and bokeh are all requested at high strength they compound into a milky wash that erases the #C4A77D to #F5E6D3 gradient and turns the backdrop into flat beige. Haze should increase with distance and stay below the mid-ground so the gradient remains readable behind the falling subject.

Chaff fragments and fine coffee dust are the cheapest realism signals available, because they are small, irregular, and high-frequency, the exact band diffusion models render most convincingly. Scattered across several depth planes, they also supply parallax cues that a single blurred background cannot provide.

Where the Recipe Breaks

The most counterintuitive failure is depth of field. A 120mm macro at f/5.6 focused on a bean at close range yields a depth of field measured in millimeters, so three foreground beans cannot all be sharp in one physical exposure. Naming "focus stacking aesthetic" resolves this: it tells the model to render impossible simultaneous sharpness rather than simulate a real lens, which is precisely what advertising photography does in post.

Aspect ratio is the second edge case. At 9:16 the gravity axis aligns with the frame's long side, so vertical blur trails read as downward motion. The same prompt in a square or landscape frame produces lateral drift and the beans appear to slide sideways rather than fall. Vertical formats are not cosmetic here; they reinforce the physics the prompt describes.

Parameter interactions close the loop. On Midjourney, --style raw reduces the aesthetic filtering that pushes commercial frames toward illustration, while a moderate --s value keeps stylization from overriding the texture specification. The specular control in the golden skeleton prompt is a useful comparison: metal and coffee oil fail the same way when highlight placement is left to chance.

None of this is about adding more words to the prompt. It is about removing ambiguity from the four variables the model cannot guess: the gravity axis, the plane of sharpness, the position and hardness of each light, and the amount of definition permitted at depth. Everything else is texture, and texture is the part models already handle well.

The single principle underneath every fix is that the model renders the physics you specify and invents the physics you omit. Freefall, blur, specular sheen, and depth are all physical claims. State each one as a measurable instruction — axis, plane, position, intensity, definition — and nothing in the frame is left for the model to guess.