The Half-Inch Problem: Six Image Models vs. the Ground Plane
Photo: VKeem, CC BY-SA 4.0, via Wikimedia Commons
A contact shadow is the cheap tell. Not the expensive tell — that is usually anatomy, text, or a window that bends in the wrong direction. The cheap tell is the half-inch gap under a coffee mug. The image looks polished, the lighting is gorgeous, the material is almost right, and then your brain notices that the object is levitating just enough to feel wrong.
At Hedra AI, we have been treating grounding as a separate capability from image quality. For this experiment, we gave six image models the same commercial product-shot brief and scored one thing: whether the object looked like it was actually sitting on the table.
The brief: one table, three objects, no drama
We used a deliberately boring setup. A matte ceramic mug, a brushed brass desk lamp, and a paperback book on a white oak table. Soft window light from camera left, 50mm lens, shallow but not silly depth of field. The prompt asked for a clean e-commerce hero image with realistic contact shadows and a continuous ground plane. No floating objects. No dramatic fog. No impossible reflections.
We generated ten images per model across the same seed range where the API allowed it. Six models, sixty images, one scoring sheet. We rated each image from 1 to 5 on three things:
- Contact: Does the shadow begin at the point where the object meets the surface?
- Cast: Does the shadow direction and softness match the light source?
- Ground: Does the table read as one continuous plane instead of a backdrop with a sticker on it?
This is not a definitive benchmark. It is a field test with a narrow brief, which is exactly how a lot of commercial work actually starts. A creator gets a product, a surface, and a deadline.
What broke, and what it looked like
The failures clustered into six familiar shapes. Model A handled simple contact well but over-darkened the cast shadow, making the mug look glued down. Model B produced lovely soft light and then forgot the shadow entirely under the lamp base. Model C was strong on the book but drifted the table grain between the foreground and background, so the ground plane felt like two photos stitched together. Model D invented a second light source for the shadow. Model E rendered a perfect shadow blob that was detached from the object by a few pixels — the half-inch problem in miniature. Model F was the most interesting: it got the contact right but made the brass lamp look like it was pressing into the table, with an ambient occlusion halo that was far too strong.
| Model | Contact | Cast | Ground | Read |
|---|---|---|---|---|
| A | 4 | 3 | 4 | Solid, slightly heavy |
| B | 2 | 3 | 4 | Floats under lamp |
| C | 3 | 4 | 2 | Table grain drift |
| D | 3 | 2 | 3 | Wrong light direction |
| E | 2 | 3 | 3 | Detached shadow blob |
| F | 4 | 4 | 3 | Crushed contact halo |
Scores are averages rounded to the nearest whole number. The useful part is not the ranking. The useful part is that each failure has a different fix.
The prompt is not a physics engine
Adding the words contact shadow to a prompt helps, but only up to a point. What helped more was describing the physical relationship in plain language: the object sits flush on the surface, the shadow begins at the point of contact, the light comes from camera left, the table is one continuous matte plane. Negative prompts for floating, gap, cutout, and sticker helped in some models and did nothing in others.
A better prompt pattern for product work looks like this:
Product shot of a matte ceramic mug on a white oak table, soft window light from camera left, realistic contact shadow under the base, cast shadow falling right, continuous table surface, no floating, no cutout, 50mm product photography
That is not magic. It is a constraint list. Models respond to constraints when the constraints are concrete and visual.
Fix it in the pipeline before you fix it in Photoshop
For commercial work in 2026, the smart move is to treat grounding as a pipeline stage, not a prompt wish. Generate the hero image, then run a second pass specifically for the ground contact. Inpainting a small strip under the object often works better than regenerating the whole scene, because you are not asking the model to re-decide the composition.
If you have a 3D or product-render workflow, use a shadow catcher. Render a clean contact shadow pass and composite it under the generated object. That sounds like more work, and sometimes it is, but it is predictable work. Predictable beats rerolling a seed twenty times while a client waits.
For pure retouching, keep three layers: contact shadow, cast shadow, and surface texture. The contact shadow should be tight and dark near the base, then fall off quickly. The cast shadow should match the light source: softer and longer for a big window, sharper for a small hard light. Do not blur both the same amount. Do not add a drop shadow with a fixed offset and call it a day.
One more practical check: zoom out. Contact shadow errors are much easier to see at thumbnail size than at 200 percent. If the object looks like it is hovering in a small preview, it will look worse on a product page.
The takeaway for creators
Image models are now good enough that the object itself is rarely the problem. The problem is the relationship between the object and the world. Grounding is where a lot of generated commercial work still gives itself away, and it is also one of the easiest things to audit. Build a contact-shadow check into your review process. Score the shadow before you score the style. The half-inch problem is small, but it is the kind of small that kills a sale.
We will keep running this brief as models update. If your pipeline has a favorite shadow fix, we want to test it.