Imagine a product image of someone holding a portable speaker. The lighting looks natural. The fingers wrap around the casing, and the surface has the right texture. Everything suggests a photograph of the product in use.
Then you check the specifications. The speaker is 15 centimeters tall, but the image makes it look nearly as long as the person's forearm. The scene is convincing until you ask a different question: is this how large the product would look in real life?
Scale is one of the AI image generation limitations that can survive an otherwise polished result. In this article, we explain why these errors happen and how teams should review an asset before it reaches publishing.
What is scale in an image?
Scale describes how large an object appears compared with its surroundings. Accurate scale means those size relationships make sense in the real world, taking perspective and distance from the camera into account.
Why image models get scale wrong
Recognizing a product does not establish its dimensions
Generative Image models learn patterns from visual examples. Those patterns help them generate a recognizable speaker, including its shape and surface details. They also learn patterns about how people hold objects.
That knowledge does not guarantee the right proportions when the model puts the two together. A product category can include many sizes. Even two speakers with almost identical designs can have different dimensions.
A reference photo provides appearance information, but a cropped product against a plain background may offer little evidence of physical size. Adding “15 centimeters tall” gives the model more information. It still has to translate that measurement into a believable relationship with the person's body.
It would be too broad to say models understand no geometry. Research has found that they learn useful spatial information. The problem is that this knowledge does not consistently produce images with correct geometry. A generated scene is not automatically a measured 3D construction with fixed dimensions and a specified camera.
The reference and generated scene use different camera geometry
A close-up photograph can make a small speaker fill the frame. That framing tells you little about how much space it should occupy in a wider lifestyle image.
When the generator places the product into a new scene, it must reconcile the source view with the new viewpoint. Its size must agree with its distance from the camera and the surface supporting it. Copying the product's appearance successfully does not settle those relationships.
For teams using AI for product photography, this means separating two questions during review: does the product look like the reference, and does it fit the generated scene at the right size?
Looking real does not mean size is right
A generation can meet the visual brief without meeting the dimensional requirement. The setting may be attractive and the product recognizable, leaving the scale error easy to overlook.
Training for convincing imagery does not, by itself, impose a physical measurement on every object. The relevant distinction is between learning what arrangements commonly look like and reliably preserving a particular product's dimensions within a new arrangement.
What current research shows
GenScale tests whether objects have plausible relative sizes
The September 2026 GenScale preprint examines relative object scale across image generation and editing. Its benchmark includes 900 image-level entries and 1,643 relationships between a reference object and a target object. One task specifically tests products shown with people.
In that task, models receive a product reference image and a prompt with metric dimensions. They must preserve the product's identity while showing it at a plausible size relative to a human body or body part.
Among the systems tested, GPT-Image-2 achieved a plausible-scale score on 78.9% of its 294 valid product-human images. Nano Banana 2 reached 75.6% on 295 valid images. Products were more often enlarged than reduced. The practical finding is that explicit size information did not eliminate errors.
“Plausible” has a specific meaning here. The study uses a five-point judgment scale, ranging from severely undersized to severely oversized. Its middle category allows an estimated size error of approximately 20% in either direction. An automated judge, calibrated against nine human raters, provides the main benchmark evaluations. These percentages therefore describe judged plausibility, not exact dimensional accuracy. Unscorable cases are excluded.
The study also examines corrections. General-purpose editors generally performed better when told which object to resize and by how much than when asked to discover the error themselves. Its proposed Rescale system improved scale plausibility through targeted editing, but correction still required evaluation.
Geometry errors extend beyond product size
The CVPR 2024 paper Shadows Don't Lie and Lines Can't Bend! examines whether generated images reproduce geometric properties found in real photographs. The researchers analyzed perspective information, detected lines, and relationships between objects and shadows.
They first selected generated images that passed a separate detector based on image appearance. Classifiers using only derived geometric information could still distinguish generated images from real ones. This supports the finding that convincing surfaces can coexist with inconsistent scene geometry.
The paper did not directly measure product scale, and its results concern the generators tested at that time. It provides supporting evidence for checking geometry separately from realism.
Why scale matters commercially
Product imagery helps shoppers imagine an item in their own lives. If the image changes its apparent size, it can change that judgment.
In product-page usability testing, 42% of users tried to determine a product's size from its images. The research also observed shoppers rejecting suitable products after misjudging their size. Images that showed products beside recognizable references helped communicate scale.
Consider the speaker again. Someone looking for an item to carry in a small bag may use the person's grip to estimate how portable it is. An oversized depiction could make the product seem unsuitable. An undersized depiction could create an expectation the real product cannot meet.
Written dimensions remain necessary, but they do not erase the impression created by an image.
Checks before publishing generated product assets
Use this checklist with the product specifications and a reliable visual reference beside the generated image. Review the version that will appear on the page, including any edits made after generation.
Compare the product with a known-size reference
Choose a reference whose dimensions you can establish. Ideally, it should sit at a similar distance from the camera as the product.
A hand offers familiar context, but hands vary in size. An AI-generated hand can also have its own proportion errors. Use it as a visual cue and compare the interaction with an accurate photograph of someone holding the real product.
Check apparent size against the listed dimensions
Confirm that the asset depicts the exact product variant. Similar-looking versions may have different measurements.
For the speaker, compare its stated height with a reliable image showing that model in use. Check width and depth as well. A product can have a convincing height while appearing too broad to hold comfortably.
Avoid treating a simple pixel ratio as a real-world measurement when the objects have different positions or orientations. If the view makes the comparison uncertain, use a clearer composition or a dimensionally controlled reference.
Check perspective and depth consistency
Look at the product in relation to the surface and background. Does it appear to occupy the position implied by the scene? Do its visible edges agree with the viewpoint?
For a speaker on a table, inspect how its base meets the tabletop. A mismatch can make the product seem pasted into the scene even when its texture looks photographic.
Remember that a closer object can legitimately appear larger. The goal is a coherent relationship between size and distance.
Inspect points of contact
Look closely at where the product touches a person or another object. Contact limits how much a depth difference can explain an unusual size relationship.
If someone holds the speaker, check whether their grip makes sense for its width. Fingers that appear to close around an implausibly large casing deserve another look.
Review the corrected image again for size and product identity. An edit can improve proportions while changing the design or disrupting contact with the hand. When you cannot establish an accurate relationship, replace the asset with a verified photograph or a controlled rendering.
How curated data can help models learn scale
Training examples can make size relationships clearer. A close-up of a speaker shows its design. Photos of someone holding that same speaker, paired with its dimensions, also show how large it should look in use.
Curated data for generative AI lets teams select examples that address specific weaknesses. If a model repeatedly makes handheld products too large, a collection of real product interactions can provide relevant examples of correct proportions. Including different viewpoints helps show how apparent size changes with distance and angle.
Teams can then test whether training on those examples improves scale accuracy on products and scenes the model has not seen. That connects data selection to a concrete goal: helping models show products at a size people can trust.
Sources
Li et al. (2026). GenScale: A Benchmark for Relative Object Scale in Image Generation and Editing. Preprint. Sections 3.2–3.3 and 5–6; Tables 5–6.
Sarkar et al. (2024). Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now. CVPR 2024. Sections 1 and 4–5.
Baymard Institute. Product Page UX: All Products Need at Least One “In Scale” Image. Verified October 12, 2026.
What are the limitations of text-to-image ai?
What are the limitations of ai-generated images?
What are the limitations of AI image generation?
What is the most important quality to check in AI generated product images before publishing?
Why do AI generated product images make products look too big or too small?
