Full-Body AI Fashion Shots: Use Ratios, Not Centimetres
Specifying height in centimetres produced a stocky figure with a 1.02 legs-to-torso ratio. Framing the prompt in internal ratios pushed legs to 1.19 and beyond.
OMNI LabsUpdated
Key takeaways
Stating physical height and inseam in centimetres fails because image models lack real-world metric scales.
A baseline prompt with centimetre values produced a stocky 1.02 legs-to-torso ratio and a waistband at 50.4% of height.
Defining anatomical proportions as internal ratios achieved waistband placements of 54.4%, 56.7%, and 56.0%.
Legs-to-torso ratios increased to 1.19, 1.31, and 1.27 when guided by a 7.5-8 heads canon.
Proportion definitions must sit at the very top of prompt text, paired with explicit visual failure criteria.
The failure of real-world centimetre prompts
A full-body prompt for one male roster model stated the height and inseam in centimetres (a 185 cm impression, an 87 cm inseam). The model ignored the numbers. The rejected frame measured the waistband at 50.4% of figure height and a legs-to-torso ratio of 1.02: legs and torso were the same length and the figure read stocky.
Diffusion models operate in a latent coordinate space bounded by canvas pixels, completely untethered from physical measuring tapes. Phrases like '185 cm tall' or '87 cm inseam' act merely as semantic tokens associated with generic male imagery, rather than instructions to scale the geometry of limbs.
The same prompt emphasised broad shoulders and a strong chest, which inflated the upper body and made the short-legged read worse.
When descriptive copy piles emphasis onto upper torso landmarks without anchoring pelvic height, the image generator expands the ribcage and shoulders downwards. The torso absorbs the majority of the vertical frame budget, leaving insufficient pixel height for the lower limbs and compressing the overall silhouette into an unnatural, heavy stance.
Replacing measurements with internal ratios
The fix states proportion as ratios of the figure itself: the waistband sits just above the halfway point, roughly 54-56% of height measured from the floor, and the legs are the longer half. We also asked for a head count of 7.5-8 heads (a tall fashion canon, not an everyday 6-head build).
Rather than assuming the model knows how tall a centimetre is, this syntax anchors every anatomical segment relative to the total height of the generated subject. By declaring that the waistband must sit above the mid-point of the total figure, the model is prevented from placing the belt line at or below the canvas median.
After the re-shoot with that block, the waistband measured 54.4%, 56.7% and 56.0% of height, and legs-to-torso measured 1.19, 1.31 and 1.27.
Specifying an explicit classical canon of 7.5-8 heads provides an immediate structural template. Everyday human anatomy averages roughly 6 heads in total height, which reads as compact and everyday under editorial fashion lighting. Elevating the head ratio elongates the lower extremities into an unmistakable runway posture, giving high-fashion garments the vertical canvas they require to drape naturally.
Evaluated via automated segmentation script on white backdrop plates.
Our health check is numeric, not an eyeball: find the figure extent from the non-white rows, find the waistband from a denim mask, print waist as a percentage of height and legs-to-torso. Healthy in our set is about 54-57% and at least 1.15.
Visual inspection of fashion stills is prone to subjective bias. Styling, high heels, and dark shadows can create an optical illusion of length that vanishes once the asset is composited into an advertising layout. Running an automated mask measurement validates that pixels match editorial specifications prior to retouching.
By measuring mask boundaries across rows, the QA script flags any render where the waistband dips toward 50.4% before artists invest time in color grading or garment retouching.
Prompt placement and negative failure criteria
The proportion block sits at the TOP of every full-body prompt, before pose and before wardrobe. Our failure test is written into the prompt: if the torso looks as long as or longer than the legs, or the waistband sits at or below the halfway point, the image is wrong.
Prompt hierarchy directly dictates model attention during early denoising passes. If wardrobe details like intricate tailored seams or heavy wool fabrics lead the prompt, the model prioritizes textile rendering before establishing global anatomical layout. Placing the proportion block first forces the model to construct the skeletal armature before draping garments.
Embedding the negative test criteria directly into the prompt text provides strong conditioning against the model's natural bias toward average proportions. It informs the attention layers that an equal limb split represents an explicit rejection state, guiding diffusion trajectories away from stocky geometry.
Relative proportions for accessory and product scale
The same principle covers scale. A product described by a real-world comparison (a van, a car) only holds at normal scale. Describing a handbag as a fraction of the body (height about one sixth of shoulder-to-ground, reaching mid-thigh) held at giant scale and at normal scale.
When commercial stills require synthetic models to interact with hero luxury products—such as leather luggage, eyewear, or footwear—external metric benchmarks collapse. Telling an image generator that a handbag or accessory has a specific centimetre measurement frequently results in objects that dwarf the talent or shrink to tiny props.
Describing accessories as fractional segments of the talent's body creates an immutable geometric relationship. The bag maintains its intended physical presence whether the frame is a tight three-quarter crop, a full-body studio plate, or a stylized billboard composition, ensuring consistent product sizing across every editorial spread.
Scope, validation limits, and transferability
We have run the ratio block on one roster model only. We have not measured it on other body types, so treat the 1.15 threshold as our number for this rig, not a general rule.
Production teams must avoid over-generalizing empirical observations from a single controlled dataset. The 1.15 legs-to-torso threshold and 54-57% waistband placement delivered consistent results for our specific tall male talent rig, but different character archetypes may require tailored calibrations.
A muscular athletic build or a petite female model will naturally exhibit different skeletal baselines. What transfers across all model architectures is the underlying methodology: abandon physical metric units, formulate anatomy as self-referential fractions, and verify outputs with automated pixel masks.
Establishing studio-wide proportion benchmarks based on measurable pixel ratios gives fashion brands predictable character consistency across entire seasonal campaigns.
In practical production pipelines, locking down proportion rules early saves hundreds of hours of downstream correction. When generative models are constrained by mathematically explicit anatomical boundaries, the resulting character plates integrate seamlessly across complex multi-shot fashion narratives.
