Why are the feet and shoes inconsistent across my AI model images?

Separate who the model is from what the model wears

A face or portrait can help anchor facial identity, but it is not evidence for shoes, foot shape, stance or floor contact. A full-body image can provide visible body proportions and pose cues. Neither reference necessarily specifies the footwear product you want to show.

When a particular shoe is advertised, keep a clear product reference separate from the approved identity reference in the production record. Use the reference roles and number of inputs the actual tool supports; this is not a promise that every generator accepts separate controls. In CartVisual’s current image workflow, product and selected model references can inform an image. The video workflow uses the approved image as its first frame rather than directly accepting an independent model reference. Check toe shape, color and sole against the real shoe afterward; inputs guide generation rather than locking pixels.

Run a crop-to-ground diagnosis

Compare like with like before changing the prompt. A shoe can appear different because one frame hides the heel, another crops the toe, or the foot is viewed from a different angle. A reference ending at the ankles leaves those relationships unresolved.

Use this short sequence on one representative image pair:

1. Check the model reference crop. If it ends above the feet, mark stance, shoe shape and ground contact as unspecified.
2. Check the shoe reference separately. Confirm the intended left and right shoes, visible color and silhouette are clear.
3. Compare camera angle, distance and lower-frame crop. Match these before judging whether the shoe changed.
4. Inspect the floor boundary and shadows. Look for a plausible contact point under each visible shoe, not a floating gap or fused sole.
5. Only then revise the instruction or reference set, changing one cause at a time.

Use a controlled comparison

Keep the model identity reference and shoe product reference fixed while you compare outputs. Change only one variable, such as the crop or camera angle, so you can tell whether the issue follows missing evidence or the composition. This is a practical review method; it is not a documented promise of deterministic generation.

For example, in a fictional exercise, suppose a campaign set has three images: two show black low-top shoes, while a third shows a pale high-top. The model reference is cropped at the ankles, and only the first two frames show the shoe reference clearly. First align the framing, then provide a clear shoe reference for all three and inspect the ankle-to-ground area again. Do not infer that the third shoe is a product variation unless the product source confirms it.

Set a review threshold before continuing

Decide which details must match before generating more images. For a footwear campaign, that may include shoe category, dominant color, toe silhouette and visible contact with the ground. Mark hidden details as not assessable rather than guessing from a cropped frame.

Stop and replace or clarify the references when the required shoe cannot be identified, the feet are cut off in every review frame, or repeated comparisons still show conflicting footwear. Do not use an image as product evidence when its shoe details are obscured or inconsistent. A review checklist can flag uncertainty; it cannot prove fit, comfort or product performance.

Footwear consistency review

If a reference contains competing visual roles, use What should I check when one model sheet contains a face close-up and three full-body views? to separate what each input actually supports.