Every diffusion model drifts the face if you regenerate too many times. Eye colour shifts, cheekbones creep, the hair gets a touch wavier in shot 47. Across a 100-SKU catalog the brand ends up showing 100 different models. Here is the technical statement of the consistent-identity problem in fashion AI, and the production-grade solution Apiway ships.
Why identity drifts at all
Diffusion models are stochastic. A generation pass starts from random noise and walks through a denoising trajectory guided by the prompt and any reference inputs. Even with identical prompt, identical seed, and identical settings, small numerical differences in the trajectory produce small differences in the output face. Across many generations, those small differences compound.
For fashion catalog work, where the same model needs to appear across dozens or hundreds of shots, this drift is the difference between brand-consistent imagery and a catalog that looks like it was shot with a rotating cast.
Why “just use the same prompt” fails
The first instinct most teams have is to lock the prompt: write a detailed description of the model, save it as a template, paste it on every generation. This does not work, because:
- Natural-language prompts are too low-resolution to specify a face. “Brown hair, brown eyes, oval face, mid-twenties” describes thousands of possible faces, not one.
- Prompt-only identity is not seed-stable. The same prompt plus a different seed produces a different face.
- Even with a fixed seed, the model may shift on minor composition changes (different garment, different framing).
Three techniques that do work
We use a combination of three approaches at different layers of the stack.
1. Preset model identities. White Studio ships ~50 female and ~10 male preset AI fashion models. Each preset is a stable identity built on a combination of learned-embedding anchors and seed-stable generation parameters. Picking a preset gives you the same face across every generation, by construction.
2. Reference-image conditioning. If the user uploads a reference photo (a real model, a brand ambassador), the generation pass conditions on the reference at the early denoising steps. The face on the output stays close to the reference identity even when garments and framings change.
3. Real-creator anchoring (Hollywood pattern). For lifestyle imagery, the most reliable approach is to skip the AI face entirely. A creator photo set from the marketplace provides a real human face that, by definition, does not drift — it is just a real photograph. AI handles the garment overlay only.
The actual numbers on drift
For from-scratch generation with prompt-only identity locking, identity drift across 100 shots is significant — we routinely measure perceptual face-distance scores that are 5–10x higher than the natural variation across photographs of the same real person.
For preset-based generation in White Studio, the drift across 100 shots is roughly comparable to photograph-of-the-same-person variance. Indistinguishable to a casual viewer. Detectable to a careful one.
For creator-photo-set anchoring, drift is zero by construction (the face is a literal photograph reused across every overlay).
When to use each technique
- Preset model: catalog and PDP work where a clean AI fashion model is acceptable.
- Reference upload: brand-ambassador work, founder portraits as the consistent face, custom casting decisions.
- Creator photo set: lifestyle, ad creative, anywhere shopper-trust signals matter most. (Background: the Hollywood VFX approach.)
Operations pattern for 100+ SKU catalogs
Lock the model approach at the start of the catalog work. Document the choice (preset name, reference photo file, creator set ID) as a brand-system asset. Onboard producers explicitly — the model is a brand decision, not a per-SKU decision.
For seasonal refreshes, treat changing the model approach the same way you would treat changing the brand colour palette: a deliberate, communicated brand evolution, not a quiet drift.
Why this stays defensible
Identity consistency is one of the few areas in fashion AI where the technical solution actually depends on product-level choices, not just on whichever LLM is best this week. The combination of preset training, reference conditioning, and marketplace anchoring is a stack that does not get obsoleted by the next foundation-model release — it sits above the foundation model and controls how it is used.
Run a 50-shot drift test
Pick one White Studio preset and run 50 generations with different garments and framings. Lay them out as a grid. The face should be stable across the grid in a way that prompt-only locking cannot match. Free accounts ship with 100 one-time credits — enough to run the full test.