How to keep the same AI model across product images: a 60-SKU case study

This is the story of a 60-SKU lookbook drop that an Indian D2C apparel brand on Myntra shipped through Relive in May 2026, with the same generated model carried across every plate of every SKU. I will walk through what the problem actually looks like in production, why studio shoots solved it expensively for years, how persona-lock works under the hood, and the failure modes we hit and fixed during the drop.

The brand has asked to not be named on the page. The numbers below are real — I have just rounded a couple to the nearest hundred for cleanliness.

The problem — model consistency across 5 SKUs × 5 plates

The brief was straightforward on paper. Sixty SKUs, women's western wear, going live on Myntra and on the brand's Shopify store within a week. Five plates per SKU — front, back, detail, lifestyle, on-figure. So three hundred plates in total.

The hard constraint was that the same model had to appear across all five plates of any given SKU, and the brand wanted the same persona — same face, same body, same vibe — across the entire collection. Not just within an SKU. Across all sixty.

This is the constraint that makes lookbooks expensive. It is not the per-plate cost — three hundred plates at a basic studio rate is roughly ₹1.2 lakh. It is the model consistency.

You cannot ship three hundred plates from sixty different mannequin shoots and call it a lookbook. The brand identity needs a face. Customers shopping a 60-piece drop want to picture the same woman across the catalog so they can see how the brand styles. Myntra's algorithm also clusters listings by face embeddings — drift across the catalog hurts discoverability, as covered in the Myntra requirements piece.

So the constraint forces you to either book the same model for multiple days of shooting, or solve the consistency problem in software.

Why agency shoots solve it (and what they cost)

The traditional answer is a booked model day.

For sixty SKUs at five plates each, you are looking at two to three model days minimum, depending on how fast the studio operates and how many garment changes the model can manage in an hour. Premium model day rate in Mumbai in mid-2026 sits around ₹25,000 to ₹40,000 for a portfolio model, plus hair and makeup at roughly ₹12,000 per day, plus stylist at roughly ₹10,000 per day, plus the studio booking at roughly ₹15,000 to ₹20,000 per day.

A realistic agency line item for this catalog:

| Line item | Cost | |-----------|------| | Model fees (2 days, mid-tier) | ₹60,000 | | Hair and makeup (2 days) | ₹24,000 | | Stylist (2 days) | ₹20,000 | | Studio booking (2 days) | ₹35,000 | | Photographer + assistant | ₹40,000 | | Retouch (300 plates × ₹250) | ₹75,000 | | Logistics (samples, returns) | ₹15,000 | | Buffer (reshoots, garment fixes) | ₹25,000 | | Total | ₹2,94,000 |

Roughly ₹3 lakh, two weeks elapsed time, and the catalog goes live with one face carried across every plate. This is the number to beat. There is broader context on Indian shoot pricing in the photography cost piece.

How AI persona-lock solves it

The naive way to generate sixty SKUs of catalog is to generate three hundred plates independently, ask the model to "match the previous outputs," and pray. This does not work. Identity drifts within a couple of generations, the front plate of SKU 1 and the front plate of SKU 30 look like two different people, and Myntra's QC catches it.

The technique that holds identity is what we call persona-lock. The architecture is:

Step 1 — Anchor

A persona is created once, at the start of the project. The brand uploads two or three reference photos of how they want the model to look, or selects a persona from a curated library. The pipeline generates an anchor identity — a small bundle of facial embeddings, body proportions, hair characteristics, and stylistic markers — and stores it under a persona ID.

This anchor is the source of truth for the entire catalog. Every subsequent generation references it.

Step 2 — Sibling generation

For each style — defined as one outfit across five plates — the pipeline does not generate five independent images. It generates five images conditioned on the same persona anchor and on each other. The plates come out as siblings of one shoot, sharing identity, lighting, background tone, and styling.

The technical effect is that the same model appears in five poses, three of which are obviously similar (front, back, lifestyle) and two of which are intentionally different (detail close-up, on-figure variation), all sharing the same face.

Step 3 — Cross-style propagation

When you move from style 1 to style 2 — same model, different outfit — the persona anchor is re-used. The pipeline does not regenerate the identity from scratch. The face, the body, the vibe carry forward. Only the garment, the pose, and where useful the background change.

This is the step that makes a 60-SKU drop coherent.

Step 4 — Identity verification

Every generated plate is run through a final identity verifier that compares its face embedding against the anchor persona. If the cosine distance exceeds a threshold, the plate is rejected and regenerated. The threshold is tuned tightly enough that human-eye drift is caught before plates reach the seller.

The 60-SKU drop, hour by hour

This is the actual timeline from the May drop. Start time was 10:00 AM on a Tuesday. End time, defined as all three hundred plates approved and downloaded, was 2:15 PM the same day.

| Time | What happened | |------|---------------| | 10:00 | Brand uploads two reference photos for the persona anchor | | 10:08 | Anchor persona generated, brand approves on second iteration | | 10:15 | Brand uploads 60 garment references — flat-lays and on-mannequin shots | | 10:30 | First 10 styles queued, generation starts | | 11:00 | First 10 styles delivered, brand reviews — 9 approved, 1 sent back for pose variation | | 11:15 | Next 20 styles queued | | 12:00 | Lunch break for the brand team | | 12:45 | 30 styles approved, next 30 queued | | 14:00 | All 60 styles delivered, brand reviewing | | 14:15 | Final approvals, plates downloaded, drop ready for Myntra upload |

Four hours and fifteen minutes from kickoff to download. The Myntra upload itself took another hour, the Shopify upload took thirty minutes. Live on both platforms by 5:30 PM the same day.

What it cost — real ₹ table

The headline number for the AI route was ₹6,000. Sixty styles at ₹100 each.

The honest comparison:

| Line item | Studio route | AI route (Relive) | |-----------|--------------|-------------------| | Generation / shoot | ₹40,000 | ₹6,000 | | Model fees | ₹60,000 | included | | HMU + stylist | ₹44,000 | included | | Studio booking | ₹35,000 | included | | Retouch | ₹75,000 | included | | Logistics | ₹15,000 | nil | | Buffer / reshoots | ₹25,000 | refund on full-style failure | | Total | ₹2,94,000 | ₹6,000 | | Time to live | 2 – 3 weeks | 1 day |

The cost difference is roughly 49x. The time difference is roughly 15x. Neither number is the headline. The actual point of the case study is that one face was carried across three hundred plates with no measurable drift — the constraint that historically forced the studio route was solved in software.

A note on GST. Relive is not GST-registered today, so the ₹6,000 was the rupees that left the brand's account — no 18 percent added. When we register, the price will be inclusive of GST or clearly called out. The detail is on /gst.

Failure modes we hit and fixed

This is the part most case studies skip. Three failure modes hit during this specific drop, and they are worth naming honestly.

Identity drift on detail plates

The first batch of 10 styles came back with one persistent issue — the detail close-up plate (plate 3 of 5) was occasionally generating a different jawline from the other four plates. The close-up framing was changing which features the model emphasised, and the identity verifier threshold was set loose enough to let some of these through.

Fix: the verifier threshold was tightened for close-up plates specifically, and plates that fail get regenerated up to three times before being escalated. After the fix, drift on detail plates dropped from roughly 8 percent to under 1 percent across the rest of the drop.

Pose collapse on lifestyle plates

Plate 4 (lifestyle) was supposed to be in a different setting from the studio plates — a coffee shop, a street, a sunlit room. For the first 20 styles, lifestyle plates were defaulting to the studio look with a faint background. The pipeline was over-prioritising identity consistency at the cost of pose and context variation.

Fix: the lifestyle plate prompt was given more explicit context separation from the other plates, and the pipeline was told to allow background variation without losing identity. After the fix, lifestyle plates looked like a lifestyle shoot, not like the studio with a different backdrop.

Garment hallucination on patterned fabric

Two garments in the drop had complex prints — a paisley and an intricate block print. On the first generation, the pipeline drifted on the print pattern between plates. Plate 1 had one paisley arrangement, plate 4 had a slightly different one. Subtle, but Myntra QC catches this.

Fix: garment-specific anchoring was added — for patterned fabrics, the print is locked separately from the persona, and the same print sample is referenced on every plate. This is now default for any garment flagged as patterned.

All three fixes are now baked into the pipeline. They came from this drop and from a handful of others that exposed similar edge cases. Catching these is the work that the ₹100 per style price covers — the QC infrastructure around the generator, not just the generator itself.

What this means for first-time AI photography buyers

A few honest things to take away if you are evaluating AI photography for the first time.

First, persona-lock is not magic. It is anchor identity + sibling generation + verification, and pipelines that do not have all three layers will drift. Before committing a catalog to a generator, generate two styles of the same persona and check whether the face holds across both. If it does not, the pipeline does not have what it needs to handle a 60-SKU drop.

Second, the price comparison is sharper than most people expect. ₹6,000 versus ₹2,94,000 for the same coherent catalog is the actual number, not a marketing number. The reason this is sustainable at ₹100 per style is that the generation cost is a small fraction of what a model day plus stylist plus retouch plus logistics adds up to.

Third, you should test on your own SKUs before believing this. The three free styles on signup are deliberately structured so that you can put a real garment from your catalog through the pipeline and see whether the output clears your internal bar. If it does not, you find out in an hour, not after a paid invoice.

Start a drop on /signup, or read the apparel-specific breakdown if you are sizing up a Myntra catalog. If something needs a human eye, the studio chat at /support is open during working hours.

The 60-SKU drop in this case study went live, sold through, and gave the brand a coherent catalog that an agency shoot would have cost fifty times as much to produce. The technique behind it is reproducible. That is the whole point.