Case study: We replaced one fashion campaign shoot with Lamina-generated product photos and reels—what stayed product-accurate, what still required a real shoot, and the cost/time trade-off

This is not a finished customer case study showing a fashion brand replacing a campaign shoot with Lamina. The evidence supplied covers a proposed one-SKU test and three per-run measurements. It does not report approval rates, fidelity scores, revision time, published assets, or fully loaded campaign savings. That distinction matters. Fast generation helps, but any campaign claim needs proof that the garment in frame matches the garment being sold.

The defensible replacement is tighter. Capture a controlled source-of-truth reference pack for the SKU, then use Lamina for the repeatable long tail: approved model selections, locations, crops, formats, lifestyle scenes, and short vertical variants. Lamina’s product workflow asks for front and back garment inputs, and the platform describes locked product references and brand-kit scoring. Product-expert approval still has to happen.

What did the recorded Lamina experiment actually prove?

The recorded experiment establishes $0.04 cost parity per run and slower measured generation latency for the Lamina variants. It does not establish lower campaign cost, faster approved-asset turnaround, or product accuracy. The reference-locked still took about 68 seconds; the reel took about 71 seconds, against about 50 seconds for the conventional-control run. For a producer, that is a small timing gap per iteration—not evidence of an end-to-end production win.

The protocol earns its keep because it spells out a credible comparison: one fashion SKU, 12 turntable images, five macro-detail crops, exact colour values, dimensions, logo-clear-space rules, and a shot list. It specifies six still concepts and three six-second vertical reel concepts per variant. What is absent is just as useful to flag: no blind accuracy scores, rejected outputs, review or editing time, or cash cost for source capture.

Which product details can Lamina-generated fashion assets carry into campaign variants?

Lamina-generated stills and reels can bring a defined product reference into campaign variants at scale, as long as every output enters approval as a candidate rather than product proof. Lamina says its brand controls can apply palette, typography, voice rules, locked product references, reference imagery, and output scoring. Its apparel workflow also requests front and back garment images, plus model and location selections.

Start with more than one clean front image. A workable fashion reference pack includes front, back, side, and texture-detail views; a field test cited in the research found that one front-facing garment image leaves too much for the model to infer. That test also found fabric feel could wander from the real garment. Cross-angle consistency was the bigger issue.

What still requires controlled source capture and closer review?

Keep controlled source capture and close human review wherever an image has to prove a sellable physical fact: exact fit across body types, construction placement, close-up material finish, reflective or transparent hardware, logo and text treatment, and authentic garment motion. You do not need to rebuild a full traditional campaign production for this. You need the evidence pack that makes generated campaign multiplication safer.

Before an asset reaches a PDP, paid placement, or campaign landing page, reviewers should check colour, texture or print, silhouette, seam and line placement, closures, logos or text, hardware, and fit. The supplied academic study found its automated garment-consistency method was more sensitive to colour and texture than shape and line dimensions. Let automation triage. A merchandiser, designer, or product specialist still needs the authority to reject any changed attribute.

Set a hard rule for logos. Diffusion systems can approximate a mark pixel by pixel rather than reproduce the intended one, so legally sensitive brand marks and readable garment text should be composited, checked against the source, or left in a photographic reference frame. For reels, use real-motion source footage where the selling claim rests on actual drape, fit, or hardware movement. Vertical captions and brand-safe audio do not prove physical behaviour.

What is the real cost and time trade-off?

There is no savings claim in the current measurement. All three recorded runs cost $0.04, while the Lamina still and reel runs came in roughly 19 and 21 seconds slower than the control. Test iteration capacity and approval throughput next. The per-run figure excludes human review, corrective work, source capture, and media costs.

Lamina’s published use-case page markets AI ecommerce images at $0.10–$2 per image, compared with $35–$165 for traditional photography, and says it can replace 80% of catalogue shoots. Those are Lamina use-case claims, not independently verified savings for this campaign scenario. Treat them as a hypothesis. Measure your own approved-asset cost against a fixed reference pack and shot list.

A vendor benchmark from Photoroom explains why that discipline matters: its strongest base editing model reached full product fidelity in 29.0% of 4,250 virtual-model generations, increasing to 38.2% with its correction layer. The benchmark still needs independent validation. It is enough, though, to make mandatory QA an operating requirement rather than a cosmetic last pass.

Why is a hybrid fashion-production workflow more credible than an all-or-nothing replacement?

A hybrid workflow is more credible because it keeps physical evidence where accuracy is costly to fake, then applies generation to variants that are costly to produce individually. Use controlled product capture for calibrated colour, fit, macro details, and proof frames. Use Lamina for on-model, locale, background, format, crop, and short-reel permutations, then route every asset to people who know the SKU for explicit approval.

Mark Scarrott’s view matters here because it describes the production model behind that split: professionally captured product evidence can anchor generated environments and creative assets without treating those as the same job.

Who should own approval of AI-generated fashion assets?

The people who understand the physical garment should own final approval. A generative system can scale a sound art direction—and a bad product assumption—with equal speed. Put merchandising or product specialists alongside creative reviewers, give them the source pack, and require a recorded reason for every rejection or correction.

Kamil Czaja’s warning gets to the operating risk: expertise has to enter before generation through better references, then stay involved through disciplined review.

What is the practical decision for fashion ecommerce teams?

Use Lamina to multiply approved creative directions, not erase the product record. Build a compact, controlled reference capture for each high-value SKU. Generate campaign still, model, location, crop, and reel variants around it; publish only frames that pass attribute-level review. Physical production stays focused on evidence, while generative production handles the broad run of on-brand creative output.

For a credible internal case study, publish contact sheets, frame-level annotations, approval and rejection rates, review minutes, revision count, and fully loaded cost per approved asset. Until those figures exist, the conclusion is plain: the workflow is promising and operationally testable, but the supplied experiment has not validated a replacement claim.