Evidence, with its limits visible

Inspect the garment, not just the image.

A useful catalogue result preserves the product. This page separates existing demonstrations from measured collection-workflow validation.

These examples are not a completed 20-garment benchmark, an independent review or a claim of guaranteed fidelity. They are existing PixelAPI demonstrations; they do not show one garment passing through all four formats.

Original garment used in the existing virtual try-on demonstration — Original from the existing public virtual try-on demonstration.

Existing virtual try-on demonstration output — Existing virtual try-on output. This illustrates a related tool, not a new Collection Studio benchmark result. Check the source and output yourself.

Existing PixelAPI patterned garment flat-lay demonstration — Existing flat-lay demonstration, published from an approved internal run. It is a separate garment; no matching source-to-output fidelity score is claimed here.

What this existing demonstration does and does not preserve

The red top retains the broad colour, two chest pockets and rolled-sleeve design. The generated neckline appears flatter, and the hem and proportions change on the model. These visible differences require a product review; this pair is not evidence of exact garment preservation. The image cannot establish physical fit, fabric weight or hidden construction.

Measured collection validation

4 November 2026: one public demo garment, three still-image formats, four attempts. These are actual outputs from the connected collection workflow, generated with a dedicated internal test account. The input was already part of PixelAPI's public demo library. Its origin as a real physical garment is not established; this is a development fixture, not a merchant-product benchmark.

Public demo source: red top with two pockets and visible neck label — Unchanged source used for all three still formats.

Actual flat-lay output with visible neck label removed — Flat lay: broad colour and shape retained; the visible neck label was removed. Needs correction for exact-product use.

Actual ghost-mannequin output with a pale border around the garment and damaged neck label — Ghost mannequin: label damage and a pale cutout border. This flat-lay input does not demonstrate hollow mannequin geometry.

Actual on-model output showing a red two-pocket top with glossier fabric appearance — On-model: garment remains recognisable, but fabric appearance is glossier and construction detail is softened. True fit and the occluded inside label cannot be assessed.

Format	Attempts	Recorded credits	Recorded wait	Processing	Review
Flat lay	1	10	234.8s	227.4s	Needs correction
Ghost mannequin	1	60	93.5s	91.0s	Needs correction
On-model	1	40	79.7s	60.5s	Needs correction
Flat lay, public repeat	1	10	64.9s	57.3s	Needs correction

All four outputs were delivered. None was approved for exact-product catalogue publication in this review. Total recorded generation credits: 120. Credits per approved output: unavailable because there are no approvals. Rejected outputs remain in the accounting; no quality-based refund was requested. Credits here are internal test usage, not money spent or a quoted customer price.

Review method: one agent compared the source and outputs visually; this is not blind or independent review. Review time was not measured reliably, so no reviewer-efficiency result is claimed. Recorded waits for this development run begin after initial input validation, include queue/execution and result synchronisation, and exclude upload time. The fresh public repeat includes initial request validation in its recorded wait. These four measurements are not a service-level promise or a comparison with another provider.

Video was not generated in this sample. The 20-garment, four-format evaluation is still incomplete. Pipeline corrections and repeat measurements are pending; no repaired-output success is claimed. These failures are shown so the limitations are visible rather than excluded from the sample.

Fresh public flat-lay repeat; the neck label is still removed — Fresh public browser repeat after deployment: 64.9 seconds recorded wait, 57.3 seconds processing. The faster, already-loaded-model run still removes the neck label; fidelity did not improve. This fourth attempt and its 10 credits are included in the totals above.

The evaluation protocol

The initial target is 20 distinct, permission-cleared garments across four formats: flat lay, ghost mannequin, on-model and video. The selection should cover solids, stripes, repeated prints, legible logos, embroidery, lace or sheer fabric, reflective detail, dark colours, light colours and complex silhouettes. A larger or repeat evaluation is needed before generalising.

Freeze each original and a SKU-specific checklist before generating. Record provenance and permission to publish.
Use common settings and record every attempt, including failures and retries. Animate the explicitly selected approved still.
Have reviewers compare source and output at full size, with provider and attempt ordering concealed where practical. Keep automated QC separate from merchant decisions.
Record colour, prints and logos, cut and shape, fabric and construction, and overall suitability. Inspect video frame by frame for temporal changes.
Report approved product formats, all-attempt credits, customer wait, reviewer time, failed details and unassessable details. Do not omit rejected products.

Known limitations to check

Hidden construction

A photograph cannot establish unseen garment geometry, exact fabric weight or physical fit. Generated interior and occluded details may be invented.

Fine details

Small text, embroidery, repeated prints, transparency and hardware can change. A visually attractive output can still be unsuitable for a listing.

Motion stability

A good still does not prove a good video. Inspect colour, edges, logos and fabric throughout the clip.

What the Studio records

Private originals, source fingerprints, immutable attempt settings, all generation attempts, merchant checklists, decisions, recorded credits, refunds, wait time and review time. A manifest supports an auditable review; it does not certify product accuracy.

Review your own collection Read the measurement definitions