The programme runs for eighteen months in seven phases. The image track fills most of the first year; the video track starts in month 4 with its capture protocol and becomes the main effort from month 9. Month 1 is indicatively October 2026. If funding starts later, the months shift and the venue plan moves to the next cycle of each venue.
4.1Phases
| Phase | Months | Key activities | Deliverables | Exit criterion |
|---|---|---|---|---|
| P0 Foundations | 1 | Company registration and DPIIT recognition; evaluation harness; consent, annotation and data protocols; licence register; cloud credit applications | Harness running on at least six systems; approved protocols | Harness reproduces published numbers within tolerance on a public split |
| P1 Pilot | 2 | Pilot shoot of 20 garments across the garment families; annotation guide; pilot audit | Pilot set; inter-rater agreement report | Krippendorff's alpha of at least 0.6 on garment attributes |
| P2 Benchmark | 3 to 6 | Main collection and annotation; audit of at least ten systems; human study; Paper 1 | Benchmark v1 frozen; Paper 1 on arXiv and submitted | Benchmark meets its size, garment-family and strata targets |
| P3 Image model | 3 to 9 | Data engine; synthetic triplets; LoRA stages; multi-view conditioning; ablations; preference tuning | Image model; ablation report | First LoRA beats the current production model on pilot garment fidelity |
| P4 Distil and transfer | 9 to 12 | Distillation; serving tests; serving-pipeline clean-up; Paper 2; image-track releases | Distilled image model; Paper 2 submitted; toolkit and benchmark released | Distilled model within twice 6.5 seconds on an L40S with no significant loss in preference |
| P5 Video model | 4 to 14 | Video capture protocol and pilot shoot; main shoots; video data engine; offline video model; Paper 3 | Video training set v1; offline video model; Paper 3 submitted | Offline model beats open video baselines on garment fidelity over time on the hold-out |
| P6 Live | 12 to 18 | Causal conversion and distillation; streaming servers; live pilot behind the existing live page; Paper 4 | Live model; pilot report; Paper 4 submitted | At least 15 fps at 512p on one GPU, first frame under 2 s, cost under a tenth of the rented engine |
4.2Schedule and milestones
Milestones
1234·56789·101112131415
WP0 Foundations: entity, harness, protocols
WP1 Pilot shoot and annotation guide
WP1 Benchmark collection and annotation
WP1 Baseline audit and human study
WP5 Paper 1 writing and submission
WP2 Data engine and synthetic triplets
WP3 LoRA training stages and ablations
WP3 Multi-view conditioning experiments
WP4 Preference tuning and distillation
WP5 Paper 2 writing and submission
WP4 Serving integration in the product
WP6 Video capture protocol and pilot shoot
WP6 Video shoots and synthetic video pairs
WP6 Offline video model training
WP5 Paper 3 writing and submission
WP7 Causal conversion and distillation
WP7 Streaming servers and live pilot
WP5 Paper 4, releases and final report
Programme month
123456789101112131415161718
- Foundations
- Image track
- Video track
- Live track
- Papers
| ID | Milestone | Month | Evidence |
|---|---|---|---|
| MS1 | Evaluation harness running on at least six systems | 1 | Harness report |
| MS2 | Pilot set and inter-rater agreement report | 2 | Pilot report |
| MS3 | Benchmark v1 frozen | 5 | Datasheet and data card |
| MS4 | Paper 1 posted to arXiv and submitted | 6 | Submission record |
| MS5 | Video pilot: capture protocol tested and first 200 consented clips | 6 | Pilot report |
| MS6 | First LoRA beats the current production model on pilot garment fidelity | 7 | Evaluation report |
| MS7 | Image ablations complete | 8 | Ablation report |
| MS8 | Paper 2 submitted | 10 | Submission record |
| MS9 | Distilled image model meets the latency target | 11 | Speed laboratory report |
| MS10 | Video training set v1: 7,000 real clips and 20,000 verified synthetic pairs | 11 | Data card and provenance ledger |
| MS11 | Image-track releases and interim report | 12 | Repositories and interim report |
| MS12 | Offline video model beats open video baselines on the hold-out | 13 | Evaluation report |
| MS13 | Paper 3 submitted | 14 | Submission record |
| MS14 | Live prototype: at least 15 fps at 512p on one GPU, first frame under 2 s | 16 | Live evaluation report |
| MS15 | Live pilot with shoppers, Paper 4 submitted and final report | 18 | Pilot report, submission record and final report |
4.3Publication calendar
| Venue | Indicative deadline | Status on 30 September 2026 | Target |
|---|---|---|---|
| CVPR 2027, main conference [63] | 16 November 2026 | Confirmed | Too early for any paper |
| CVPR 2027 workshops | Around March 2027 | Not yet announced | Paper 1 |
| ACM Multimedia 2027 | Around late March to April 2027 | Not yet announced | Paper 1 or Paper 2 |
| NeurIPS 2027 Evaluations and Datasets [64] | Around May 2027 | Not yet announced | Paper 1, extended |
| BMVC 2027 | Around late May 2027 | Not yet announced | Paper 2 if results are early |
| WACV 2028 | Around June and August 2027 | Not yet announced | Paper 2 or the cost report |
| CVPR 2028 | Around mid-November 2027 | Not yet announced | Paper 3 |
| SIGGRAPH 2028 | Around January 2028 | Not yet announced | Paper 4 if results are early |
| ECCV 2028 | Around March 2028 | Not yet announced | Paper 4 |
Deadlines marked 'around' follow each venue's usual cycle and will be confirmed when calls open. These venues allow arXiv preprints if the submitted paper is anonymous.
4.4Governance and ways of working
- A weekly research meeting with a shared agenda, and a monthly steering review with the founders.
- One code repository, one experiment tracker and versioned configurations, so every reported number can be reproduced.
- A named data steward who owns the provenance ledger, consent records and release decisions.
- An authorship and contribution policy agreed in writing before work starts.
- Quarterly progress reports to funders against the milestones and the key performance indicators.
4.5Risk register
| Risk | Likelihood | Impact | Mitigation |
|---|---|---|---|
| A licence contaminates the model | Medium | High | Apache-2.0 or MIT bases, teachers and annotation tools only; licence register and provenance ledger; legal review before any release |
| Consent or privacy failure | Low | High | Written consent and releases covering AI training and synthetic derivatives; no scraping; anonymised public release |
| Another group publishes first | Medium | Medium | Post to arXiv early; keep the focus on garment coverage, fairness strata and garment-faithful live try-on |
| Data collection is slower than planned | Medium | Medium | Pilot first; catalogue partners in parallel; a 1,000-pair minimum image benchmark as fallback |
| Diverse subjects and garments are hard to source from one country | Medium | Medium | Catalogue, stock and footage partners in several regions; partner shoots abroad; coverage reported honestly for every stratum |
| Video data is costly and slow | High | Medium | Pilot shoot before Tier A, Tier A before Tier B; shooting in India lowers cost; licensed footage covers the unpaired stage |
| Live garment accuracy falls short | Medium | High | Ship offline video first; keep the rented engine for pilots; report the gap against the hold-out honestly |
| Serving cost of large models | High | Medium | Distil the image model into a 4B student and the video model into a 1.3B live student; keep the current production model until the student wins |
| GPU capacity for training and live serving | Medium | Medium | Cloud credits across providers; the portable worker container already built; one GPU per live stream by design |
| Live video is misused to fake someone's appearance | Medium | High | The guardrail plan: age and consent checks at session start, sampled output moderation, session limits and labels |
| Volunteer attrition | Medium | Medium | Funded fellowships, scoped tasks and clear authorship |
Sources in this chapter
- [63]CVPR 2027. Key dates. cvpr.thecvf.com/Conferences/2027/Dates
- [64]NeurIPS 2026. Call for Evaluations and Datasets. neurips.cc/Conferences/2026/CallForEvaluationsDatasets