Status: Living document
Last updated: July 21, 2026
Owner: TryOn Labs
Scope: OpenTryOn (open-source Fashion AI toolkit) + commercial layer (TryOn Studio / Playground)
OpenTryOn is the open-source Fashion AI toolkit — models, training, fine-tuning, prompts, agents, efficient inference, and an open studio UI — so builders can create virtual try-on, catalog, and fashion content systems without starting from scratch.
Fashion-first now. Platform-generic later.
Fashion is one of the hardest domains for generative AI and one of the highest-ROI for commerce:
| Pain | Who feels it | Why it persists |
|---|---|---|
| Fit & appearance uncertainty drives returns | Shoppers, DTC brands | Online apparel returns often 20–40%+; try-on is still rare (~1% of stores) |
| Catalog / PDP production is slow and expensive | Brands, agencies, marketplaces | Photoshoots don’t scale with SKU velocity; seasonal churn multiplies cost |
| Research ≠ product | ML engineers, startups | Academic VTON repos (IDM-VTON, CatVTON, OOTDiffusion) are powerful but fragmented, hard to train, hard to run efficiently |
| Vendor lock-in & API sprawl | Product teams | Many closed APIs (FASHN, Revery, Vue.ai, Nightjar, Botika…) with different schemas, credits, and quality trade-offs |
| No shared “fashion OS” for agents | Builders of agentic workflows | Chat UIs and raw APIs don’t compose into catalog → try-on → lookbook → video pipelines |
| Efficient deployment is expert-only | Teams with limited GPU budget | Quantization, distillation, pruning, KV-cache, speculative decoding rarely packaged for fashion models |
Market signal (directional, sources vary by definition): virtual try-on market estimates around ~$15B in 2025 → ~$46–48B by 2030 (~26% CAGR). Reported retail outcomes include ~20–35% conversion lift, ~25–40% return reduction, AOV up to ~33% — while consumer demand far exceeds adoption. AI/ML-based try-on is among the fastest-growing segments vs classic AR overlays.
OpenTryOn exists because the gap is not “another demo” — it is infrastructure: training + inference + prompts + agents + studio, open enough to own, commercial enough to ship.
┌─────────────────────────────────────────────────────────────────┐
│ TRYON STUDIO (open UI + commercial hosted) │
│ Visual workflows, brand kits, agent tasks, export to PDP/video │
├─────────────────────────────────────────────────────────────────┤
│ PLAYGROUND / API (developer closed-source surface) │
│ Credits, orgs, logs, evals — already live for API testing │
├─────────────────────────────────────────────────────────────────┤
│ OPENTRYON (open-source toolkit) │
│ CLI · MCP · adapters · local models · train/finetune · agents │
│ prompts · efficiency recipes · datasets · docs │
└─────────────────────────────────────────────────────────────────┘
| Layer | Nature | Job |
|---|---|---|
| OpenTryOn | Open source (CC BY-NC today; revisit license for commercial OSS strategy) | The toolkit researchers and engineers fork, extend, and run locally or against any provider |
| Playground | Closed / hosted | Developers test APIs, compare models, inspect outputs — exists today |
| TryOn Studio | Open UI code path + hosted commercial product | Non-dev fashion teams: try-on, model swap, catalog, campaigns, agents — without assembling the stack |
This matches the existing TryOn Labs split: shared core, developer Playground, outcome-focused Studio/agentic product — OpenTryOn is the open heart of that core.
- Generation & try-on — VTON, image gen/edit, video, bg remove (cloud adapters + local weights)
- Preprocessing — garment/human segmentation, pose, captioning
- Training & fine-tuning — LoRA/QLoRA, Unsloth-style recipes, fashion datasets, brand style adapters
- Prompt collections — curated, versioned prompts for try-on, catalog, lookbook, video
- Efficient inference — quantization, distillation, pruning, KV-cache, speculative decoding, batching — scripts that a fashion ML eng can actually run
- Agents — task-scoped agents (PDP optimizer, catalog generator, try-on QA, lookbook director, return-risk advisor…)
- Studio UI — open-source studio for local/self-host; commercial hosted Studio for teams
- Use-case segmented, not vendor-segmented (CLI/MCP registry pattern already in place)
- One invoke path for CLI and MCP (
invoke_model) so tools never drift - Fashion fidelity over generic demos — garment preservation, identity, pose, lighting
- Efficiency is a first-class feature, not a blog post
- Open to learn / closed to operate at scale — OSS grows the ecosystem; Studio + Playground fund the company
- Agents are workflows, not chatbots with fashion stickers
| Player | Focus | Strength | Weakness vs OpenTryOn |
|---|---|---|---|
| FASHN | Fashion API + studio (try-on, model create, edit, video) | Strong fashion-native API; agent skill for coding agents; credit pricing | Closed models; not a training/finetune/efficiency toolkit |
| Vue.ai | Enterprise retail AI suite | Full stack for large retailers | Heavy enterprise sales; not OSS toolkit |
| Revery.ai | Virtual dressing room APIs | E-comm developer focus | Narrower surface; closed |
| Nightjar, Botika, SellerPic, Genlook | Brand catalog / Shopify imagery | Shopify-native workflows, catalog consistency | Brand tools ≠ open ML platform |
| Veesual | High-end photo compositing | Photoreal for premium brands | B2B, not self-serve OSS |
| Google Shopping VTO / Walmart Zeekit | In-platform try-on | Distribution | Locked to their catalogs |
| DressX / digital fashion | High-fidelity / digital garments | Luxury aesthetics | Different ICP (digital fashion) |
| Generic genAI (Gemini, GPT Image, FLUX, Veo, Sora) | Horizontal models | Quality & speed | Not fashion-ops; no VTON training stack |
| Project | Focus | Notes |
|---|---|---|
| IDM-VTON | Diffusion VTON (ECCV 2024) | High fidelity; popular HF demos; CC BY-NC-SA |
| CatVTON | Efficient VTON (ICLR 2025) | <8GB VRAM path; CatVTON-FLUX LoRA |
| OOTDiffusion | Outfitting fusion VTON | Foundational open model; large community |
| OrthoTryOn | Unified fashion gen (try-on + pose + reconstruction) | Research-efficient multi-task LoRA |
| OpenVTO | Studio avatar + try-on + short video loops | Early toolkit; aesthetics-first |
| ComfyUI workflows / HF Spaces | Ad-hoc pipelines | Powerful but not a productized toolkit |
| OpenTryOn (us) | Multi-provider CLI/MCP + local models + docs + agents path | Broadest ops surface among fashion OSS; still thin on train/efficiency/studio |
Research repos Closed fashion APIs
(IDM/Cat/OOT) (FASHN, Revery…)
\ /
\ /
\ /
▼ ▼
┌─────────────────────────────┐
│ OPENTRYON = fashion AI OS │
│ train · serve · prompt · │
│ agent · studio · any model │
└─────────────────────────────┘
We are not “another VTON model.”
We are the toolkit and studio layer that wraps open models and closed APIs, adds training/efficiency/agents, and ships a path from notebook → CLI/MCP → Playground → Studio.
Differentiation bets:
- Provider-agnostic fashion registry (already shipping)
- Training + fine-tuning recipes for brand/SKU adaptation
- Efficiency pack for running fashion models on consumer/prosumer GPUs
- Prompt + eval collections as shared assets
- Agents for fashion ops outcomes (catalog, PDP, QA)
- Open Studio + commercial hosted Studio/Playground dual flywheel
| Segment | Persona | Jobs to be done |
|---|---|---|
| Fashion-tech startups | Founding engineer / ML lead | Ship MVP try-on or catalog AI without rebuilding adapters |
| Agency / creative tech | Technical producer | Script batch lookbooks; swap providers by cost/quality |
| Academic / indie researchers | Grad student, HF contributor | Train/finetune VTON; publish with reproducible scripts |
| Internal ML at DTC / marketplace | Applied ML eng | Self-host sensitive garments; evaluate closed APIs side-by-side |
Anti-ICP (OSS): pure marketers with no engineer; enterprises needing only a Shopify button (send to Studio later).
- Developers integrating fashion APIs
- Teams comparing model quality before contracting
- Hackathon / prototype builders
| Tier | Who | Buying trigger |
|---|---|---|
| Primary | DTC apparel brands ($1M–$50M GMV) with thin photo budgets | Seasonal catalog velocity, return rates, SKU count |
| Secondary | Fashion e-comm / marketplaces | On-model imagery at scale; multi-brand consistency |
| Secondary | Creative agencies serving fashion | Faster client turnaround; reusable brand models |
| Enterprise | Large retailers | Security, SSO, SLA, private models, on-prem/VPC |
Buyer: Head of E-comm / Creative Ops / Digital Product (economic); ML/Eng as champion when self-host matters.
OpenTryOn remains the community and adoption engine. Revenue sits on hosted products and enterprise.
| Stream | Model | Notes |
|---|---|---|
| Playground + API | Credits / usage | Already aligned with “developers test our APIs” |
| TryOn Studio SaaS | Seat + usage (Starter / Pro / Studio / Enterprise) | Outcome workflows; mapped internally to credits |
| Enterprise | Annual contract + VPC / private weights / support | Fine-tuned brand models, SSO, audit logs |
| Professional services | Optional | Custom agents, training on brand data (keep productized) |
| OSS | Free (non-commercial today) | Consider dual-license or commercial OSS later without breaking community trust |
Pricing principles:
- Credits for compute-heavy gen (Playground/API)
- Subscription for workflow/agent value (Studio)
- Never charge for “access to docs” — docs stay open to grow OpenTryOn
Unit economics levers: model routing (cheap vs quality), caching, batch, efficient local inference for self-host Enterprise, multi-provider failover.
- GitHub + docs as source of truth (docs)
- Discord / LinkedIn / WhatsApp — ship notes like the recent 4-API drop, not press releases
- MCP + CLI — meet developers where agents already work (Cursor, Claude, etc.)
- HF Spaces / notebooks — one-click try-on and efficiency demos
- Content: “compare FASHN vs Flux VTO vs CatVTON on the same garment” eval posts
- Sign up → run try-on in minutes → export code snippet that uses OpenTryOn adapters
- Credits for power users; free tier for exploration
- Partner listings (API marketplaces) once stable
- Convert Playground power users and Discord brands into Studio trials
- Shopify / catalog integrations as distribution (phase later)
- Case studies: return-rate and time-to-PDP metrics
- Agency channel: white-label workflows
| Audience | Message |
|---|---|
| Engineers | “Fashion AI toolkit: train, serve, swap providers, agentize.” |
| Brands | “Studio that produces on-model and try-on content without a full reshoot.” |
| Investors | “OSS wedge → usage → Studio/API revenue in a $15B→$48B try-on market.” |
Horizon assumes fashion-only through Phase 3; generic verticalization after proof.
- Multi-provider CLI + MCP registry
- Cloud VTON / gen / edit / video / understand / bg-remove adapters (incl. Pruna, Nano Banana 2 Lite, FASHN, Gemini Omni)
- Docs site + Discord community
- Playground for API testing
- Early agents / Gradio demos; Studio spun out as separate app talking over MCP
Exit criteria: Developers can discover and dry-run every registered model; docs match registry.
Theme: From adapters to a real Fashion AI toolkit.
| Workstream | Deliverables |
|---|---|
| Local OSS VTON | First-class adapters for CatVTON / IDM-VTON / OOTDiffusion (or FLUX-fill LoRA paths) under tryon.models |
| Train / finetune | Documented LoRA/QLoRA recipes; fashion dataset loaders; brand-style fine-tune notebook |
| Prompt collections | Versioned prompt packs (try-on, catalog, lookbook, video) in-repo |
| Evals | Minimal garment-fidelity / identity-preservation checklist + scripted side-by-side runner |
| Studio OSS MVP | Open UI that drives OpenTryOn via MCP/CLI for try-on + generate + export |
| License clarity | Publish commercial-use policy for adapters vs weights vs Studio |
Exit criteria: A new contributor can fine-tune a small fashion adapter and run local VTON with opentryon[local] without reading 5 research READMEs.
Theme: Run serious models on realistic hardware.
| Workstream | Deliverables |
|---|---|
| Quantization | 8-bit / 4-bit recipes for priority local models; VRAM tables in docs |
| Distillation / pruning | Starter scripts or wrappers where upstream supports them; clear “supported / experimental” labels |
| Serving tricks | KV-cache guidance, batching, speculative decoding where applicable (LLM/VLM paths first; diffusion as research track) |
| Benchmarks | Latency / VRAM / quality scorecards published with each efficiency release |
| Agents v1 | 3–5 fashion agents: Catalog PDP, Try-on QA, Lookbook director, Model-swap ops, Prompt librarian |
| Playground maturity | Orgs, usage meters, model comparison UI |
Exit criteria: Published “efficient CatVTON / Flux path on 12–16GB GPU” guide with numbers; ≥2 agents used in Studio/Playground weekly by design partners.
Theme: Outcomes for brands, not knobs for engineers.
| Workstream | Deliverables |
|---|---|
| Hosted TryOn Studio | Auth, projects, brand kits, batch jobs, exports |
| Billing | Credits + subscription tiers |
| Integrations | Shopify / DAM / S3 export; webhooks |
| Private models | Brand LoRA hosting; data retention controls |
| Enterprise | SSO, VPC option, SLA |
Exit criteria: Paying Studio or API customers; NPS/retention from design partners; clear CAC payback narrative.
Theme: Fashion-proven → adjacent verticals.
- Beauty, eyewear, home soft-goods (same try-on/catalog patterns)
- Shared efficiency + agent framework with vertical prompt/eval packs
- Keep fashion as the flagship reference vertical
Prioritize task agents with measurable outputs:
| Agent | Input | Output |
|---|---|---|
| PDP Optimizer | Product images + copy | Score + rewrite + image set recommendations |
| Catalog Generator | Flat-lay / ghost mannequin | On-model gallery + metadata |
| Try-On QA | Person + garment + result | Pass/fail + defect tags (bleed, pose, logo) |
| Lookbook Director | Brand brief + assets | Multi-shot sequence + video prompts |
| Provider Router | Constraints (cost, latency, quality) | Chosen model + fallback chain |
| Return-Risk Advisor | Garment attributes + fit notes | Risk score + try-on suggestion |
| Fine-Tune Coach | Brand images | Dataset checklist + train config |
Agents should call OpenTryOn tools (MCP/CLI), not reimplement providers.
Package as opentryon efficiency docs + scripts:
- Baseline — FP16/BF16 reference latency & VRAM
- Quantize — bitsandbytes / GGUF / AWQ where model family allows
- Compile / kernels — torch.compile, FlashAttention where relevant
- Distill — student models for edge try-on previews
- Prune / sparsity — experimental; only when quality gates pass
- LLM stack — KV-cache, speculative decoding for understanding/agent planners
- Ops — batch queues, warm models, result caching
Every recipe must answer: GPU class, VRAM, ms/image, quality delta vs baseline.
| Risk / decision | Mitigation |
|---|---|
| CC BY-NC limits commercial OSS adoption | Dual-license or Apache for toolkit code; keep research weights under upstream licenses |
| Competing with FASHN while integrating FASHN | Stay provider-agnostic; differentiate on train/efficiency/agents/studio |
| Training data / likeness / consent | Strict dataset docs; SynthID / provenance guidance for gen video |
| Scope creep into “generic AI platform” too early | Fashion-only OKRs until Phase 3 exit |
| Studio vs Playground confusion | Playground = capabilities; Studio = tasks/outcomes (existing naming decision) |
- GitHub stars / forks / unique contributors
- Docs traffic; Discord active weekly users
- Registry model count + MCP tool adoption
- External PRs for adapters / prompts / efficiency scripts
- Playground signups → first successful API call
- Credit burn and retention
- Studio trials → paid conversion
- Design-partner case studies (time-to-PDP, return proxy metrics)
- Ship Phase 1 spine: one local VTON path + one fine-tune recipe + prompt pack v0
- Efficiency v0: VRAM/latency table for that local path
- Studio OSS MVP: try-on + generate wired to MCP
- GTM: monthly “toolkit drop” posts (APIs, recipes, evals) — same tone as the 4-API LinkedIn post
- Align license + commercial boundary in README/VISION so contributors know what’s open vs hosted
- OpenTryOn docs: https://tryonlabs.github.io/opentryon/
- OpenTryOn repo: https://github.com/tryonlabs/opentryon
- Market: Mordor / Grand View style VTON market reports (~$15.18B 2025 → ~$48.1B 2030, ~26% CAGR); retailer ROI writeups citing conversion/return/AOV ranges
- Open models: IDM-VTON, CatVTON, OOTDiffusion; OrthoTryOn; OpenVTO
- Closed: FASHN, Vue.ai, Revery, Nightjar, Botika, Genlook, Google/Walmart in-platform VTO
- Internal: TryOn AI Strategy & Product Plan (Playground vs Studio split, tryonlabs.ai domains)
| Name | Role |
|---|---|
| OpenTryOn | Open-source toolkit |
| TryOn AI Playground | Developer hosted API testing / credits |
| TryOn Studio (aka TryOn AI product UX) | Agentic / workflow product for fashion teams |
| TryOn Labs | Company / org |
Do not rename or re-architect prematurely. Generalization = extract shared packages (efficiency, agents, registry) once fashion Studio has paying users and repeatable evals — then add vertical packs (beauty, eyewear) rather than diluting fashion depth.