Skip to main content

Fashion ML Engineer Path

This page is the outline for AI/ML engineers who want to train, evaluate, and operate fashion models with OpenTryOn — not only call cloud APIs.

Status: target path for v0.1.0 (Fashion ML Toolkit Core). Invoke-layer pieces below work today on v0.0.3; train/eval/workflow sections are planned. See the Roadmap.

Who this is for

  • Fashion ML engineers (LoRA / local VTON / VLM finetunes)
  • Fashion agent builders (MCP + CLI workflows)
  • Fashion companies customizing models for brand style and catalog quality

If you only need to call providers, start with InstallationCLIMCP instead.

Path overview (target)

data / prompts  →  train (LoRA)  →  eval / bench  →  invoke (CLI/MCP)  →  fashion workflow
StepGoalToday (v0.0.3)Toward v0.1.0
1. Data & promptsTrain packs + versioned prompt packsDataset loaders (VITON-HD, …)Schema, validate CLI, prompt packs
2. TrainBrand-style / VTON LoRATryOnDiffusion research codeopentryon train + recipes/notebooks
3. EvalGarment/identity qualityManual / ad-hocFashion Bench v0 + opentryon eval
4. InvokeSame path for cloud & localCLI + MCP registry+ productized local OSS VTON
5. WorkflowTask agents, not chatbotsEarly LangChain agentsTry-On QA or Fine-Tune Coach via MCP

What you can do now

pip install -U opentryon
# local/GPU models when needed:
pip install -U "opentryon[local]"

opentryon vton --model flux-vto --help
opentryon generate --model p-image --help
# MCP: see getting-started/mcp
  • Cloud try-on / generate / edit / video / understand via one registry
  • Local extras: FLUX.2-dev Turbo, Kimi-VL, LLaVA-NeXT, BEN2
  • Docs per provider under API Reference

Planned package layout

tryon/train/       # recipes, configs, runners
tryon/eval/ # metrics, Fashion Bench, reports
tryon/workflows/ # agentic fashion graphs
tryon/prompts/ # versioned prompt packs
tryon/datasets/ # loaders + train packs / cards

Suggested reading order

  1. Roadmap — next slices A–F
  2. CLI · MCP · Configuration
  3. New model checklist (when adding local/train adapters)
  4. VISION.md (product system)

Contribute

High-value PRs for this path: prompt packs, bench pairs, LoRA recipe hardening, first local VTON adapter, Try-On QA / Fine-Tune Coach workflow.

Join Discord or open a GitHub issue tagged for train/eval/workflows.