Skip to main content

Roadmap

Last Updated: 11 November 2026 · Current release: v0.0.9 · Next: v0.1.0 Fashion ML Toolkit Core

Canonical file: ROADMAP.md · Strategy: VISION.md

Audience priority: fashion AI/ML engineers · agent builders · app builders · fashion companies (train + MCP) · CLI-first developers.

v0.1.0 exit criteria: train pack → LoRA finetune → garment/identity eval + Fashion Bench v0 → invoke via CLI/MCP → one agentic fashion workflow — without assembling five research repos. Outline: Fashion ML Engineer Path.

Shipped — v0.0.3 – v0.0.9 (Phase 0: invoke layer)​

  • Unified CLI + FastMCP (shared registry / invoke_model)
  • Broad cloud try-on / generate / edit / video / understand adapters
  • v0.0.4: LTX-2.5 (API + local), Hailuo 2.3, Wan (API + local 2.2), Runway Gen-4.5, Qwen3.8 (API + local)
  • v0.0.5: dedicated VTON (Google Vertex, OutfitAnyone-Plus, Photoroom, Leffa, CatVTON), ChatGPT Images 2.5, MiniMax H3 / H3 Max + Fal H3 Max, NVIDIA NIM, Hy4, Muse Image, Qwen-Image, Wan 3.0, planner as registry super-agent
  • v0.0.6: new decide (Jev, Liquid d1, THX-01) and embed (EmbeddingGemma 2, pplx-embed-v2-late) services, tts (ElevenLabs, Gemini), Gemini 3.8 Flash, Claude Haiku 5.5, Nano Banana 2.1, FLUX 3 Image, Qwen-Image-2.1, FASHN VTON 1.5 (local, Apache-2.0), Veo 3.1 / Omni 1.1 / Seedance references, Cosmos 3 local, DeepSeek
  • v0.0.7: FASHN catalogue endpoints (product-to-model, model swap, face-to-model, reframe, image-to-video), fashion retrieval embeddings (Marqo-FashionSigLIP, MODA, ZooClaw), Fashion Florence tagger, Extract-Outfit, klein-4B try-on (Apache-2.0 stack), try-on safety classifier (new classify service); root wrapper scripts removed
  • v0.0.8: Clad body measurement (new measure service), FLUX 3 Image Klein (API + local) and FLUX 3 Video, NVIDIA LipSync (audio-to-video), TRELLIS (generate-3d), SANA-Sprint, Nemotron 3.5 Super VL, VideoJev, Gemini Embedding 2
  • v0.0.9: 20 decision models (Perplexity Decider, Cloudflare Clef, Kev, AutoTrust JEV, vLLM-SR D3, Mercury Decide), new robot-act (Dream-1, Cosmos 3 Nano-DROID), detect (Moondream) and parse (Nemotron Parse) services, Gemini 4 Argon, Gemini Omni 1.2 (default), Kimi K3.1 Pro, Moondream 3.1, Mercury 3, FLUX 3 Max, Kling 4.0, Seedream 5.0 Flash, Nemotron 3.5 Embed
  • OpenAPI / Postman media snapshots, docs, Gradio demos
  • Local extras (FLUX.2 Turbo, Kimi-VL, LLaVA-NeXT, BEN2, LTX-2.5, Wan 2.2, Qwen3.8, Qwen-Image, Leffa, CatVTON, MiniMax H3)
  • Web UI in TryOn Studio over MCP

Next — Fashion ML Toolkit Core → v0.1.0​

SliceFocus
A — Data & promptsVersioned prompt packs, train-pack schema, data CLI helpers
B — TrainBrand-style LoRA + one VTON/local LoRA path; opentryon train
C — EvalFashion Bench v0, side-by-side reports; opentryon eval
D — Local VTONLeffa + CatVTON shipped in v0.0.5 (opentryon[local])
E — WorkflowsMCP-native Try-On QA or Fine-Tune Coach (task agent, not chatbot)
F — Efficiency cardVRAM/latency table for the path we actually ship

v0.0.4 – v0.0.9 shipped as intermediate tags; v0.1.0 when A–E work end-to-end.

Fashion-only for this phase. Prompt datasets, fashion datasets, workflows, and agentic fashion workflows are in scope. Generic multi-domain / full LLM–VLM platform work waits until fashion patterns prove out.

Later​

  • Broader quantization / distillation / serving recipes
  • Full fashion agents suite (PDP, lookbook, router, return-risk, …)
  • Async / batch / caching DX · video VTON / 3D VTON
  • Adjacent domains only after fashion train/eval/agent recipes are solid

Contribute​

Good first issues: prompt packs, bench pairs, docs, dry-run tests.
High-value: LoRA recipes, Fashion Bench metrics, first local VTON, first MCP workflow.

Vendor / model queue (separate from these slices): Integrate next.

Follow the new-model checklist or join Discord.