New services backlog

These capabilities need a new CLI service (and usually a new output_kind) before model adapters land. Integrate one category at a time; start with platform plumbing (segmentation output), then the first model in that category.

Status key: next = ready to implement after design sign-off · watch = licence, API maturity, or scope blocker · skip = defer · shipped = in the registry.

---

measure — shipped

Id	Path	Notes
clad-body-photo	A	Clad API photo + height
clad-body-questionnaire	A	Clad eight-question flow
clad-body-local	B	clad-body on Anny/MHR params

Docs: Clad API, local.

---

1. segment — human / garment parsing (masks)

Service contract (target v1)

Field	Choice
CLI	opentryon segment --model <id> --image …
Method	segment
output_kind	segmentation (new) — indexed label PNG + labels JSON; optional per-class binary masks
MCP	Base64 PNG(s) + JSON metadata

Integration queue

Priority	Registry id (draft)	Path	Vendor / weights	Licence / notes	Status
1	fashn-human-parser	B	fashn-ai/fashn-human-parser (SegFormer-B4, 18 classes)	SegFormer / FASHN terms — confirm commercial use before default	watch (licence)
2	aitryon-parsing-v1	A	Alibaba DashScope OutfitAnyone parsing companion	Same DASHSCOPE_API_KEY as outfitanyone-plus; Beijing region	next after segmentation output_kind
3	segformer-b2-clothes	B	HF mattmdjaga/segformer_b2_clothes	Research / “other” licence	watch
4	alibaba-aitryon-parsing	A	Documented with aitryon-plus flows	Already referenced in OutfitAnyone-Plus	watch (alias of #2)
—	Leffa ONNX parsing	—	Internal to leffa adapter only	Not a registry id unless unified	skip (for now)

Downstream: klein-4b-tryon (--mask-image), inpainting edits, catalogue QA.

Blocker: Product/legal must pick a commercial-clean default parser (survey note in integrate-next).

---

robot-act — shipped (Dream-1)

Id	Path	Notes
dream-1	A	Dream Labs WAM via dreamlabs-client
dream-1-flash	A	Faster tier

Docs: Dream-1. No local Dream-1 policy weights (hosted API only).

---

Frontier MoE / long-context (understand)

See integrate-next — Frontier long-context MoE survey: Kimi-K3, Muse Spark Open, Ling-3.1-flash, Beam, Nemotron 3.5 Super VL.

---

2. detect — e-commerce object detection

Service contract (target v1)

Field	Choice
CLI	opentryon detect --model <id> --image …
Method	detect
output_kind	text (JSON file) in v1: {detections: [{label, score, bbox_xyxy}]}

Integration queue

Priority	Registry id (draft)	Path	Vendor / weights	Licence / notes	Status
—	moondream-3.1, moondream-3-preview, moondream-3.1-local	A / Photon	Moondream skills detect / point / segment	Moondream License. Open-vocab, not a fixed class list. SVG path, not a label PNG	shipped
1	trendyol-yolov9	B	Trendyol YOLOv9 (84 product classes, survey Oct 2026)	CC BY-SA 4.0 — share-alike for derivatives	watch (paper/weights GA)
2	trendyol-dinov2-embed	B	DINOv2 256-d embedder (same survey)	CC BY-SA — consider embed service instead if boxes not needed	watch
3	—	A	No first-party “fashion detect” API in Wave 2 table	Use understand VLMs for prose only	skip

Blocker: Stable HF repo + class list; NMS/threshold CLI flags frozen in docs.

---

Suggested category order

1. segment — highest leverage for VTON preprocess (after licence + segmentation output_kind).
2. detect — catalogue / agent automation.

---

Related (not in this doc)

• music-generate + Lyria 3.5 — separate from tts; see integrate-next survey line.
• 3D garments / sewing — pygarment / GarmentCode — large scope; own design pass.