Skip to content

v0.0.6 — decide + embed services and the Oct 2026 model wave

Choose a tag to compare

View all tags
@kailashahirwar kailashahirwar released this 12 Oct 11:44
· 20 commits to main since this release
v0.0.6
e158f41
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

OpenTryOn v0.0.6 Release Notes

Two new services (decide, embed), a second TTS vendor, and the Oct 2026 model wave

Release Date: 12 October 2026

OpenTryOn v0.0.6 adds 42 registry models on top of v0.0.5 and two new service categories. Same CLI / MCP / registry surface — 132 models, 137 MCP tools (132 models + 5 discovery / key / planner), up from 90 models / 95 tools. Nine services: vton, generate, edit, understand, video-generate, bg-remove, tts, decide, embed.

What's New

New services

Service What it does Models
decide Typed, calibrated answers (yes/no, choice, score) about text / images / speech — no text output jev (Typesafe API), d1-3b, d1-omni-600m (Liquid AI, local), thx-01 (hosted), thx-01-local
embed Multimodal embeddings saved to .npz (--query adds similarity scores) embeddinggemma-2, pplx-embed-v2-late-0.6b, pplx-embed-v2-late-9b (all local)
tts (introduced after v0.0.5) Text-to-speech eleven-v4, eleven-v4-turbo, gemini-3.8-flash-tts, gemini-3.8-flash-lite-tts

decide and embed are named-model-only in the planner (no chat intent, no default changes).

Virtual try-on

Model Path CLI
FASHN VTON v1.5 Local, Apache-2.0, maskless pixel-space fashn-vton-1.5
Qwen-Image-2.1 DashScope + local Diffusers (also on generate / edit) qwen-image-2.1-pro / -turbo / -local / -turbo-local

Image generate / edit

Model Path CLI
Nano Banana 2.1 Gemini API (gemini-nano-banana-2.1); adds 1:4 / 4:1 / 1:8 / 8:1 nano-banana-2.1
FLUX 3 Image BFL API, up to 4K, up to 10 reference images on edit flux3-image
Qwen-Image-2.1 See above qwen-image-2.1-*

Video

Change CLI
Cosmos 3 Nano local (Diffusers, BF16/CUDA) cosmos3-local
H3 Max Relight (Fal; video + lighting sphere) fal-h3-max-relight
H3 Max Lip Sync (Fal) fal-h3-max-lipsync
P-Video-2-Pro (Pruna) p-video-2-pro
Veo 3.1 Generate / Fast / Lite, 4K, --last-image, --reference-image veo
Gemini Omni 1.1 Flash with --resolution and --video (extend) gemini-omni
Seedance 2.5 reference-to-video (--reference-image/-video/-audio) seedance

Understanding

Model Path CLI
Gemini 3.8 Flash Image, video, audio, PDF in; --thinking-level gemini-3.8-flash
Claude Haiku 5.5 Anthropic Messages API (text + image) claude-haiku-5-5
DeepSeek-V4.1-Flash DeepSeek Platform (image only) deepseek-flash
DeepSeek-VL2 / DeepSeek-OCR Local only deepseek-vl2 / deepseek-ocr
GLM-5.3-FlashX Z.ai glm-5.3-flashx
Qwen3.8-Omni-Flash DashScope (adds audio) qwen3.8-omni-flash
Ternary Bonsai 2 27B Local llama.cpp / MLX server ternary-bonsai-2-27b
Limite 1B - Violetto Local; deliberately out of fashion/media scope limite-1b-violetto

Breaking / behaviour changes

  • Sora is deprecated. OpenAI removed the Sora video API on 24 Sep 2026. --model sora is kept for backwards compatibility but emits a DeprecationWarning and live calls fail.
  • Planner default video model is now veo (was sora).
  • gemini-omni now targets gemini-omni-1.1-flash; the old gemini-omni-flash-preview id (deprecated 30 Sep 2026) is rejected.
  • veo is Veo 3.1 only; the Veo 3.0 ids (shut down 30 Jun 2026) are rejected.
  • nano-banana-2 is kept, but Google deprecated its model id (gemini-3.1-flash-image, no shutdown date yet) — prefer nano-banana-2.1.
  • New core dependency: anthropic>=0.75.0. New [local] dependency: thx01>=0.1.0 (for thx-01-local).

Install / Upgrade

pip install -U opentryon
# optional local/GPU models (Leffa, CatVTON, Qwen-Image, Cosmos 3, H3, LTX, Wan 2.2, …):
pip install -U "opentryon[local]"

Several local models need libraries newer than the shared [local] pins, so install them in a separate environment: Liquid d1 (transformers>=5.15), pplx-embed-v2-late (sentence-transformers>=6.0.0), EmbeddingGemma 2 (sentence-transformers), and FASHN VTON 1.5 (pip install "git+https://github.com/fashn-AI/fashn-vton-1.5.git", git-only).

From source:

git checkout v0.0.6
pip install -e .

Quick examples

opentryon generate --model nano-banana-2.1 --prompt "Panoramic lookbook banner" --aspect-ratio 4:1
opentryon vton --model fashn-vton-1.5 --person-image person.jpg --garment-image shirt.jpg   # local GPU
opentryon understand --model gemini-3.8-flash --video runway.mp4 --prompt "Summarise the looks"
opentryon understand --model claude-haiku-5-5 --image garment.jpg --prompt "Fabric, colour, fit?"
opentryon video-generate --model veo --prompt "Transition" --image a.jpg --last-image b.jpg --duration 8
opentryon video-generate --model fal-h3-max-relight --video runway.mp4 --reference-image sphere.png
opentryon tts --model gemini-3.8-flash-tts --text "Welcome to the spring collection." --voice Kore
opentryon decide --model d1-3b --questions questions.json --image dress.jpg                   # local GPU
opentryon embed --model embeddinggemma-2 --text "red gown" --image dress.jpg --query "formal dress"

Docs

Notes

  • Restart MCP after upgrade so Studio and other clients see the new tools. TryOn Studio has no capability screens yet for decide or embed; the tools work from any MCP client and the Agent chat (named model).
  • Typesafe Jev is early access (sign-ups were reported paused on 2026-09-22). Several local models (Liquid d1, EmbeddingGemma 2, pplx-embed-v2-late, FASHN VTON 1.5, Cosmos 3 local) were integrated without a GPU, so their weight loading was not exercised end to end.
  • Not integrated: NVIDIA PivotOPD (a training method with no released checkpoints).
  • 132 models across services → 137 MCP tools.

Links

Assets 4
Loading