Skip to content

v0.0.7 — FASHN catalogue endpoints, fashion models, and the classify service

Latest
Latest

Choose a tag to compare

View all tags
@kailashahirwar kailashahirwar released this 13 Oct 12:45
· 5 commits to main since this release
v0.0.7
f37f011
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

OpenTryOn v0.0.7 Release Notes

FASHN catalogue endpoints, fashion-specific models, and a new classify service

Release Date: 13 October 2026

OpenTryOn v0.0.7 adds 16 registry models on top of v0.0.6 and one new service category, classify. Same CLI / MCP / registry surface — 148 models, 153 MCP tools (148 models + 5 discovery / key / planner), up from 132 models / 137 tools. Ten services: vton, generate, edit, understand, video-generate, bg-remove, tts, decide, embed, classify.

This release is about fashion-specific work: catalogue photography endpoints, fashion retrieval and tagging models, garment extraction, a commercially clean local try-on, and a gate for garments before try-on.

What's New

FASHN catalogue endpoints (hosted, same FASHN_API_KEY as try-on)

CLI model Service What it does
fashn-product-to-model edit Product photo → a person wearing it (optional pose/background/face references)
fashn-model-swap edit Change the model's identity, keep the outfit and pose
fashn-face-to-model edit Headshot → try-on-ready upper-body avatar
fashn-reframe edit Expand or crop to a target aspect ratio
fashn-edit edit Prompt edit with optional mask / context image
fashn-model-create generate Fashion model photo from a prompt
fashn-image-to-video video-generate Animate a still, 5 or 10 s
fashn-bg-remove bg-remove Transparent-background cut-out

Product-to-model generates a new person and does not replace try-on (fashn-tryon-max). Integrated from FASHN's published docs without live calls — confirm credit costs on your account.

Fashion models (local, open weights)

Model Service What it does Licence
marqo-fashionsiglip, moda-fashion-distilled (image-only), moda-pro-lite, zooclaw-fashionsiglip2 embed Fashion catalogue search by text or by photo Apache-2.0 / MIT
fashion-florence understand Garment photo → JSON {category, primary_color, material, style_tags} MIT
qie-2511-extract-outfit edit Worn outfit photo → flat studio mockup (garment extraction) Apache-2.0 LoRA on Qwen-Image-Edit-2511
klein-4b-tryon vton Try-on LoRA on FLUX.2-klein-4B; person (clothing-agnostic) + top + bottom Apache-2.0 base and LoRA

New classify service

Model What it does
tryon-safety-classifier Genlook's SigLIP2 classifier: from all photos of one product, what a try-on would show (level ok/revealing/lingerie/adult, blocked, subject, swimwear, ravewear). Re-implements the model's inference instead of running its repo code; weights pinned to tag v1.0. Apache-2.0

classify is named-model-only in the planner (no chat intent, no default change).

Breaking changes

  • Removed 16 deprecated root-level wrapper scripts — bg_remove.py, caption_images.py, gpt_image.py, image_gen.py, luma_image.py, run_ootd.py, sora_video.py, veo_video.py, video_gen.py, vton.py, fashion_agent.py, model_swap_agent.py, vton_agent.py, luma_video_test.py, test_gpt_image.py, test_luma.py. They only printed a deprecation warning and forwarded to copies under examples/legacy/, examples/agents/ and tests/legacy/. Anyone running python vton.py (etc.) from a checkout now gets file-not-found: use the opentryon CLI (for example opentryon vton --model ...) or run the copies under examples/ and tests/legacy/ directly. main.py, run_demo.py and api_server.py are real scripts and stay. The installed opentryon package never shipped these root scripts, so pip users are unaffected.
  • No changes to existing model ids, flags or defaults, and no new required dependencies (install_requires is unchanged from v0.0.6).

Install / Upgrade

pip install -U opentryon
# optional local/GPU models (Leffa, CatVTON, Qwen-Image, klein try-on, Fashion Florence, …):
pip install -U "opentryon[local]"

Several of the new local models need libraries beyond the shared [local] pins — install them in a separate environment:

Model Extra install
marqo-fashionsiglip, moda-* pip install open_clip_torch
tryon-safety-classifier pip install open_clip_torch safetensors huggingface_hub
qie-2511-extract-outfit pip install -U peft diffusers transformers accelerate
klein-4b-tryon pip install git+https://github.com/huggingface/diffusers.git peft accelerate (Flux2KleinPipeline is not in a stable diffusers release yet)
fashion-florence none beyond [local]; loads with trust_remote_code against the transformers 4.4x line it pins

From source:

git checkout v0.0.7
pip install -e .

Quick examples

opentryon edit --model fashn-product-to-model --image jacket.jpg --prompt "studio background" --resolution 2k
opentryon edit --model fashn-model-swap --model-image look.jpg --prompt "woman with red hair"
opentryon video-generate --model fashn-image-to-video --image look.jpg --duration 5 --resolution 720p
opentryon embed --model marqo-fashionsiglip --image catalogue/*.jpg --query "black leather ankle boots"
opentryon understand --model fashion-florence --image dress.jpg
opentryon edit --model qie-2511-extract-outfit --image model_wearing_look.jpg
opentryon vton --model klein-4b-tryon --person-image person_agnostic.png --top-image shirt.png --bottom-image jeans.png
opentryon classify --model tryon-safety-classifier --image front.jpg back.jpg

Docs

Notes

  • Restart MCP after upgrade so Studio and other clients see the new tools. TryOn Studio has no capability screens yet for decide, embed or classify; the tools work from any MCP client and the Agent chat (named model).
  • The new local models were integrated without a GPU or real weights: they are tested against stubbed pipelines, tokenizers and safetensors files, not end-to-end runs. Vendor-reported benchmark numbers in the docs are not reproduced here. Run a quick sanity check on your own data before relying on them.
  • klein-4b-tryon is a v1 community LoRA (273 training tuples, no published metrics) and does no clothing segmentation: the person image must be clothing-agnostic, or pass --mask-image.
  • The safety classifier is one layer of moderation: it classifies products (not people), was trained mostly on western e-commerce photos, and the right threshold depends on your audience.
  • 148 models across services → 153 MCP tools.

Links

Assets 4
Loading