`undress-service`: The Repo That Looks Like a Wardrobe App and Behaves Like Two Projects

A self-hosted FastAPI fashion assistant, a local AI styling stack, and a completely separate downloader package living in the same tree.

8 min read • View on GitHub • More from GangTailorUpgrade

A wardrobe cabinet split down the middle. One side is organized with labeled clothing, weather tags, and a phone showing outfit suggestions. The other side is peeled back to reveal hidden cabling, proxy nodes, and downloader machinery. It explains the repository’s split identity: a clean fashion assistant on the surface and a second, unrelated toolchain underneath.
The repo’s public face is a local fashion assistant. Its internal shape tells a different story.
Key Takeaways

The Repo With Two Public Faces

At first glance, `undress-service` reads like a straightforward local AI wardrobe app. It promises a self-hosted way to catalog clothes, classify items, and generate outfit suggestions without sending everything to a third-party cloud. The surprise is that the repository also contains a separate package, `coomertool`, that points in a completely different direction.

Dress AI Sponsor.

GangTailorUpgrade, Project Maintainer · GangTailorUpgrade/undress-service

That tension is the whole story. The clean, consumer-facing wardrobe assistant is technically interesting on its own, but the second package changes how you read the repo. It stops being just a product and starts looking like a bundle of adjacent ambitions, some polished and some deliberately off-brand.

The Clean System: A Local-First Fashion Assistant

The main application lives in `app/`, where FastAPI, SQLAlchemy, and Pydantic form a conventional service architecture. Wardrobe items become database records, outfits become relationships, and the UI layer is backed by a set of APIs for ingesting photos, generating metadata, and returning styled recommendations.

The system is less a closet manager than a translation pipeline, turning a photo into structured wardrobe data and then back into a styled visual.

How the Wardrobe Brain Works

The ingestion path is pragmatic. An uploaded image is resized, dominant colors are clustered with K-Means, and CLIP does zero-shot classification against text prompts like a shirt, top, or jacket. The result is a structured wardrobe record with category, colors, season, and fabric attached to it.

A close-up assembly line made of three stations. A shirt photo enters one end, passes through a swatch-making color cluster, then a CLIP classification station, then an outfit builder that emits a styled look. In a side corridor, a separate conveyor belt disappears into another room. It explains how the app turns visual input into structured wardrobe data while hinting at a second unrelated subsystem.
The technical center is a translation chain: image in, wardrobe metadata out, styled result back in.

That pipeline matters because it avoids the usual fashion-app dependency chain. Instead of treating clothing analysis as a cloud service, the repo leans on local models and local storage. In this category, that is not just a deployment detail. It is part of the value proposition.

Why Local Models Matter Here

Self-hosting changes the economics and the trust model. Users do not have to upload private wardrobe photos to a vendor, and the maintainer does not have to pay per-request API costs for every classification or generated look. The tradeoff is maintenance, because local models raise the bar for setup and hardware.

ApproachData custodyModel hostingCost profileUX polishMaintainabilityEditorial risk
Cloud fashion SaaSImages leave the machineVendor-hostedRecurring subscriptionUsually polishedLow for the userLow
Self-hosted wardrobe assistantUser-controlledLocal or optional localHardware and ops burdenDepends on the installHigher for the userModerate
Single-purpose OSS repoUsually user-controlledUsually localMinimal direct costVaries widelyHigher, but narrowerLower
`undress-service`User-controlled on the surfaceLocal-first stackLocal compute plus setupMixed, because it is two projectsHigher, because the repo is splitHigh, because branding and scope diverge

The repo’s local posture also explains why the technical choices feel familiar to builders. CLIP gives broad zero-shot labeling without a custom dataset. K-Means gives a cheap, explainable way to extract color palettes. Generative models then turn that structured wardrobe state back into a visual preview.

The Shadow Package

Then there is `coomertool`. It is not a small helper module or a hidden utility. It is a separate scraper and downloader package sitting beside the fashion app, with its own code path and its own implied audience. That makes the repository harder to read and more interesting to analyze.

This is where the article stops being a simple architecture tour. The repo is technically competent, but its bundle of identities creates friction. One part says private wardrobe assistant. Another part says something else entirely. In open source, that kind of mismatch is rarely accidental enough to ignore.

What This Repo Says About Open Source Now

Bundled repos are common. Ambiguous branding is common too. What makes `undress-service` notable is that it turns those ordinary habits into a visible artifact. The codebase is not just a service. It is a record of how developers sometimes package adjacent tools, different users, and conflicting intentions under one roof.

Compared with polished cloud products, it is rougher and more revealing. Compared with clean single-purpose OSS projects, it is more confusing but also more honest about the messiness of real repo formation. That is why it deserves attention: not because the fashion assistant is novel, but because the repository tells on itself.