Semantic Search, Agent-Ready API, Image Provenance, and Better Video Workflows: What's New at Uwear
Since launching Qwen Intimate, most of the backend work has had one theme: making Uwear feel more like production infrastructure for fashion teams and AI agents. Generated assets are easier to find, the API is easier for agents to use, video clips are more reusable, and the platform has better foundations for provenance, analytics, and reliability.
What shipped
- -Semantic search: Search generated product photos by meaning, not just file names or tags.
- -Agent-ready API: A simpler generation workflow, downloadable agent reference, API keys, polling, and signed webhooks.
- -Image provenance: Uwear now has the backend groundwork for invisible watermarking, C2PA metadata, and public verification.
- -Video workflows: Extract frames from videos, build richer Montage proposals, and export with better fit controls.
Search Generated Product Photos by Meaning
Fashion teams generate a lot of visual variants: clean studio photos, lifestyle scenes, close-ups, alternate angles, campaign tests, and video frames. File names and manual tags are useful, but they do not always match how people remember an image.
Uwear now has semantic search support for generation results. New generated assets can be indexed with Gemini embeddings and stored in pgvector, so a search can match the meaning of an image instead of only exact metadata. That means queries like "black dress on a city street", "close-up denim texture", or "studio swimsuit full body" can become much more useful inside a large generated library.
What changed under the hood
- -Generation result embeddings are stored in Postgres with pgvector.
- -New generation results queue indexing work asynchronously so search does not slow down generation completion.
- -Weak semantic matches are filtered out so search results stay useful.
This is especially important for teams that use Uwear as a production asset library. Once you are creating hundreds or thousands of AI product images, retrieval becomes part of the workflow.
A More Agent-Ready Fashion Generation API
Uwear's API is being simplified around a clearer agent-friendly contract: create a generation job, wait for the result through polling or webhooks, then download the generated assets. A new downloadable UWEAR.md reference gives coding agents the rules they need to work with the API directly.
What changed for API users and agents
- -Open REST API: Use bearer API keys against `https://api.uwear.ai`.
- -Unified generation route: Use `POST /generation` for images, edits, upscales, and videos with a `use_case` parameter.
- -Agent reference: Give UWEAR.md to ChatGPT, Codex, Claude, Cursor, or another coding agent.
- -Polling or webhooks: Poll `GET /generation/{generation_id}` for simple scripts, or receive signed `generation.completed` and `generation.failed` events for production automations.
- -Curated fashion models: Access image, edit, upscale, and video models chosen for apparel workflows through one API surface.
We wrote a dedicated agent-ready API guide for this, because the larger story is not just webhooks. It is Uwear becoming a practical AI fashion generation API that agents can discover, understand, and use.
Use the Agent Skill
Download the current UWEAR.md reference from this blog and give it to a coding agent to plan bulk collection generation, virtual try-on architecture, image edits, upscales, and video workflows with the public API.
Download UWEAR.mdImage Provenance and Public Verification Groundwork
As AI product imagery becomes more common, provenance becomes part of the infrastructure. Brands need a way to distinguish a generated or edited asset from a camera-original photo, and partners may need a way to verify where an asset came from.
Uwear now has the backend foundation for a Uwear-owned provenance path. The current approach is based on Uwear-applied markers, not provider-native signals: an invisible TrustMark watermark, a C2PA manifest assertion, a server-side provenance log keyed by the final asset hash, and a public verification endpoint.
What verification can return
- -Verified: Uwear markers are present and match a verified provenance log entry.
- -Not verified: No Uwear provenance signal was found.
- -Tampered: One or more Uwear signals exist, but full verification failed.
- -Unsupported: The submitted media type is not supported for verification.
This is technical provenance infrastructure, not a replacement for visible disclosure. Brands are still responsible for how generated images are displayed in product pages, ads, marketplaces, emails, and other shopper-facing channels. For the full compliance-oriented explainer, read the EU AI Act image provenance guide, or use the public Uwear image verifier.
Better Video and Montage Workflows
Uwear's video work is moving from isolated generations toward reusable production workflows. The goal is simple: a generated clip should not be a dead end. It should be something you can inspect, trim, reuse, combine, and turn into other assets.
Video workflow improvements
- -Frame extraction: Pull a still frame from an owned video generation result and save it as an image result with inherited tags and clothing context.
- -Montage proposals: Moùza can propose montages using a mix of existing videos and images that still need video generation.
- -Export fit mode: Montage export now supports `cover` for centered crop and `contain` for padded fit.
- -Pipeline reliability: Video and upscale steps now preserve target resolution, audio, prompts, and last-frame settings more consistently.
- -Thumbnails: Video generation thumbnails were cleaned up so generated clips are easier to scan.
If you are building a campaign from AI-generated fashion content, this matters more than any single video feature. The useful workflow is not "make one clip". It is "generate, review, reuse the best moment, combine clips, and export for the channel".
Camera Edits Are More Predictable
Camera angle control is now driven by a backend-owned catalog and shared camera semantics. That gives generation, edits, batch flows, API schemas, and Moùza a single source of truth for camera behavior.
A few important fixes landed with that: normal generation requests no longer accidentally infer camera-edit presets from a plain camera value, batch generation cameras are normalized more consistently, and edit prompts resolve camera edit IDs through the canonical backend catalog.
Moùza Is More Controllable
Longer agent turns should not depend on a single request surviving from start to finish. Moùza now has an async turn flow with durable turn state, event logs, polling, and websocket status events. That makes heavier agent work easier to resume and observe.
Stop support also shipped for active agent chat streams. If you interrupt a turn, Uwear saves a partial assistant draft where possible so the next message can continue from the stopped point instead of losing the conversation state.
Team, Admin, and Billing Improvements
- -Granular usage analytics. Company usage analytics can now break down generation activity by model pricing-rule settings, making team spend easier to understand.
- -Company negative credit policy. Admins can approve a controlled negative credit line for a company while still enforcing a central credit floor.
- -Verification approval emails. When a company is approved, Uwear can notify the owner and billing contacts without resending duplicate verified-to-verified emails.
- -GPT Image 2 quality variants. Backend model data now supports the new GPT Image 2 quality variants.
Reliability Fixes
A handful of backend fixes shipped alongside the larger platform work:
- -Shopper SMS OTP is more reliable. Twilio Verify is now available as a configurable SMS provider, with the Verify Service friendly name handled by Twilio instead of per request.
- -Tag removal now persists. Generation result tag removal now awaits async deletes correctly, so removing a tag actually removes the row.
- -Username conflicts show correctly. Duplicate shopper usernames now return the intended conflict response instead of being wrapped as an internal server error.
- -Sizing errors keep their structure. The HTTP exception sanitizer now preserves structured 4xx details while continuing to hide internal 5xx details.
- -Shared Qwen worker deploys are safer. Normal staging and production deploys are guarded from mutating the shared Qwen worker, with a dedicated deploy path for intentional worker changes.
- -Image size mappings were fixed. Backend payload size handling was cleaned up so model requests map dimensions more reliably.
All Platform Updates Are Rolling Into Uwear
Semantic search, an agent-ready API, provenance groundwork, video workflow improvements, and reliability fixes are part of the same direction: Uwear as production infrastructure for AI fashion imagery.
Questions or feedback? Reach out. We read everything.