styling.help — Devpost submission draft

Submission fields

Project name: styling.help

Tagline: Search it. Try it on. Choose the look. Step into the scene. Or simply ask when you need styling help.

One-sentence description: styling.help is a two-speed fashion platform: the working Guided Mode leads anyone from “What should I wear?” to a complete look, while Advanced Search is being built as a subscription try-on engine where power shoppers can search across retailers, preview a limited number of eligible products free, choose a favorite, and turn it into a social-ready scene with Gemini image generation.

Positioning: Guided styling for anyone. The ultimate try-on search engine and creative studio for power shoppers.

Vision: Search the world of fashion, see yourself in what you find, and step into the scene.

Implementation status: Guided Mode (“Style me”) works end to end in the current prototype. Advanced Search (“Search fashion”), free-try metering, subscription access, saved try-on selection, Gemini/Nano Banana scene generation, and social export are in active development; planned capabilities are labeled throughout this draft.

Project link: https://styling.help

Project summary

styling.help is one fashion intelligence platform with two levels of control and a new creative payoff.

When someone wants the system to lead, Guided Mode turns an occasion and desired feeling into clarified preferences, Google-grounded styling directions, retailer-linked candidate products, a YouCam Apparel VTO preview, conversational refinement, and a professional-ready brief.

When someone already knows what they are hunting for, Advanced Search becomes the power-user experience: describe the item naturally, search across retailers, understand why each result matches, control the ranking through explicit feedback, and try on eligible products with YouCam. A limited free allowance lets users experience the value before a subscription unlocks continued advanced try-ons.

The journey does not have to end at “this garment looks promising.” The planned Scene Studio lets the user choose a successful try-on, select or describe an aspirational setting—an editorial rooftop, a tropical coastline, an alpine resort, or another imaginative location—and use Gemini image generation, commonly known as Nano Banana, to create a labeled social-ready image. The output is a creative artifact, not evidence that the user visited that location or that the garment physically fits.

These are not separate products for separate kinds of people. The same shopper might need guidance for a wedding, use Advanced Search to find the exact blazer that completes the direction, and then create a share-worthy image of the final choice. The user chooses how much control they want; preference state, product evidence, visualization, and creative output carry the journey forward.

The strategic loop is: understand an intention, retrieve candidates, explain the match, visualize promising garments, learn from Save, More like this, Not for me, and pinned choices, choose the winning try-on, then transform it into something worth sharing. That is more valuable than either an undifferentiated product grid or a one-shot virtual try-on effect.

Inspiration

Most fashion products force shoppers into one of two extremes. A catalog search assumes they already know the right garment terms and filters. An automated styling flow assumes they want the system to make every decision.

Real people move between those states. For an unfamiliar or consequential event, they may ask, “What will help me feel right in this room?” For a specific purchase, they may know exactly what they need: “A relaxed navy blazer under £250 for an outdoor wedding.”

Even a precise request can lead to fragmented retailer tabs, inconsistent filters, duplicated listings, uncertain availability, and product images that make it difficult to imagine the garment on the shopper.

Virtual try-on solves an important part of that uncertainty: “Can I see myself in this?” But fashion is also identity and aspiration. After choosing a look, people often want an image that expresses the story around it: “What would this feel like on a Tokyo rooftop, a Mediterranean coast, or at the event I am imagining?”

styling.help meets the user at every level:

Guided Mode — “Style me.” Help me figure out what to wear.
Advanced Search — “Search fashion.” I know what I want; give me the ultimate search-and-try-on controls.
Scene Studio — “Create the moment.” I chose the look; help me turn it into a social-ready creative image.

Our goal is a personalized fashion engine that starts with human intent rather than catalog structure, explains why a result belongs, uses YouCam as personal visual evidence, and gives the user a controlled creative payoff after the decision is made.

What it does

Guided Mode — implemented

The current Style me journey works end to end:

Describe the moment. Choose or upload a portrait, then enter the occasion and desired feeling in everyday language.
Clarify the decision. Gemini analyzes the validated portrait and request, surfaces material unknowns such as dress code, season, location, and budget, and makes assumptions visible.
Generate complete directions. Gemini uses Google Search grounding to compose distinct styling strategies with references and retailer-linked candidate products.
Try on the primary garment. Selecting a direction can send its primary garment and the portrait to Perfect Corp YouCam S2S Apparel VTO.
Refine the Living Look. Requests such as “less traditional—keep the jacket” become structured preferences, KEEP/AVOID constraints, aesthetic changes, and history. Refinement can select another existing candidate; when the primary garment changes, the client attempts a new VTO realization.
Export the thinking. The current state becomes a printable professional brief containing context, candidates, references, constraints, visualization, and unresolved questions.

The prototype also includes an in-memory professional-review queue where a reviewer can record a verdict, alteration guidance, notes, and sign-off.

Advanced Search — in active development

The Search fashion experience is the ultimate try-on search engine for shoppers who want direct control:

Search naturally. For example: “A relaxed navy blazer under £250 for an outdoor wedding.” A portrait is optional for search and needed only if the shopper requests virtual try-on.
Search across retailers. Retrieve normalized, retailer-linked candidates without requiring the user to choose a store first.
Rank personally. Apply declared preferences, hard exclusions, budget, current context, and saved feedback—not generic popularity alone.
Explain every match. Show why each result fits, which constraints it satisfies, and which commerce details remain uncertain.
Control the result set. Shortlist, compare, dismiss, request more like a result, and open the retailer page.
Try before subscribing. Meter a clearly disclosed limited allowance of eligible YouCam try-ons so the shopper can experience the core value.
Subscribe for continued access. After the free allowance, require an authenticated subscription entitlement for additional Advanced Search try-ons and premium creative features.
Retain completed try-ons. Save successful results to the user’s account rather than replacing a single transient image.
Choose the winner. Let the user mark one completed try-on as the preferred result for purchase consideration or creative export.
Learn and search again. Feed explicit reactions and pinned choices back into retrieval and ranking.

The dedicated Advanced Search UI, discovery/ranking API, account system, free-try ledger, subscription entitlement, saved result gallery, feedback actions, and per-result VTO control are not yet present in the current prototype. The size and reset policy of the free allowance have not been set and should not be invented in submission copy.

Scene Studio — in active development

After the user selects a preferred successful try-on:

Choose a scene. Select a curated setting or write an approved prompt for an aspirational or exotic location.
Use the chosen result. Send only the user-selected YouCam output—not every attempt—into the creative generation pipeline.
Generate with Gemini image generation. Use a pinned and verified Nano Banana model to place the user and chosen look into the scene while aiming to preserve identity and garment appearance.
Review before publishing. Let the user accept, regenerate, or discard the result.
Export for social. Produce labeled 4:5 and 9:16 renditions with download and native-share controls.
Preserve truth and provenance. Mark the image as AI-generated, retain provider/job metadata, and never imply the depicted location or event actually occurred.

The current code does not yet call Gemini for image output, pin a Nano Banana model, generate scenes, persist preferred try-ons, render social aspect ratios, or download/share a generated image.

Why all three experiences belong in one engine

Guided styling, advanced product search, and creative scene generation solve different moments in one fashion decision.

Guided Mode establishes context the shopper may not have articulated: occasion, dress code, aesthetic direction, budget, exclusions, and confidence signals. Advanced Search can use that state to retrieve better products without making the user rebuild every filter. Search feedback can improve the next guided recommendation. Scene Studio gives the final selected look a controlled self-expression outcome.

The shared engine is designed to:

Understand an imprecise fashion intention.
Retrieve relevant products from multiple retailers.
Explain and personalize the ranking.
Visualize promising products on the shopper with YouCam.
Learn from explicit product and try-on choices.
Persist the preferred successful result.
Transform that chosen result into a labeled aspirational image with Gemini image generation.
Search, style, or create again using the evolving preference state.

The provider roles remain deliberately separate:

YouCam is the garment-decision layer. It previews an eligible fashion product on the user.
Gemini image generation is the creative-publication layer. It places the selected result into an imaginative scene after the user has chosen it.

A generated scene is not additional evidence of garment fit, product accuracy, or real-world travel. It is an opt-in creative image.

How we built it

The implemented Guided Mode is a React 19 and TypeScript client built with Vite. A single Express and TypeScript server owns the API, provider integrations, uploads, traces, in-memory state, and built client.

Gemini consultation, vision, and semantic state

We currently use the pinned @google/genai SDK and Gemini Interactions API for three implemented jobs:

analyzing the request and validated portrait input to produce known context, visible observations, explicit assumptions, and clarification questions;
composing structured styling directions and candidate products with Google Search enabled;
translating natural-language feedback into the smallest supported semantic mutation against existing directions and products.

Current Gemini requests use images as inputs and return structured text/JSON. They do not generate image output and should not be described as Nano Banana integration yet.

The central object is the Living Look. Instead of treating every interaction as a disconnected prompt, it carries the selected direction, candidate products, aesthetic axes, preserved constraints, negative constraints, references, and trajectory history through the current in-memory session.

Advanced Search will extend this state model with a direct product query, normalized results, ranking evidence, shortlist and comparison state, explicit shopper-feedback events, persisted try-on attempts, and a preferred-result reference.

Google Search grounding and product evidence

In Guided Mode, Gemini uses Google Search grounding to research occasion guidance, editorial references, and retailer-linked garment candidates. The server captures Search queries and citations for an Operations Trace, then normalizes the returned directions and products into stable application state.

Today, candidate ordering and the primary-garment choice come from the model response. The application does not yet run an independent relevance-ranking or deduplication pipeline. Advanced Search will add explicit ranking signals, hard-filter evaluation, cross-retailer normalization, result explanations, and feedback-driven reranking.

Search evidence improves discovery but does not guarantee live stock, current price, authenticity, retailer coverage, image rights, or VTO compatibility. Those distinctions remain visible in the product and submission.

YouCam apparel visualization

The current YouCam integration is a real server-side workflow rather than a static demo effect:

Resolve the selected direction’s primary garment and user portrait.
Register or upload provider assets when needed.
Submit a YouCam S2S v2.0 cloth task.
Poll asynchronously for success, failure, or timeout.
Return the provider result, task ID, and trace evidence for the before/after experience.
Preserve the original portrait and show a clearly labeled unavailable/fallback state when the provider cannot complete.

Advanced Search will reuse this realization service behind an account, eligibility check, transactional allowance or subscription-entitlement check, and per-result Try it on action. Successful outputs will become immutable try-on records so the user can compare attempts and select a preferred result.

Current product rows do not expose individual YouCam actions, and the current client retains only one transient Living Look result.

Gemini aspirational scene generation — in active development

The planned scene pipeline begins only after the user selects a successful YouCam result:

Validate user ownership, consent, try-on status, and scene entitlement.
Accept a curated scene choice or safety-checked user prompt.
Send the selected result and scene instruction to a pinned Gemini image-generation model in the Nano Banana family.
Aim to preserve the user’s identity and the chosen garment while changing the environment and editorial treatment.
Moderate the prompt and output, reject deceptive or unsafe uses, and let the user review before export.
Store model, provider-job, prompt-policy, consent, and generation metadata.
Render visibly labeled 4:5 and 9:16 social formats for download or native sharing.

No exact Nano Banana model is named under Built with until the source pins it and a live trace verifies image output. Identity and garment fidelity are product goals, not guaranteed outcomes.

Subscription and try-on allowance — in active development

The planned Advanced Search business model is limited free try-ons, then subscription:

Each authenticated user receives a clearly displayed introductory allowance for eligible Advanced Search try-ons.
Every accepted YouCam job consumes usage through a durable, transactional ledger with idempotency protection.
Failed or provider-rejected jobs follow a documented refund policy rather than silently consuming value.
After the allowance is exhausted, a verified subscription entitlement unlocks continued Advanced Search try-ons and premium Scene Studio access, subject to a transparent usage policy.
Checkout, recurring status, cancellation, renewal, and entitlement changes are driven by verified server-side webhooks.
Provider-cost and abuse controls protect the service without hiding limits from the user.

There is no consumer subscription, payment provider, webhook, entitlement service, account, or quota ledger in the current build. The existing in-memory professional-review request is a separate prototype and is not subscription billing.

Explainable personalization and feedback

Guided Mode already turns natural-language refinement into structured changes and can select among products returned in the existing consultation. It does not run a fresh search or calculate a new product ranking.

Advanced Search will make feedback an explicit discovery control:

Save records positive product intent.
More like this identifies which attributes should influence another retrieval.
Not for me captures a negative signal without inventing a reason.
Compare exposes trade-offs instead of hiding them in one score.
Select this try-on marks the successful result allowed into Scene Studio.
Pinned choices and exclusions become durable ranking inputs.

The saved-feedback, preferred-result, and reranking layers are planned, not yet implemented.

Operations trace, provenance, and professional handoff

The current Operations Trace exposes recent Gemini, Search, application-state, and YouCam events with evidence such as queries, citations, latency, image modalities, and provider task IDs. It makes live provider work distinguishable from fallback behavior.

The expanded trace will need to cover user entitlement, try-on usage, payment-webhook state, preferred-result selection, Gemini scene jobs, moderation decisions, and export provenance without leaking prompts, portraits, credentials, or payment data.

For Guided Mode, brief generation deterministically serializes the current context, constraints, references, candidates, visualization, and unresolved questions. The professional queue demonstrates the intended handoff, while authentication, durable storage, payment, assignment, and feedback delivery remain future work.

Business model — in active development

The subscription is attached to the high-cost, high-control Advanced Search experience rather than making the entire product a paywall.

Experience the value first: users receive a limited number of eligible Advanced Search try-ons at no charge.
Subscribe for continued creation: subscribers receive continued try-on access under a transparent allowance or fair-use policy, saved search and try-on history, preference learning, and Scene Studio access.
Keep search useful without VTO: product discovery, explanations, and retailer links should remain valuable even when a product is incompatible or the user has no remaining try-ons.
Separate optional professional value: professional review can remain a separate service rather than being confused with the consumer subscription.

Exact prices, free-try counts, reset periods, and subscriber limits are intentionally omitted until provider economics and user testing support defensible numbers.

Challenges we ran into

Making guidance useful without taking away control

A short request such as “make me look good” is emotionally clear but operationally incomplete. The challenge was to identify only the decisions that materially affect the recommendation, explain why each question matters, and preserve enough control that the user can redirect the result.

Combining grounded research with stable product state

Search-assisted responses must become reliable UI objects. We separate structured directions and candidates from grounding evidence, normalize the result, retain source links, and expose the research trace. We also avoid treating a Search-backed suggestion as a verified commerce record.

Orchestrating asynchronous virtual try-on

YouCam tasks require image validation, provider asset handling, task submission, polling, timeout behavior, and truthful provider-state reporting. Local portraits and remote garment images arrive in different forms, and an unavailable provider must not become a fabricated transformation.

Building power-search controls without losing trust — in progress

Advanced Search introduces a different class of problem: normalize and deduplicate products across retailers, distinguish hard filters from preferences, calculate an explainable relevance rank, preserve negative feedback, update results without creating a filter bubble, and verify enough commerce evidence to earn trust.

Metering paid generative workflows correctly — in progress

A try-on allowance cannot be a browser counter. It requires authenticated identity, transactional usage records, idempotent provider jobs, failed-job refund rules, verified payment webhooks, entitlement transitions, rate limits, and cost controls that remain correct across retries and multiple server instances.

Preserving identity and garment intent across two AI providers — in progress

The selected YouCam output becomes input to a separate creative generator. Scene creation must preserve the person and chosen fashion direction closely enough to remain recognizable while clearly separating creative imagery from fit evidence. Prompts and outputs need moderation, user consent, provenance, deletion controls, and labels that remain attached to downloaded media.

Accomplishments that we’re proud of

Built Guided Mode end to end: plain-language input, multimodal clarification, Search-grounded directions, retailer-linked candidate products, optional primary-garment VTO, semantic refinement, and a printable brief.
Integrated YouCam through real image registration/upload, asynchronous task creation and polling, task evidence, and visibly distinct fallback behavior.
Made assumptions, Search evidence, model actions, and provider work inspectable instead of presenting AI confidence as certainty.
Modeled styling as evolving state with explicit constraints rather than a sequence of disposable prompts.
Allowed refinement to select another existing candidate and rerun VTO when the primary garment changes.
Established reusable foundations for Advanced Search and Scene Studio: structured products, retailer links, preference constraints, a realization endpoint, generated-content labeling, and an operations trace.

Advanced Search, subscriptions, free-try metering, persisted try-on selection, Nano Banana scene generation, and social export are not listed as completed accomplishments until each is working and verifiable.

What we learned

Different shoppers do not need different products—they need different levels of control. The same person may want the system to lead for an unfamiliar event and want advanced search controls for a product they already understand.

The hardest fashion-search problem is not query syntax. It is preserving the meaning behind the query: the room someone is entering, how they want to feel, what they refuse to compromise on, and what previous results taught us.

Search grounding and virtual try-on solve different parts of uncertainty. Search can expand and explain the candidate set. YouCam can make one promising option personally legible. Neither automatically proves stock, value, fit, comfort, or suitability.

Our next product hypothesis is that choosing a try-on and creating a scene are two different acts. Selecting a try-on supports a fashion decision. Turning that selection into an aspirational image supports self-expression and sharing—and therefore needs stronger consent, moderation, and truth-label requirements.

What’s next for styling.help

Add authenticated accounts and durable private storage for preferences, search sessions, and try-on results.
Build the Advanced Search query API, result experience, cross-retailer normalization, deduplication, ranking, and feedback controls.
Add per-result YouCam eligibility checks and immutable try-on records.
Implement a transactional free-allowance ledger with idempotency, failed-job refund policy, and visible usage state.
Integrate subscription checkout, verified webhooks, recurring entitlements, cancellation, and renewal handling.
Add try-on history, comparison, and explicit preferred-result selection.
Pin and integrate a Gemini/Nano Banana image-generation model for safety-checked aspirational scenes.
Add labeled 4:5 and 9:16 rendering, download, native sharing, and portable provenance metadata.
Add scene consent, prompt/output moderation, identity and garment-fidelity evaluation, retention/deletion controls, rate limits, and provider budget controls.
Strengthen product evidence with field-level citations and independently verified price and availability where possible.
Add automated API, entitlement, billing-webhook, provider-failure, moderation, and browser end-to-end tests.

Built with

The stack below powers the implemented Guided Mode:

React 19
TypeScript
Vite
Node.js 22
Express
@google/genai
Gemini Interactions API for multimodal reasoning and structured text/JSON output
Google Search grounding
Perfect Corp YouCam S2S Apparel Virtual Try-On
MediaPipe Tasks Vision
Docker
Google Cloud Run-compatible deployment

Planned/in-development additions:

Authenticated identity and durable private storage
Consumer subscription checkout, verified webhooks, and entitlement ledger
Transactional try-on allowance metering
A pinned Gemini image-generation/Nano Banana model
Scene moderation, provenance, social rendering, download, and native sharing

Nano Banana is not listed as currently built until a concrete model call exists and a live trace verifies image output.

Current implementation boundaries

For an accurate submission and demo:

Guided Mode is implemented; Advanced Search and Scene Studio are in active development.
There is no consumer account, free-try ledger, subscription, payment integration, recurring entitlement, or persisted usage state.
There is no dedicated advanced-search UI/API, independent ranking pipeline, shortlist, comparison view, product-feedback store, or per-result VTO control.
There is no saved try-on history or preferred-result selection; the current client retains one transient Living Look result.
Current Gemini calls accept image input and return structured text/JSON. They do not generate images, invoke a pinned Nano Banana model, replace backgrounds, or create scenes.
There is no social image renderer, 4:5/9:16 export, download/share path, or scene-specific moderation and provenance workflow.
The implemented Gemini flow uses validated portrait pixels as multimodal input during clarification and research. Refinement can also receive a successful VTO result as visual context.
Current product candidates are retailer-linked suggestions returned during grounded styling research. The application does not independently guarantee cross-retailer coverage, deduplicate listings, or calculate a relevance rank.
Product links, prices, availability, authenticity, image rights, and compatibility require verification; “observed” does not mean guaranteed current.
YouCam currently visualizes one selected direction’s primary garment. It does not predict size, physical fit, comfort, tailoring, or exact drape.
Natural-language refinement can update constraints and aesthetic axes or select an existing direction/product. It does not retrieve a new catalog or rerank results. When an existing primary garment changes, the client attempts another VTO realization.
Briefs, traces, review requests, and professional feedback are stored in process memory.
The professional-review queue has no real payment processing and is separate from the planned consumer subscription.

Suggested gallery captions

Use these for screens that exist in the current Guided Mode:

Guided Mode: start with the outcome — Begin with a desired feeling and occasion instead of catalog terminology.
Clarify before recommending — Gemini surfaces material unknowns, portrait observations, and explicit assumptions.
Grounded directions, visible evidence — Styling strategies and retailer-linked candidates are paired with Search evidence in the Operations Trace.
Provider-labeled virtual try-on — A successful YouCam task powers the before/after comparison, while unavailable states remain visibly distinct.
A Living Look, not a disposable prompt — KEEP/AVOID constraints, product selection, aesthetic axes, and history preserve the conversation.
Designed for professional judgment — The printable brief packages context, evidence, candidates, visualization, and unresolved questions.

Capture these only after Advanced Search, subscriptions, and Scene Studio are implemented and verifiable:

Choose your level of control — Get guided with Style me or take over with Search fashion.
Power-search fashion naturally — A detailed human request becomes explainable cross-retailer matches.
Try before subscribing — The interface clearly displays and decrements the introductory try-on allowance.
Choose the winning try-on — Compare persisted YouCam results and select the one allowed into Scene Studio.
Step into the scene — Place the chosen look into a user-approved aspirational setting with Gemini image generation.
Created with AI — Download or share a social-ready image with a visible label and retained provenance.

Before publishing

Replace all three URL placeholders at the top.
Verify the implementation-status line immediately before submission. Remove “in active development” only after the relevant UI, APIs, identity, persistence, payment, and provider paths are working.
Do not screenshot or narrate Advanced Search, subscriptions, preferred-result selection, Scene Studio, or social export as live before they exist.
Demonstrate transactional free-allowance decrement and failed-job handling before claiming limited free tries.
Demonstrate a verified payment webhook granting and revoking subscription entitlement before claiming paid access.
Verify persisted selection among multiple successful try-ons before claiming the user can choose a favorite.
Pin the exact Gemini image-generation model in source and show provider evidence before naming Nano Banana under Built with.
Test scene prompt/output moderation, user consent, deletion, identity preservation, garment fidelity, and provider failure states.
Preserve a visible AI-generated label and provenance metadata in exported media.
Never imply that a generated location, event, body appearance, garment fit, or product result is physically real.
Confirm portrait, product-image, generated-output, location-reference, logo, and music rights before publication.
Record the current flow in DEMO.md and verify that any shown VTO run has a successful youcam_vto provider state and task evidence.
Call current products “retailer-linked suggestions from grounded research,” not comprehensive or verified cross-retailer search results.
Keep .env, credentials, project identifiers, personal information, sensitive raw traces, and developer tools out of captures.

Launch-ready description after Advanced Search, subscriptions, Scene Studio, and social export ship

styling.help is the ultimate personalized try-on search engine: search fashion across retailers, preview a limited number of eligible products free, subscribe for more, choose your winning YouCam result, and turn it into a labeled social-ready scene with Gemini image generation.

Built With

antigravity
gcp
gemini
javascript
node.js
npm
youcam

Styling.Help