Lookset
Lookset 由 ThreadTruth Studio 改名而来。
Development source: public releases of Lookset are published at https://github.com/denggui-ai/lookset (latest v0.1.0-beta.3). This repository holds the development history, tests and verification records.
AI Fashion Studio · 服饰 AI 影棚
0.1.0-beta.3 — plan what each image should show. Start with real garment and model references, choose among 36 visual entries, and plan with 52 actions across 60 reference views. Rank pose candidates around product goals and selected task history; review the plan before approving generation.
This release includes the shared 176-entry pose pool, knowledge and style-rule resolution, and exact or limited text-based near-repeat checks. The same runtime has been backed up, installed and checked on the maintainer’s Mac. Visual consistency, fresh-host behavior and actual rollback remain incompletely verified. Usage and limits · Offline guide.
单件或整套服饰，按服饰编排九张 AI 模特图，36 个视觉入口可选。Turn a garment or coordinated outfit into a garment-driven set of nine AI model portraits. Choose from 36 visual entries.
下载Lookset插件 · Download Lookset · 开始安装 · Install · 查看成片 · See results · 探索风格 · Explore styles
English guide · 中文说明 · 产品首页 / Website · 能力清单与拓扑 / Capability topology · 当前能力状态 / Capability status
When clear front and front-side photos are available but the back is missing, no back-view shot is planned; the set uses front, front-side and detail arrangements and states the limit. Accurate rear construction still requires a real rear photo.
Japanese-home workflow tutorial (中文) (six-image era): materials, first-image approval, six standalone images and an honest case review. The tutorial identifies released and development behavior separately.
One garment or a complete outfit, nine garment-driven shots
Upload a garment or coordinated outfit, choose one of 36 visual entries (24 style entries plus 12 execution entries), and let the plan lay out nine independent AI model portraits around what this garment needs to show, each with a stated purpose. A labeled nine-panel direction preview comes first; after you confirm the direction, images are generated one by one within the authorized count. Review and confirm the first image before the remaining eight, which reuse its identity reference for consistency within the set. This does not promise to reproduce a person in your input photo. Explicit requests for six images, the classic six poses or a single image are honored as asked; a choice of 36 entries does not mean 36 sets in one request.
Prepare your photos · Nine-shot planning and generation prompts · Garment: 24 previews (historical) · Outfit: 24 previews (historical)
Both collections are historical direction previews from the six-image era: one six-panel sheet per style, not independent finals. The six styling cases (a floral dress, a sage-shirt home outfit, two styles of a contrast-trim T-shirt, an external Codex user’s black-jacket outfit, and our same-input Office Commute set; 36 separate AI images) were all produced under the six-image flow, each with its own disclosed review scope. Public nine-shot cases will follow the entry-by-entry sweep.
Why create with Lookset?
Keep your real garment at the heart of every creative decision.
- Garment details you can review. Review color, construction and styling before generation, then compare each result with your source photos. See the reviewed cases.
- Explore more looks for one outfit. Find a direction across 36 visual entries, from ecommerce studio to fashion editorial. See one outfit in three styles.
- Preview the direction, then create nine images. Choose a direction, approve one nine-panel direction preview, confirm it, then create nine independent images. Explicit requests for six or a single image are honored as asked. Start with recognition.
What you can do
- Identify the garment. Read visible details from a single item, an outfit or multiple views.
- Choose a direction. Get recommendations or choose freely from all 36 visual entries; 207 knowledge records (source catalog, Chinese-market phrasing, Chinese cultural dress) sit beside the entries for the assistant to consult when a direction is ambiguous. They are not entries themselves.
- Set the shoot. Choose studio, location or hybrid, plus presentation and canvas. Alternate forms have limited public validation.
- Choose the output. Request one test, a labeled nine-panel direction preview, nine independent images (or an explicit six), or prompts only.
- Choose the route. Use Codex by default or request a ChatGPT web handoff; automation needs host browser tools.
- Review and refine. Compare with the source, check dimensions, and authorize a specific retry when needed.
Full capabilities, copyable prompts and verification scope →
Install Lookset 0.1.0-beta.3 in Codex. Generate there by default, or explicitly choose the optional ChatGPT web handoff described below.
Get started
What you need: macOS, Python 3, a terminal, and a Codex CLI that supports codex plugin add. This is the only tested setup; Windows, Linux, and fresh machines are unverified. The first recognition step generates no images. Generating images later needs a Codex account with native image generation, or a ChatGPT account with image access if you choose the web route. Fixed-pose wardrobe edits additionally need Node.js with the sharp module already installed on the host; without them that branch reports tool-blocked and nothing is installed automatically.
- Follow the English installation guide to download the 0.1.0-beta.3 Plugin ZIP and matching
.sha256, check prerequisites, verify the archive, and enable the Plugin. Use the named Plugin download, not GitHub's automatic source-code ZIP.
- Open Lookset in a new Codex task and say “hello”, or upload garment photos you may use. Include a person reference if you want to reuse one. No long starter prompt is needed.
- Lookset checks visible garment facts, available views and the intended use of person references, then helps prepare a shoot plan. You can say “Plan the shoot first; do not generate images yet.” Only missing information that affects the result needs clarification.
Uploading materials or confirming a plan does not authorize generation. Existing style, preview, independent-image and budget gates still apply. Reception instructions are statically checked; no new model behavior evaluation was run for this release.
If recognition does not appear, follow installation troubleshooting. Share a brief success or failure through the installation feedback form.
Choose where to generate
| Route | How it works | What you need |
|---|
| Codex · default | Generate within Codex after approval; review the first image before continuing the set. | A Codex host with native image generation and available quota. |
| ChatGPT web · explicitly selected | Codex prepares numbered references and prompts. Transfer them manually, or authorize Codex to operate a supported browser. | A ChatGPT account with image access and quota. Automatic execution also requires host browser tools for uploading and downloading original files. |
Install Lookset in Codex. The ChatGPT route uses the web handoff described in the tutorial (中文); installing this Plugin does not install browser tools. Uploading references to ChatGPT requires permission. Failed or uncertain requests are checked before any separately authorized retry, and used attempts remain recorded. Web workflow details (中文).
Examples and what they demonstrate
Six sets, 36 separate images. See all images and source notes. Dress and home-outfit visuals were accepted; the T-shirt sets retain fine-detail and action review. These AI-model images support styling previews, not personal identity, size or fit guarantees.
Product references → AI scene styling. The comparison shows the actual primary shirt and pants references alongside a home-scene result. The pants are displayed as a crop; additional construction and identity references were used. An external Codex user also completed six images and reported satisfaction; the version used was not recorded.
See the home Before / After, all six originals and source notes
One outfit input → six Office Commute images. The same authorized input also appears in the external trial. Compare two presentation directions, rather than versions; the external version is unknown. Slot 6 uses stationary front standing because no real rear photo was supplied. Fine garment details remain subject to review.
| Actual input | AI · Office Commute |
|---|
| |
See both directions, all six new images and review scope · Download case sharing materials
Same jacket, different mood. Compare the French and American single-image tests: the same outfit, AI identity and first pose. French v3's gentler expression was visually accepted; American retains a restrained urban mood with acceptance pending. One test image per style, with fine garment review still open.
| | | | |
|---|
| Office Commute · same-input comparison | Floral dress | Home outfit | American Street | Japanese Lifestyle |
Case notes and media terms.
Four source photos → six independent AI portraits. This reviewed historical case is separate from the paired gallery and was not generated under beta.8. Explore the complete case and its media terms.
The 24-style paired gallery compares Codex and ChatGPT web results, with source credits, rights disclosures, and per-case reviews. These are AI-generated research comparisons; they do not establish a model ranking or permission to reuse every image. Read the gallery's source and rights records.
| Collection | Published scope | What it establishes |
|---|
| New styling cases | 6 sets × 6 separate images = 36 images | Visual acceptance and remaining review items are recorded separately; excluded from the frozen historical style index. |
| Paired comparison gallery | 24 styles × 2 routes = 48 images | Same-input comparisons with reviews and rights disclosures. These are not 24 complete six-image sets or verification of every style under beta.8. |
| White-vest complete case | 1 style × 6 independent images | One reviewed complete case; the frozen historical primary-case style index is 1/24. |
| Historical direction previews | 2 collections × 24 styles | One six-pose preview sheet per style, with layout checks and maintainer acceptance. The white-vest beta.3 and outfit beta.4 collections retain their original evidence. |
| Real coordinated-outfit source | AI-generated six-pose direction preview |
|---|
| |
coordinated outfit · 24 styles — the beige blazer, white top, dark denim, olive tote, and brown loafers form one locked outfit. This is a single preview sheet. Browse all 24 outfit sheets (GitHub lists the image files; the folder's index.html is a gallery page for a downloaded copy) · Media rights.
single garment · 24 styles — the thumbnails below show the frozen white-vest collection. Each opens a larger six-pose sheet. Style evidence index · Historical work register.
| | | |
|---|
| American Street | Athleisure | Balletcore | British Heritage |
| Cityboy | Clean Fit | Coquette Ladylike | E-commerce Studio |
| French Effortless | Gorpcore | Guochao Street | Italian Luxe |
| Japanese Lifestyle | Korean Cold Editorial | Korean Menswear | Neo Chinese |
| Nordic Minimal | Office Commute Women | Old Money | Preppy |
| Quiet Luxury | Resort Vacation | Workwear Vintage | Y2K Millennium |
These historical previews retain their source records, reviews, and hashes. They do not prove universal garment or outfit coverage or establish current-runtime visual quality, additional complete cases, or non-maintainer adoption.
Scope and limits
Lookset is independent and community-maintained, not an OpenAI product or endorsement. Installation and enablement have been verified on the maintainer's macOS machine; a fresh non-maintainer environment remains unverified. See compatibility and verification scope.
- Source details guide the output. Color, material appearance, silhouette, construction, pattern, and logo placement come from the source. For an outfit, this also includes layering, proportions, shoes, bags, and accessories. Check the generated result against your source; exact small text, logos, and fit are not guaranteed.
- Human review remains part of delivery. This workflow supports apparel imagery, not identity-preserving try-on on a supplied person, CAD fit simulation, non-apparel products, text-only concepts, API integration, or unattended commercial delivery. It does not guarantee platform approval or sales results.
- Style names describe visual direction. Gendered or culturally named styles do not infer a person's identity, ethnicity, nationality, body, or gender.
- Use authorized material and keep rights separate. Images are AI-generated and need applicable labeling. Gallery publication does not mean all third-party rights are cleared; follow the per-case disclosures. Apache-2.0 covers code and documentation, not demo media.
- Generation depends on the chosen host. The Plugin adds no API-key flow, third-party generation service, MCP server, or telemetry. Without the required tools, recognition and prompt preparation can continue; generation pauses or you can explicitly choose the web handoff route.
Documentation and support
Install, upgrade, and rollback · Compatibility · Offline guide (bundled in the Plugin ZIP; open it locally, since GitHub shows its HTML source) · Changelog · Beta progress
Current prerelease: 0.1.0-beta.3. Choose one of 36 visual entries, authorize one labeled nine-panel preview, confirm the direction, then authorize independent images; the first portrait still needs model confirmation. Every generation goes through the ledger scripts. Earlier releases and their galleries remain historical records; use each old archive's bundled guide when installing that version. Beta exit targets and remaining validation are tracked in the Beta register.
Share sanitized installation results through the feedback form. Report bugs in Issues and ask questions in Discussions. Keep private garments, customer data, credentials, and full logs out of public posts. Maintainer: DENGGUI · WeChat: Lvmusic0930.
python3 -m unittest discover -s tests -v
python3 tools/pack-lint.py --strict skills/lookset/references/styles/*.pack.yaml
python3 tools/trigger-eval.py
python3 tools/build-release.py
Runtime lives under skills/lookset/; tests, evals, release tooling, and evidence stay outside it. See Contributing, Security, Migration, and Provenance. Project Beta targets are not OpenAI admission rules; application details are kept in Codex for Open Source.
