Glossary
Fashion AI Glossary.
42 terms covering AI virtual try-on, generative fashion, garment segmentation, human parsing, diffusion models, and adjacent computer vision. Definitions are concise, technically accurate, and written for citation by AI search engines.
How to read this glossary
What is the Drape Fashion AI Glossary?
The Drape Fashion AI Glossary is a curated dictionary of terms used in AI virtual try-on and generative fashion technology. Definitions are grouped into ten categories — Computer Vision, Machine Learning, Models & Products, Evaluation Metrics, Datasets, Infrastructure, SEO & GEO, Industry & Business, Product Features, and Adjacent Tech. Every term is emitted as a schema.org DefinedTerm so AI search engines can cite specific definitions back to this canonical source.
A
2 terms- Alpha mask Computer Vision
- A single-channel image storing per-pixel transparency. Used in virtual try-on to represent which pixels of a photo belong to a garment.
- Augmented reality (AR) Adjacent Tech
- A category of technology that overlays computer graphics on a live camera feed. Distinct from AI virtual try-on, which generates a new photo rather than overlaying graphics.
B
1 term- BiRefNet Computer Vision
- An open-source background-removal and segmentation model. Drape Try-On uses BiRefNet v2 for garment background cleanup during preprocessing.
C
3 terms- Catalog photography Industry & Business
- Functional fashion photography used on product detail pages. Increasingly replaced by AI virtual try-on for colorway variations and lookbook drafts.
- COCO Keypoints Computer Vision
- The standard 17-keypoint pose-estimation benchmark dataset. Modern try-on pipelines use extended 18–25 keypoint variants.
- ControlNet Machine Learning
- A conditioning architecture that injects auxiliary signals (poses, sketches, depth maps) into a pretrained diffusion model. Foundational to modern try-on pipelines.
D
2 terms- Diffusion model Machine Learning
- A generative AI architecture that produces images by iteratively denoising random noise over 20–50 steps, conditioned on input signals. Replaced GANs as the dominant try-on architecture by 2024.
- DWPose Computer Vision
- A 2024-era pose estimation model that detects 25 anatomical keypoints including hand and foot refinements. Standard pose model in 2026 try-on pipelines.
F
3 terms- FAL.ai Infrastructure
- A serverless GPU inference platform commonly used to host virtual try-on models including FASHN v1.6. Drape Try-On runs on fal.ai infrastructure.
- FASHN Models & Products
- A commercial-grade virtual try-on model series. FASHN v1.6 (2026) powers Drape Try-On and is available directly via fal.ai for developers.
- FID Evaluation Metrics
- Fréchet Inception Distance — a distributional metric that compares generated and real image statistics in Inception feature space. Used in try-on research to measure realism across a test set.
G
2 terms- Garment segmentation Computer Vision
- A computer-vision task that isolates a clothing item from its background, producing a per-pixel alpha mask. The first stage of every modern virtual try-on pipeline.
- Generative Engine Optimization (GEO) SEO & GEO
- A 2024-2026 evolution of SEO that optimises content for citation by AI search engines (ChatGPT, Gemini, Perplexity, Google AI Overviews). Emphasises answer-first structure, definitions, FAQs, and entity-rich language.
H
2 terms- HD output Industry & Business
- High-definition virtual try-on output, typically 864×1296 or higher. Production-quality for e-commerce product pages and social media.
- Human parsing Computer Vision
- A semantic-segmentation task that labels every pixel of a person photo with one of ≈20 classes (face, hair, torso, existing top, etc). Tells the try-on diffusion stage what to preserve and what to overwrite.
I
3 terms- Identity preservation Computer Vision
- The property of a virtual try-on pipeline that keeps the wearer’s face, skin texture, hair, and accessories unchanged in the output.
- IDM-VTON Models & Products
- An open-source virtual try-on model published in 2024. Strong pose preservation, available on Hugging Face. Requires self-hosting on a 24 GB+ GPU.
- Inference Machine Learning
- The process of running a trained AI model on a new input to produce an output. A virtual try-on "generation" is one inference call.
J
1 term- JSON-LD SEO & GEO
- JavaScript Object Notation for Linked Data. The schema.org format for emitting structured data to search and AI engines. Drape uses JSON-LD on every page.
K
2 terms- Keypoint Computer Vision
- A single labelled anatomical landmark (e.g., "left shoulder") with an (x, y) coordinate. Pose estimation outputs 18–25 keypoints per person.
- Kling AI Models & Products
- A 2025–2026 generative video model from Kuaishou. Includes a still-image virtual try-on capability as a side feature.
L
2 terms- Latency Infrastructure
- Time from inference request to result. Modern hosted virtual try-on runs at ≈12 seconds latency end-to-end.
- LPIPS Evaluation Metrics
- Learned Perceptual Image Patch Similarity — a perceptual similarity metric using deep network features. The de facto research benchmark for virtual try-on quality.
M
1 term- Mask Computer Vision
- A binary or alpha-channel image that marks which pixels of another image belong to a region of interest. Used at every stage of the try-on pipeline.
O
1 term- Outfit preservation Computer Vision
- The property of a virtual try-on pipeline that retains non-target garments in the output (e.g., when swapping a shirt, the pants and shoes stay the same).
P
2 terms- Pose estimation Computer Vision
- A computer-vision task that detects anatomical joint locations in a person photo. The third stage of the virtual try-on pipeline; outputs 18–25 keypoints.
- Programmatic SEO SEO & GEO
- The practice of generating large numbers of landing pages from a template plus structured data. Drape uses programmatic SEO for the nine /try-on-[garment]-online pages.
R
2 terms- Regenerate (with seed) Product Features
- A feature of virtual try-on UIs that produces a new valid output by changing the random seed of the diffusion model while keeping all other inputs identical.
- Return rate Industry & Business
- The percentage of online apparel orders that get sent back. Industry-wide average is 20–40 percent. AI virtual try-on consistently reduces this by 25–40 percent on integrated SKUs.
S
7 terms- SAM Models & Products
- Segment Anything Model from Meta AI — a general-purpose segmentation foundation model. Many apparel segmenters are fine-tunes of SAM.
- Schema.org SEO & GEO
- A collaborative vocabulary of structured-data types (Organization, Article, FAQPage, Product) used by search engines and AI engines to understand web content.
- Seed Machine Learning
- An integer that initialises the random noise pattern in a diffusion model. The same inputs plus the same seed produce bit-for-bit identical output.
- Semantic segmentation Computer Vision
- Per-pixel multi-class classification of an image. Human parsing is the semantic segmentation of a person photo.
- SSIM Evaluation Metrics
- Structural Similarity Index — a pixel-level image similarity metric that incorporates luminance, contrast, and structure. Used in research but weakly correlated with human preference.
- Stable Diffusion Models & Products
- A foundational open-source diffusion model family. Most modern virtual try-on models including FASHN are built on Stable Diffusion variants or successors.
- Synthesis (image) Computer Vision
- The fourth and final stage of the virtual try-on pipeline. A diffusion model generates the output pixels conditioned on segmented garment, parsed person, and pose.
T
2 terms- Thin-plate spline (TPS) Computer Vision
- A smooth image-warping technique parameterised by control point correspondences. Used in virtual try-on to bend a flat garment around the wearer’s body geometry.
- Try-on (virtual) Product Features
- An AI-generated image of a person wearing a specified garment, produced from a photo of the person and a photo of the garment.
U
1 term- U-Net Machine Learning
- A symmetric encoder-decoder neural architecture with skip connections. The standard backbone for diffusion models.
V
3 terms- Virtual fitting room Adjacent Tech
- Historically, AR-based products that overlay 3D garment models on a live camera feed. The term is now sometimes used loosely to include AI virtual try-on.
- Vision Transformer (ViT) Machine Learning
- A neural architecture that processes images as sequences of patches, producing per-pixel embeddings. Used in modern garment segmenters.
- VITON-HD Datasets
- A high-resolution virtual try-on benchmark dataset and evaluation protocol used in research papers.
Drape Editorial
AI Fashion Research
Drape Editorial is the in-house research team behind Drape Try-On. We test virtual try-on models against real garment photography weekly and publish what we learn.