Glossary · Docs | FabricVTON

Virtual try-on (VTON): Generating a photograph of a person wearing a garment from separate photographs of the person and the garment
Diffusion model: A generative model that creates an image by removing noise step by step
MMDiT: Multimodal diffusion transformer, which processes image and condition tokens in shared attention
LoRA: Low-rank adaptation, a way to fine-tune a large model by training small added matrices
Try-off: Generating a flat product image of a garment from a photograph of someone wearing it
Distillation: Training a smaller or faster model to reproduce a larger model's outputs
FID and KID: Fréchet and Kernel Inception Distances, which measure how close generated images are to real ones as a set
SSIM and LPIPS: Structural and learned perceptual similarity between an output and its ground truth
Kendall τ: A rank correlation used here to measure agreement between metrics and human raters
Krippendorff's alpha: A measure of agreement between human annotators
Monk Skin Tone scale: A ten-point skin tone scale created by Dr Ellis Monk and published by Google for inclusive evaluation
CIEDE2000: A standard formula for perceived colour difference
Garment family: A group of garments that fail in similar ways, such as tops, outerwear or draped garments
Garment presentation: How the garment photograph shows the item: flat, folded as sold, on a hanger or mannequin, or worn by another person
Draped or wrapped garment: A garment whose worn shape comes from wrapping, folding or tying on the body rather than from its cut alone, such as a saree, sarong, shawl or wrap dress
DPIIT: Department for Promotion of Industry and Internal Trade, which recognises startups in India
Video try-on: Dressing a person in a garment across a whole video, keeping identity, motion and the garment consistent from frame to frame
Live try-on: Video try-on on a camera stream as it arrives, with each frame drawn in about 33 to 66 milliseconds
Frames per second (fps): How many new images a video system produces each second; live video needs about 15 to 30
Causal generation: Drawing each frame only from past frames and the current input, never future ones, which live video requires
Self-forcing: Training a video model on its own generated frames so that small errors do not build up over time
KV cache: Stored attention keys and values from earlier frames, reused so that each new frame is fast to compute
WebRTC: The open standard browsers use for live video calls, used here to stream the camera to the GPU and back
VFID: Video Fréchet Inception Distance, a measure of how close generated videos are to real ones as a set