Continuous source prior
Noise around category text embeddings expands isolated CLIP anchors into a continuous Gaussian mixture.
SIGGRAPH Asia 2026
SketchFlow
Zero-Shot Vector Sketch Generation via GMM Prior Flow in CLIP Latent Space
Shenzhen University
The problem
Learn an open semantic transport from category-only vector sketches.
QuickDraw offers millions of human stroke trajectories but only 345 discrete labels. SketchFlow turns those sparse text anchors into a continuous GMM prior, transports it onto the rendered-sketch CLIP manifold with OT-CFM, and decodes the transported feature into a vector stroke sequence.
Method
From a text prototype to a drawable sketch feature
The learned vector field bridges the CLIP text-sketch modality gap before the diffusion decoder generates geometry.
Semantic OT transport
Category-consistent OT-CFM learns straight trajectories toward the rendered-sketch embedding distribution.
Vector decoding
A hybrid 1D U-Net and Transformer diffusion decoder produces 256-point stroke trajectories with pen states.
345QuickDraw categories
~50Mhuman-drawn sketches
256points per trajectory
ViT-B/32shared CLIP space
Results
Concepts outside the training taxonomy
Distinctive visual concepts can follow the learned transport and retain the sparse, human-like style of QuickDraw trajectories.
Drawing process
Stroke order remains part of the result
Each animation follows the generated vector trajectory from its first point to the completed sketch.
SIGGRAPH
8 generated trajectories
ASIA 2026
8 generated trajectories
Malaysia landmarks
6 generated trajectories
Scope
Best for concise concepts with a strong visual prototype.
SketchFlow is trained from the 345 discrete QuickDraw categories. It is not a general image model and does not reliably parse long, compositional prompts.
Its strongest zero-shot behavior appears on visually distinctive concepts that CLIP represents clearly, such as Kirby, Mickey, ghosts, rockets, emotions, symbols, and landmarks.
Citation
SketchFlow
Camera-ready paper and final bibliographic metadata will be added after publication processing.
@inproceedings{zhou2026sketchflow,
title = {SketchFlow: Zero-Shot Vector Sketch Generation
via GMM Prior Flow in CLIP Latent Space},
author = {Zhou, Jin and Yang, Hongliang and
Xu, Pengfei and Huang, Hui},
booktitle = {ACM SIGGRAPH Asia 2026 Conference Papers},
year = {2026}
}