History-aware brush physics
A particle-based brush captures bristle deformation, friction, and contact history to model how motion shapes ink.
CALLIMASTER / FROM IMAGE TO ACTION
The ink
remembers
the brush.
A single image. A recovered 6-DoF trajectory.
A real brush bringing it back to life.
REAL ROBOT EXECUTION01 / 05
IMAGE SUPERVISION.
PHYSICAL EVIDENCE.
6-DoFDirectly executable brush poses
+7.16 ppReal-robot clDice vs. CalliRewrite
5 test setsFour scripts + the Collected set
01 / The idea
Expert motion is rare.
Its traces are everywhere.
Expert brush trajectories are costly to capture. Calligraphy images are abundant. CalliMaster uses these visual outcomes to learn the motion that produced them.
A differentiable brush model connects position and orientation to the resulting ink, allowing image supervision to refine full 6-DoF trajectories.
6-DoF
Beyond the centerline
Full brush position and orientation
1K
Real demonstrations
Motion-captured calligraphy characters
300K
A bridge to scale
Synthetic characters with stroke geometry
2M
Images become teachers
Real images without action annotations
02 / The method
Physics connects
the dots.
A differentiable loop connects the desired image, the predicted motion, and the ink produced by brush–paper contact.
01
See the ink
Single character image
02
Recover motion
6-DoF brush trajectory
03
Model contact
Differentiable brush physics
04
Render the trace
Reconstructed ink image
↳VISUAL OUTCOME SUPERVISION↵
Differentiable contact
A neural surrogate predicts brush–paper contact. Differentiable ink rendering sends image feedback back to the trajectory.
Progressive learning
Training progresses from real strokes to synthetic characters, then to diverse calligraphy images without action labels.
02.1 / Contact modeling
Beyond a line
on a page.
A path describes where the brush goes.
Contact determines the ink it leaves.
Explore eight characters, four ways.
THE CONTACT STUDY08 / 08 — Xuan
Brush trajectory Reference ink missed by the midline reconstructionIndividual source views · click to inspect
03 / In simulation
Simulation galleryBefore the ink.
Inside the simulator.
Recovered trajectories, executed in simulation across a range of characters and scripts.
EXAMPLE 01 / AI6-DoF trajectory → simulated execution
The nine demonstrations in this gallery show simulated brush execution.
03.1 / Across scripts
Many scripts.
One model.
Generalization beyond regular-script action demonstrations. One reference, four methods. Every result, side by side.
Regular Script 01
01 / 10
GT is the source image. All method outputs are rendered in simulation.
Colors distinguish strokes. Click any image to inspect.Swipe to compare. Tap to inspect.
03.2 / Simulation benchmark
The difference
is in the details.
+13.27pp
Mean simulation clDice gain over CalliRewrite¹
CalliMaster CalliRewrite G-HTR BaselineHigher is better ↑
All methods evaluated using simulation rendering.
¹ Simulation clDice from manuscript Table II, equally averaged across all five sets: CalliMaster 0.94746; CalliRewrite 0.81472.
View the full results table +
04 / Real-world execution
FIVE DEMONSTRATIONSFrom pixels.
To paper.
The recovered motion leaves the simulator. A physical robot executes CalliMaster’s predicted brush positions and orientations, turning a calligraphy image into ink on paper.
01 Observe→02 Recover→03 Write
SELECT A DEMONSTRATION
All five recordings show physical execution. Videos retain the supplied 3× playback speed.
Compare the resulting ink04.1 / The physical evidence
10 SELECTED EXAMPLESThe same ink.
Four ways to write it.
One target image, four methods, a physical brush. Compare stroke continuity, width, and character structure across four scripts and the Collected set.
Running script
09 / 10
Target: input image. All four outputs: real robot writing.
Ink bounding boxes · normalized scale. Click to enlarge.
04.2 / Real-robot benchmark
Measured
on paper.
25 samples per method.
5 samples in each of 5 test sets.
Five metrics across four methods.
CALLIMASTER / MEAN clDice
0.8229+7.16 pp vs. strongest baseline
Stroke-skeleton topology. Higher is better.
clDice · All five setsHIGHER IS BETTER ↑
Relative viewBaseline
Means use equal weight across the five test sets. Metrics cover all 25 evaluated samples per method; the gallery shows 10 selected examples. Labels and tables report the original scores.
Compare every test set +
05 / The data
From a thousand
demonstrations.
To millions of traces.
CalliTraj-1k captures real human brush motion. Synthetic geometry and image-only supervision extend these demonstrations to diverse calligraphic forms.
Explore CalliTraj-1kCalliTraj-1k
1KReal motion-captured characters with 6-DoF poses and approximately 7,000 individual strokes.
Synthetic characters
300KStroke masks, centerlines, and writing order connect real action priors to whole-character motion.
Real calligraphy images
2MDiverse calligraphic forms supervise action learning through the ink they leave behind.
06 / Resources