Skip to main content
Back to timeline
arXivSource publication:

Han slip characters get redrawn as editable SVG paths: ReDraw reaches 0.9225 SSIM on synthetic glyphs while leaving observed strokes untouched

Synopsis

Ink loss and broken strokes on Han Dynasty bamboo and wooden slips have been handled mainly by image-to-image restoration of whole glyphs, which can alter observed strokes and gives little direct control over stroke geometry. ReDraw recasts this as generation and editing of Scalable Vector Graphics (SVG) code: a first stage faithfully vectorizes or conservatively traces existing strokes, and a second stage predicts skeletons before contours, appending new paths without changing existing ones. On 600 synthetic samples it reaches SSIM 0.9225 and LPIPS 0.0850, above retrained DocRes and PromptIR, and after one-shot adaptation on 320 real Hanjian glyphs unseen glyphs reach SSIM 0.9721.

Source-provided article image: From Restoration to Redrawing: Revitalizing Ancient Chinese Characters by SVG Code Generation and Editing
arXiv

Interpretation

Ancient-character restoration moves from whole-image pixel reconstruction to vector-code operations on character geometry, keeping observed strokes and added completions separate at the code level. Earlier image-to-image restoration rebuilt the whole image, could modify observed strokes, and exposed no editable geometric primitives for localized completion of severely incomplete strokes. On 600 synthetic samples that share source glyphs with training but differ in damage pattern and background, ReDraw reaches SSIM 0.9225 and LPIPS 0.0850, versus SSIM 0.8732 for DocRes and LPIPS 0.1321 for PromptIR, with all methods retrained for 10,000 steps on the same 60k synthetic samples.

Restoration is split into two stages: first obtain a vector representation of existing strokes, then append paths for missing strokes while leaving given paths unchanged. Stage one offers two targets, faithful vectorization that keeps worn contours and small holes, and conservative tracing that repairs only mild geometric degradation and leaves severe incompleteness to stage two. On the same 60k synthetic samples with 10,000 steps each, two-stage supervised fine-tuning reaches SSIM 0.9055 against 0.8916 for single-stage direct prediction of a complete SVG, with LPIPS falling from 0.1210 to 0.1120 and MSE from 0.04153 to 0.03881.

Completion predicts a stroke skeleton before its contour and is refined with reinforcement learning using structure-aware rewards, reducing redundant strokes and wrong connections. The skeleton acts as an explicit geometric intermediate, so contour generation follows an inspectable structure instead of emitting a whole path at once. Starting from the supervised model trained on 60k synthetic samples, 1,000 additional GRPO updates raise N2F SSIM from 0.9055 to 0.9225 and cut MSE and LPIPS by roughly 27% and 24%.

Perspective

The demonstrated scope is synthetic restoration of Hanjian glyphs plus few-shot adaptation to real glyphs: synthetic evaluation uses 600 samples sharing source glyphs with training, and the real adaptation set is 80 character classes with four glyphs each, 320 glyphs in total, manually completed by four volunteers over about 200 working hours and reviewed by experts. The authors propose later integrating ReDraw into a full-page document restoration agent with an agentic verifier for iterative refinement.

Generated paths are candidate reconstructions, not verified historical appearance; restoration is unreliable when more than 50% of strokes are missing or ink degradation is more diverse; an 8k-token context window limits support for highly complex characters; and only foreground characters are restored, excluding backgrounds. The authors also note that test subsets differ across shot counts, preventing a controlled comparison of the effect of additional shots.