TL;DR
ReDraw is a novel two-stage vision-language model (VLM) that revitalizes ancient Chinese characters by treating restoration as a structured SVG code generation and editing task. By predicting stroke skeletons before their contours and using reinforcement learning with structure-aware rewards, it faithfully preserves existing historical strokes while accurately redrawing severely damaged ones.
What Problem Does This Paper Address?
Ancient Chinese manuscripts, particularly those written on Han Dynasty bamboo and wooden slips (Hanjian), are invaluable historical records. However, centuries of decay have left them with "damaged ink"—faded, broken, or entirely missing strokes.
Traditional restoration methods typically rely on image-to-image (I2I) translation (like GANs or Diffusion models). While these produce visually plausible results, they suffer from two major flaws:
- Lack of Control: They operate on pixels, making it hard to manipulate specific stroke geometries.
- Over-Restoration: They often "hallucinate" or alter the geometry of strokes that are actually well-preserved, violating the archaeological principle of "minimal intervention."
The authors argue that restoration should be more like an expert's manual redrawing process: identify what's there, trace it faithfully, and then carefully fill in the gaps.
Methods
The researchers introduce ReDraw, a framework that reformulates restoration into two distinct SVG-based stages:
- Stage 1: SVG Generation: The model takes a degraded image and generates an initial SVG. It handles two subtasks: Faithful Vectorization (capturing the exact geometry and degradation of observed strokes) and Conservative Tracing (cleaning up mild noise while leaving major gaps for the next stage).
- Stage 2: Structure-aware SVG Editing: This stage performs Stroke Completion. Crucially, it uses a Skeleton Chain-of-Thought (CoT) approach. Instead of jumping straight to the final shape, the model first predicts the "skeleton" (the central path) of the missing stroke and then generates the "contour" (the outline). This mimics the way a calligrapher thinks about stroke structure.
To refine the model's output, the authors employed Group Relative Policy Optimization (GRPO), a form of reinforcement learning. They designed "structure-aware rewards" that check if the predicted skeleton matches the predicted contour and if the new strokes overlap correctly with the existing ones.
Datasets
Because high-quality pairs of "degraded vs. perfect" ancient characters are rare, the authors built a massive multi-layer data synthesis pipeline:
- Ink Layer: 1.2M clean references derived from 120k seal and clerical script vectors.
- Degradation Layer: Simulated stroke erasure, edge perturbation, and color jitter.
- Material Layer: Textures from 400 real character-free bamboo slip images.
For real-world evaluation, they curated an expert-verified Hanjian dataset consisting of 320 real glyphs, which they used for few-shot adaptation.
Key Insights and Ideas
- SVG as a "Structured" Prior: By using SVG code instead of pixels, the model is forced to treat characters as a collection of discrete, editable paths. This naturally prevents the "blurry" or "mushy" artifacts common in pixel-based restoration.
- The "Skeleton" is the Key: Predicting the skeleton first acts as a geometric constraint. It ensures that even if the ink is missing, the underlying calligraphic structure remains sound.
- Localized Editing: Unlike I2I models that regenerate the whole image, ReDraw's second stage only appends new SVG paths. This guarantees that the original, observed strokes remain 100% unchanged at the code level.
Key Results
ReDraw significantly outperformed existing state-of-the-art image restoration models (like SwinIR and DocRes) and general-purpose VLMs (like GPT-4o) across several metrics:
- Structural Similarity (SSIM): ReDraw achieved 0.9225, compared to the next best baseline at 0.8732.
- Perceptual Fidelity (LPIPS): It reduced error by 35.7% relative to PromptIR.
- RL Impact: The GRPO refinement reduced Mean Squared Error (MSE) by approximately 27% compared to standard supervised fine-tuning (SFT), leading to much cleaner stroke connections.
Limitations and Future Work
- Complexity Limits: The VLM's context window (8k tokens) limits its ability to handle extremely complex characters with dozens of strokes.
- Severe Damage: The model's reliability drops when more than 50% of a character's strokes are missing.
- Contextual Blindness: Currently, ReDraw restores characters in isolation. Future work aims to integrate it into a full-page restoration agent that uses surrounding text as context to help "guess" missing strokes more accurately.
- Human-in-the-loop: The authors suggest that since these are "candidate reconstructions," future versions should allow interactive expert assessment to choose between multiple completion hypotheses.
Summary updated