Research map / Image generation / Editing and style transfer
LineArt: A Knowledge-guided Training-free High-quality Appearance Transfer for Design Drawing with Diffusion Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LineArt, a training-free, knowledge-guided appearance transfer framework that renders detailed design line drawings using multi-frequency line fusion and a two-stage (base layer shaping, surface layer coloring) diffusion process, plus a new ProLines dataset.
Problem
Transferring complex material and lighting appearance onto professional, detailed design line drawings while preserving fine structure, without paired training data or 3D modeling; existing text-based and image-translation methods lose accuracy, consistency, and realism.
Approach
Combines a three-level edge representation (single lines for regions, double lines for details, soft edges from Haar-filtered generated output for texture) with an oil-painting-inspired Imprimatura pipeline: a Retinex-based brightness prior sets the initial noise distribution (underpainting), then texture-patch reference images are injected into selected attention layers (glazing), all on ControlN
Key results
- On the Bronze subset, LineArt reaches FID 86.31 versus 88.26 for the ControlNet+IP-Adapter baseline and SSIM 0.83, with best or second-best values on most metrics.
- On ImageNet-Sketch subset, LineArt obtains FID 100.71 versus 111.04 for the baseline and PSNR 20.93 vs 19.85.
- A 20-participant user study over nine examples ranked LineArt highest on edge fidelity, appearance perception, and overall perception.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LineArt instead of ControlNet-Canny + IP-Adapter baseline | Stable Diffusion · Bronze | FID: 88.26 -> 86.31 | Helped |
| Use LineArt instead of ControlNet-Canny + IP-Adapter baseline | Stable Diffusion · Bronze | GLCM: 29.23 -> 2.42 | Helped |
| Use LineArt instead of ControlNet-Canny + IP-Adapter baseline | Stable Diffusion · DifferSketching | SSIM: 0.95 -> 0.96 | Helped |
| Use LineArt instead of ControlNet-Canny + IP-Adapter baseline | Stable Diffusion · ImageNet Sketch | CD: 7.10 -> 5.89 | Helped |
| Use LineArt instead of ControlNet-Canny + IP-Adapter baseline | Stable Diffusion · DeepPatent | CLIPi: 0.77 -> 0.84 | Helped |
Limitations
- Struggles with line drawings containing excessive shadows or complex textures.
- Fails on tasks requiring precise semantic correspondence between drawing and appearance.
- Relies on a reference image and multi-stage pipeline with hand-tuned thresholds and patch sizes.
Experimental setup
Datasets: Bronze DifferSketching ImageNet-Sketch DeepPatent
Benchmarks: ProLines
Metrics: SSIM Chamfer Distance GLCM PSNR Color Histogram loss FID LPIPS CLIP image similarity user study ranking
Baselines: ZeST Cross-Image Attention (CIA) StyleID DreamBooth InstructPix2Pix + IP-Adapter AesPA-Net ControlNet-Canny + IP-Adapter (Baseline) T2I-Adapter-SDXL
Builds on: ControlNet IP-Adapter Stable Diffusion Cross-Image Attention
Closest papers by meaning
- One-Shot Reference-based Structure-Aware Image to Sketch Synthesis (AAAI 2025 · similarity 0.86)
- DrawingsDreamer: A Unified Multi-View Engineering Drawings Generation Model (NeurIPS 2026 · similarity 0.85)
- Graphic Design with Large Multimodal Model (AAAI 2025 · similarity 0.85)
- AirSketch: Generative Motion to Sketch (NeurIPS 2024 · similarity 0.85)
- SketchDNN: Joint Continuous-Discrete Diffusion for CAD Sketch Generation (ICML 2025 · similarity 0.85)
- Text to Sketch Generation with Multi-Styles (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.