FASTER

Font-Agnostic Scene Text Editing and Rendering Framework

FASTER (Font-Agnostic Scene Text Editing and Rendering Framework) addresses scene text editing — replacing text in natural images while preserving the original font style, background texture, and visual coherence. Published at WACV 2025.

Problem

Existing text editing methods require font labels or rely on style transfer that degrades quality at diverse fonts and backgrounds. FASTER removes this requirement via a font-agnostic rendering pipeline.

Approach

  • Style extractor: encodes the glyph style of source text independent of content
  • Text renderer: renders new target text using the extracted style embedding
  • Background inpainting: fills occluded regions naturally without visible seams
  • End-to-end training: jointly optimized on paired synthetic and real-world data

Results

Quantitative comparison against state-of-the-art scene-text-editing methods (↑ higher better, FID ↓ lower better):

Metric SRNet MOSTEL DiffSTE FASTER
PSNR ↑ 18.66 20.75 13.40 20.84
SSIM ↑ 0.614 0.707 0.389 0.790
FID ↓ 48.16 37.55 65.06 11.79
OCR Acc ↑ 14.79 20.10 11.90 30.30

FASTER cuts FID (perceptual distance, lower is better) to less than a third of the next-best method:

{
  "tooltip": { "trigger": "axis", "formatter": "{b}: {c} FID" },
  "grid": { "left": "3%", "right": "4%", "bottom": "3%", "containLabel": true },
  "xAxis": { "type": "category", "data": ["SRNet", "SwapText", "MOSTEL", "DiffSTE", "FASTER"] },
  "yAxis": { "type": "value", "name": "FID (lower better)" },
  "series": [
    {
      "type": "bar",
      "data": [
        { "value": 48.16, "itemStyle": { "color": "#b9c4d0" } },
        { "value": 35.62, "itemStyle": { "color": "#b9c4d0" } },
        { "value": 37.55, "itemStyle": { "color": "#b9c4d0" } },
        { "value": 65.06, "itemStyle": { "color": "#b9c4d0" } },
        { "value": 11.79, "itemStyle": { "color": "#4f8ef7" } }
      ],
      "barMaxWidth": 50,
      "itemStyle": { "borderRadius": [4, 4, 0, 0] },
      "label": { "show": true, "position": "top" }
    }
  ]
}

Qualitative Results

Scene-text-editing results — target text rendered into natural images while preserving font style and background texture.

Publication

Das, A. et al. FASTER: A Font-Agnostic Scene Text Editing and Rendering Framework. WACV 2025. (arXiv:2308.02905)

Work done at CVPRU, Indian Statistical Institute Kolkata.