| """Transforms that simulate what a scraper/training pipeline does to an image. |
| |
| `scraper_sim` (JPEG Q85 + light blur) is baked into fitness so the overlay is |
| selected to survive realistic preprocessing. `hard_transforms` (heavy blur, |
| downscale, rotation) are the robustness *ceiling* used only in the eval harness. |
| """ |
|
|
| from __future__ import annotations |
|
|
| import io |
|
|
| from PIL import Image, ImageFilter |
|
|
| from veil_pgd.config import RobustnessSim, get_settings |
|
|
|
|
| def jpeg_recompress(image: Image.Image, quality: int) -> Image.Image: |
| buf = io.BytesIO() |
| image.convert("RGB").save(buf, format="JPEG", quality=quality) |
| buf.seek(0) |
| return Image.open(buf).convert("RGB") |
|
|
|
|
| def scraper_sim(image: Image.Image, cfg: RobustnessSim | None = None) -> Image.Image: |
| cfg = cfg or get_settings().robustness |
| out = image |
| if cfg.gaussian_blur_radius > 0: |
| out = out.filter(ImageFilter.GaussianBlur(cfg.gaussian_blur_radius)) |
| out = jpeg_recompress(out, cfg.jpeg_quality) |
| return out |
|
|
|
|
| def hard_transforms(image: Image.Image) -> dict[str, Image.Image]: |
| """The robustness ceiling: transforms known to weaken typographic overlays.""" |
| w, h = image.size |
| return { |
| "heavy_blur": image.filter(ImageFilter.GaussianBlur(2.5)), |
| "downscale_half": image.resize((max(1, w // 2), max(1, h // 2))).resize((w, h)), |
| "rotate_15": image.rotate(15, resample=Image.BICUBIC, expand=False), |
| "jpeg_q50": jpeg_recompress(image, 50), |
| } |
|
|
