Run every source image through every JSON-described transformation pipeline. Build the transform engine, the directory plumbing, the cross product, then the edge cases and the decode-once optimisation.

The five steps

The transform engine 10 min
Reading the two directories 8 min
Every image, every pipeline 12 min
The awkward inputs 8 min
Stop redoing the same work 12 min

A transformation pipeline is an ordered list of steps. Apply them to one image.

def apply_transformations(image, transformations: List[dict]) -> Image.Image

Each step is a dict with a type and sometimes a parameter. Use exactly these Pillow calls:

| type | call | | --- | --- | | grayscale | ImageOps.grayscale(image) | | flip_horizontal | ImageOps.mirror(image) | | flip_vertical | ImageOps.flip(image) | | scale | ImageOps.scale(image, factor) | | blur | image.filter(ImageFilter.BoxBlur(radius)) | | rotate | image.rotate(angle) — Pillow defaults, so expand=False |

Rules:

Steps run in order; each one consumes the previous one's output.
The image you are handed must not be modified — start from a copy.
An empty list returns an independent copy of the input.
An unknown type raises ValueError.

# a 2x2 greyscale image holding   10 20
#                                 30 40
apply_transformations(img, [{"type": "flip_horizontal"}])   # 20 10 / 40 30
apply_transformations(img, [{"type": "rotate", "angle": 90}])  # 20 40 / 10 30

Image Transform Pipeline