vibrant leaves

Synthetic images for training image restoration networks, generated in your browser.

VibrantLeaves generates synthetic images on which image restoration networks (denoising, super-resolution, …) can be trained instead of natural images. Random shapes occlude each other (the dead leaves model), filled with pseudo-periodic and 1/f textures, in three depth-of-field layers. Choose the parameters, and drop a few photos to sample the colours from. See the paper and the code, which also generates full datasets. Images are generated in your browser; your photos are not uploaded.

Shapes
Shape type
Radius
1 px101001000 px
Power-law exponent
Higher values give more small shapes.
Textures
Off: large shapes get colour gradients.
Texture types click to add / remove
Share of each type
1/f slope of the noise
01234
Colours random
Drop images here or browse
Colours are sampled from these images; without images, colours are random.
Depth of field & output
Image size
The images are fixed for given parameters, so you can compare them when you change a parameter. Click an image to enlarge and download it.