FlyingFrog/Transmute2K-Library-of-Image-Transformations-Simulations-and-Augmentations

Transmute2K: Library of Image Transformations Simulations and Augmentations, based 2K methods, models and simulation from science, art and math.

Transmute2K is a large-scale library and dataset containing 2,000 image transformation methods based on techniques, simulations and models from various fields of science, math and art that have been adapted for image manipulation and image-processing.

In general, the transformations can be applied to standard RGB images, as well as to images or maps of other types, such as PBR materials and hyperspectral maps.

Each transformation follows a model or function. This can range from simple operations such as rotation and blurring to more complex physical and scientific models involving diffusion, clustering, topology,camera effect, optical, physical and chemical proccess and other forms of simulation, with the image used as the working medium on which the transformation operates on.

The transformations are divided into two groups:

1. Image-to-image transformations, which transform one image into another according to a specific model, such as cross-fading.
2. Single-image transformations, which modify one image using a specific process, such as blurring or diffusion.

Each transformation function receives either one image or a pair of images and outputs a sequence of images corresponding to different stages of the transformation.

The input can be a standard RGB image or another type of map, such as a PBR material map or a spectral map.

The transformations were collected from a wide range of fields. Some use standard image-augmentation or image-modification functions, while others adapt models and simulations from different scientific and artistic domains and apply them to images in creative way.

The dataset was created/collected using an agentic AI pipeline based on several large language models, (GLM, KIMI) with human manual inspection and filtering of the results.

Examples of Image-to-Image Transformations

Examples of Single-Image Transformations

File Structure

Code Files

Python scripts for all approximately 2,000 transformations are available in:

Code_image2image_transformations.zip
Code_single_image_transformations.zip

The dataset-generation pipeline is available in:

dataset_generation_code.zip

Image Transformation Sequences

Sequences of images generated by each transformation are available in files whose names begin with transformation_sequence

Each file contains an sequence of images of every transformation in the dataset applied to a random image or pair of images.

Image Transformation As Animated GIFs

Files whose names begin with GIF contain examples of each transformation as animated GIFs.

PBR Material Transformation Sequences

Files whose names begin with PBR contain examples of the transformations applied to PBR materials (As sequence of PBRs)

Code Structure

Each transformation code folder contains:

description.txt
generate.py

The description.txt file contains a description of the transformation.

The generate.py script contains the transformation code and the transformation function.

Image-to-Image Transformations

For image-to-image transformations, the transformation is implemented in the following function:

transform(start_map, end_map, params=None, numsteps=None)

Parameters

start_map

The starting image or map as a NumPy array.

end_map

The target image or map as a NumPy array.

params

An optional dictionary containing transformation-specific parameters.

numsteps

The optional number of steps in the generated transformation sequence.

start_map and end_map must have identical shapes.

The expected array layout is:

[height, width, channels]

Return Value

The function returns a list of maps with the same shape as start_map.

Each item in the list represents one stage of the gradual transformation from start_map to end_map.

Example run

See: run_img2img_transform.py main for example on running the transformation on images and PBRs

Single-Image Transformations

For single-image transformations, the transformation function is:

transform(input_map, params=None, numsteps=None)

The parameters are the same as for image-to-image transformations, except that the function receives a single input map instead of two.

Return Value

The function returns a list of maps with the same shape as input map.

Each image in the output list represents one stage of the transformation.

Running transformation

See run_single_im_transform.py __main__ for examples of running transformations on images or PBRs.

Additional Data

For each transformation, we provide several pieces of additional information:

• Gradual flag (gradual = True/False in Setting.json): Indicates whether the transformation is progressive and gradual. In a gradual transformation, step 1 is always more similar to step 2 than to step 3. This flag is True for almost all transformations.

• Step count: If the default number of steps does not produce a complete transformation, a manually selected number of steps is offered. The default number of steps is sufficient for the majority of transformations (for 512×512 images).

• Related methods: While all methods in Transmute2K are unique, most methods have related methods that share certain concepts and mechanisms. overlapping_models.json in Extra_Data.zip contains, for each method, a list of related models.

• Model fidelity: model_fidelity.json in Extra_Data.zip contains, for each method, a measure of how closely it resembles the model or system it represents. Since the image transformations in Transmute2K aim to be relatively fast (seconds to minutes) and many of the processes they are based on are computationally heavy (e.g., fluid dynamics), most methods use some kind of shortcut or approximation that replicates the core mechanism as a toy model rather than as a detailed simulation.