Awesome

🎨 texture-synthesis

A light Rust API for Multiresolution Stochastic Texture Synthesis [1], a non-parametric example-based algorithm for image generation.

The repo also includes multiple code examples to get you started (along with test images), and you can find a compiled binary with a command line interface under the release tab.

Also see our talk More Like This, Please! Texture Synthesis and Remixing from a Single Example which explains this technique and the background more in-depth:

Maintenance note

We at Embark are not actively using or developing these crates and would be open to transferring them to a maintainer or maintainers that would be more active. See #166.

Features and examples

1. Single example generation

Generate similar-looking images from a single example.

API - 01_single_example_synthesis

use texture_synthesis as ts;

fn main() -> Result<(), ts::Error> {
    //create a new session
    let texsynth = ts::Session::builder()
        //load a single example image
        .add_example(&"imgs/1.jpg")
        .build()?;

    //generate an image
    let generated = texsynth.run(None);

    //save the image to the disk
    generated.save("out/01.jpg")
}

CLI

cargo run --release -- --out out/01.jpg generate imgs/1.jpg

You should get the following result with the images provided in this repo:

2. Multi example generation

We can also provide multiple example images and the algorithm will "remix" them into a new image.

API - 02_multi_example_synthesis

use texture_synthesis as ts;

fn main() -> Result<(), ts::Error> {
    // create a new session
    let texsynth = ts::Session::builder()
        // load multiple example image
        .add_examples(&[
            &"imgs/multiexample/1.jpg",
            &"imgs/multiexample/2.jpg",
            &"imgs/multiexample/3.jpg",
            &"imgs/multiexample/4.jpg",
        ])
        // we can ensure all of them come with same size
        // that is however optional, the generator doesnt care whether all images are same sizes
        // however, if you have guides or other additional maps, those have to be same size(s) as corresponding example(s)
        .resize_input(ts::Dims {
            width: 300,
            height: 300,
        })
        // randomly initialize first 10 pixels
        .random_init(10)
        .seed(211)
        .build()?;

    // generate an image
    let generated = texsynth.run(None);

    // save the image to the disk
    generated.save("out/02.jpg")?;

    //save debug information to see "remixing" borders of different examples in map_id.jpg
    //different colors represent information coming from different maps
    generated.save_debug("out/")
}

CLI

cargo run --release -- --rand-init 10 --seed 211 --in-size 300x300 -o out/02.png --debug-out-dir out generate imgs/multiexample/1.jpg imgs/multiexample/2.jpg imgs/multiexample/3.jpg imgs/multiexample/4.jpg

You should get the following result with the images provided in this repo:

3. Guided Synthesis

We can also guide the generation by providing a transformation "FROM"-"TO" in a form of guide maps

API - 03_guided_synthesis

use texture_synthesis as ts;

fn main() -> Result<(), ts::Error> {
    let texsynth = ts::Session::builder()
        // NOTE: it is important that example(s) and their corresponding guides have same size(s)
        // you can ensure that by overwriting the input images sizes with .resize_input()
        .add_example(ts::Example::builder(&"imgs/2.jpg").with_guide(&"imgs/masks/2_example.jpg"))
        // load target "heart" shape that we would like the generated image to look like
        // now the generator will take our target guide into account during synthesis
        .load_target_guide(&"imgs/masks/2_target.jpg")
        .build()?;

    let generated = texsynth.run(None);

    // save the image to the disk
    generated.save("out/03.jpg")
}

CLI

cargo run --release -- -o out/03.png generate --target-guide imgs/masks/2_target.jpg --guides imgs/masks/2_example.jpg -- imgs/2.jpg

NOTE: Note the use of -- to delimit the path to the example imgs/2.jpg, if you don't specify --, the path to the example will be used as another guide path and there won't be any examples.

You should get the following result with the images provided in this repo:

4. Style Transfer

Texture synthesis API supports auto-generation of example guide maps, which produces a style transfer-like effect.

API - 04_style_transfer

use texture_synthesis as ts;

fn main() -> Result<(), ts::Error> {
    let texsynth = ts::Session::builder()
        // load example which will serve as our style, note you can have more than 1!
        .add_examples(&[&"imgs/multiexample/4.jpg"])
        // load target which will be the content
        // with style transfer, we do not need to provide example guides
        // they will be auto-generated if none were provided
        .load_target_guide(&"imgs/tom.jpg")
        .guide_alpha(0.8)
        .build()?;

    // generate an image that applies 'style' to "tom.jpg"
    let generated = texsynth.run(None);

    // save the result to the disk
    generated.save("out/04.jpg")
}

CLI

cargo run --release -- --alpha 0.8 -o out/04.png transfer-style --style imgs/multiexample/4.jpg --guide imgs/tom.jpg

You should get the following result with the images provided in this repo:

5. Inpaint

We can also fill-in missing information with inpaint. By changing the seed, we will get different version of the 'fillment'.

API - 05_inpaint

use texture_synthesis as ts;

fn main() -> Result<(), ts::Error> {
    let texsynth = ts::Session::builder()
        // let the generator know which part we would like to fill in
        // if we had more examples, they would be additional information
        // the generator could use to inpaint
        .inpaint_example(
            &"imgs/masks/3_inpaint.jpg",
            // load a "corrupted" example with missing red information we would like to fill in
            ts::Example::builder(&"imgs/3.jpg")
                // we would also like to prevent sampling from "corrupted" red areas
                // otherwise, generator will treat that those as valid areas it can copy from in the example,
                // we could also use SampleMethod::Ignore to ignore the example altogether, but we
                // would then need at least 1 other example image to actually source from
                // example.set_sample_method(ts::SampleMethod::Ignore);
                .set_sample_method(&"imgs/masks/3_inpaint.jpg"),
            // Inpaint requires that inputs and outputs be the same size, so it's a required
            // parameter that overrides both `resize_input` and `output_size`
            ts::Dims::square(400),
        )
        // Ignored
        .resize_input(ts::Dims::square(200))
        // Ignored
        .output_size(ts::Dims::square(100))
        .build()?;

    let generated = texsynth.run(None);

    //save the result to the disk
    generated.save("out/05.jpg")
}

CLI

Note that the --out-size parameter determines the size for all inputs and outputs when using inpaint!

cargo run --release -- --out-size 400 --inpaint imgs/masks/3_inpaint.jpg -o out/05.png generate imgs/3.jpg

You should get the following result with the images provided in this repo:

6. Inpaint Channel

Instead of using a separate image for our inpaint mask, we can instead obtain the information from a specific channel. In this example, the alpha channel is a circle directly in the middle of the image.

API - 06_inpaint_channel

use texture_synthesis as ts;

fn main() -> Result<(), ts::Error> {
    let texsynth = ts::Session::builder()
        // Let the generator know that it is using 
        .inpaint_example_channel(
            ts::ChannelMask::A,
            &"imgs/bricks.png",
            ts::Dims::square(400),
        )
        .build()?;

    let generated = texsynth.run(None);

    //save the result to the disk
    generated.save("out/06.jpg")
}

CLI

cargo run --release -- --inpaint-channel a -o out/06.png generate imgs/bricks.jpg

You should get the following result with the images provided in this repo:

7. Tiling texture

We can make the generated image tile (meaning it will not have seams if you put multiple images together side-by-side). By invoking inpaint mode together with tiling, we can make an existing image tile.

API - 07_tiling_texture

use texture_synthesis as ts;

fn main() -> Result<(), ts::Error> {
    // Let's start layering some of the "verbs" of texture synthesis
    // if we just run tiling_mode(true) we will generate a completely new image from scratch (try it!)
    // but what if we want to tile an existing image?
    // we can use inpaint!

    let texsynth = ts::Session::builder()
        // load a mask that specifies borders of the image we can modify to make it tiling
        .inpaint_example(
            &"imgs/masks/1_tile.jpg",
            ts::Example::new(&"imgs/1.jpg"),
            ts::Dims::square(400),
        )
        //turn on tiling mode!
        .tiling_mode(true)
        .build()?;

    let generated = texsynth.run(None);

    generated.save("out/07.jpg")
}

CLI

cargo run --release -- --inpaint imgs/masks/1_tile.jpg --out-size 400 --tiling -o out/07.bmp generate imgs/1.jpg

You should get the following result with the images provided in this repo:

8. Repeat texture synthesis transform on a new image

We can re-apply the coordinate transformation performed by texture synthesis onto a new image.

API - 08_repeat_transform

use texture_synthesis as ts;

fn main() -> Result<(), ts::Error> {
    // create a new session
    let texsynth = ts::Session::builder()
        //load a single example image
        .add_example(&"imgs/1.jpg")
        .build()?;

    // generate an image
    let generated = texsynth.run(None);

    // now we can apply the same transformation of the generated image
    // onto a new image (which can be used to ensure 1-1 mapping between multiple images)
    // NOTE: it is important to provide same number of input images as the 
    // otherwise, there will be coordinates mismatch
    let repeat_transform_img = generated
        .get_coordinate_transform()
        .apply(&["imgs/1_bw.jpg"])?;

    // save the image to the disk
    // 08 and 08_repeated images should match perfectly
    repeat_transform_img.save("out/08_repeated.jpg").unwrap();
    generated.save("out/08.jpg")
}

CLI

1. First, we need to create a transform that can be reused

The notable bit here is the --save-transform out/multi.xform which creates the file that can be used to generate new outputs with.

cargo run --release -- --rand-init 10 --seed 211 --in-size 300x300 -o out/02.png generate --save-transform out/multi.xform imgs/multiexample/1.jpg imgs/multiexample/2.jpg imgs/multiexample/3.jpg imgs/multiexample/4.jpg

2. Next, we use the repeat subcommand to repeat transform with different inputs

The important bits here are the use of the repeat subcommand instead of generate, and --transform out/multi.xform which tells what transform to apply to the inputs. The only restriction is that the number of images you specify must match the original number of examples exactly. If the input images have different dimensions than the example images, they will be automatically resized for you.

cargo run --release -- -o out/02-repeated.png repeat --transform out/multi.xform imgs/multiexample/1.jpg imgs/multiexample/2.jpg imgs/multiexample/4.jpg imgs/multiexample/3.jpg

Also note that the normal parameters that are used with generate don't apply to the repeat subcommand and will be ignored.

9. Sample masks

Sample masks allow you to specify how an example image is sampled during generation.

API - 09_sample_masks

use texture_synthesis as ts;

fn main() -> Result<(), ts::Error> {
    let session = ts::Session::builder()
        .add_example(
            ts::Example::builder(&"imgs/4.png").set_sample_method(ts::SampleMethod::Ignore),
        )
        .add_example(ts::Example::builder(&"imgs/5.png").set_sample_method(ts::SampleMethod::All))
        .seed(211)
        .output_size(ts::Dims::square(200))
        .build()?;

    // generate an image
    let generated = session.run(None);

    // save the image to the disk
    generated.save("out/09.png")
}

CLI

cargo run --release -- --seed 211 --out-size 200 --sample-masks IGNORE ALL --out 09_sample_masks.png generate imgs/4.png imgs/5.png

You should get the following result with the images provided in this repo:

10. Combining texture synthesis 'verbs'

We can also combine multiple modes together. For example, multi-example guided synthesis:

Or chaining multiple stages of generation together:

For more use cases and examples, please refer to the presentation "More Like This, Please! Texture Synthesis and Remixing from a Single Example"

Additional CLI functionality

Some functionality is only exposed through the CLI and not built into the library.

flip-and-rotate

This subcommand takes each example and performs flip and rotation transformations to it to generate additional example inputs for generation. This subcommand doesn't support target or example guides.

Example: cargo run --release -- -o out/output.png flip-and-rotate imgs/1.jpg

Command line binary

• Download the binary for your OS.
• Or Install it from source.
  • Install Rust - The minimum required version is 1.37.0
  • Clone this repo
  • In a terminal cd to the directory you cloned this repository into
  • Run cargo install --path=cli
  • Or if you wish to see the texture as it is being synthesized cargo install --path=cli --features="progress"
• Open a terminal
• Navigate to the directory where you downloaded the binary, if you didn't just cargo install it
• Run texture_synthesis --help to get a list of all of the options and commands you can run
• Refer to the examples section in this readme for examples of running the binary

Notes

• By default, generating output will use all of your logical cores
• When using multiple threads for generation, the output image is not guaranteed to be deterministic with the same inputs. To have 100% determinism, you must use a thread count of one, which can by done via
  • CLI - texture-synthesis --threads 1
  • API - SessionBuilder::max_thread_count(1)

Limitations

• Struggles with complex semantics beyond pixel color (unless you guide it)
• Not great with regular textures (seams can become obvious)
• Cannot infer new information from existing information (only operates on what’s already there)
• Designed for single exemplars or very small datasets (unlike Deep Learning based approaches)

Additional Dependencies

If you're compiling for Linux, you'll need to have libxkbcommon development libraries installed. For ubuntu this is libxkbcommon-x11-dev.

Links/references

[1] [Opara & Stachowiak] "More Like This, Please! Texture Synthesis and Remixing from a Single Example"

[2] [Harrison] Image Texture Tools

[3] [Ashikhmin] Synthesizing Natural Textures

[4] [Efros & Leung] Texture Synthesis by Non-parametric Sampling

[5] [Wey & Levoy] Fast Texture Synthesis using Tree-structured Vector Quantization

[6] [De Bonet] Multiresolution Sampling Procedure for Analysis and Synthesis of Texture Images

[7] All the test images in this repo are from Unsplash

Contributing

We welcome community contributions to this project.

Please read our Contributor Guide for more information on how to get started.

License

Licensed under either of

• Apache License, Version 2.0, (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
• MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)

at your option.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.