Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

A White-Box Photo Post-Processing Framework

ACM Transactions on Graphics (presented at SIGGRAPH 2018)

Yuanming Hu1,2, Hao He1,2, Chenxi Xu1,3, Baoyuan Wang1, Stephen Lin1

[Paper] [PDF Slides] [PDF Slides with notes] [SIGGRAPH 2018 Fast Forward]

1Microsoft Research 2MIT CSAIL 3Peking University

Change log:

  • July 9, 2018: Minor improvements.
  • May 20, 2018: Inlcuded user study UI.
  • May 13, 2018: Minor improvements.
  • March 30, 2018: Added instructions for preparing training data with Adobe Lightroom.
  • March 26, 2018: Updated MIT-Adobe FiveK data set and treatments for 8-bit jpg and png images.
  • March 9, 2018: Finished code clean-up. Uploaded code and some instructions.
  • March 1, 2018: Added some images.


Requirements: python3 and tensorflow. Tested on Ubuntu 16.04 and Arch Linux. OS X may also work, though not tested.

sudo pip3 install tensorflow-gpu opencv-python tifffile scikit-image
git clone --recursive
cd exposure

Using the pretrained model

  • python3 example pretrained models/sample_inputs/*.tif
  • Results will be generated at outputs/

Training your own model on the FiveK dataset

  • python3
    • This script will automatically setup the MIT-Adobe FiveK Dataset
    • Total download size: ~2.4GB
    • Only the downsampled and data-augmented image pack will be downloaded. Original dataset is large as 50GB and needs Adobe Lightroom to pre-process the RAW files. If you want to do data pre-processing and augmentation on your own, please follow the instructions here.
  • python3 example test
    • This command will load,
    • and create a model folder at models/example/test
  • Have a cup of tea and wait for the model to be trained (~100 min on a GTX 1080 Ti)
    • The training progress is visualized at folder models/example/test/images-example-test/*.png
    • Legend: top row: learned operating sequences; bottom row: replay buffer, result output samples, target output samples
  • python3 example test models/sample_inputs/*.tif (This will load models/example/test)
  • Results will be generated at outputs/

Training on your own dataset

Please check out

Visual Results

All results on the MIT-FiveK data set:


  1. Does it work on jpg or png images?

To some extent, yes. Exposure is originally designed for RAW photos, which assumes 12+ bit color depth and linear "RGB" color space (or whatever we get after demosaicing). jpg and png images typically have only 8-bit color depth (except 16-bit pngs) and the lack of information (dynamic range/activation resolution) may lead to suboptimal results such as posterization. Moreover, jpg and most pngs assume an sRGB color space, which contains a roughly 1/2.2 Gamma correction, making the data distribution different from training images (which are linear).

Therefore, when applying Exposure to these images, such nonlinearity may affect the result, as the pretrained model is trained on linearized color space from ProPhotoRGB.

If you train Exposure in your own collection of images that are jpg, it is OK to apply Exposure to similar jpg images, though you may still get some posterization.

Note that Exposure is just a prototype (proof-of-concept) of our latest research, and there are definitely a lot of engineering efforts required to make it suitable for a real product. Like many deep learning systems, usually when the inputs are too different from training data, suboptimal results will be generated. Defects like this may be alleviated by more human engineering efforts which are not included in this research project whose goal is simply prototyping.

  1. The images from the datasets are 16-bit. Have you tried 8bit jpg as input? If so, how about the performance? I did. We have some internal projects (which I cannot disclose right now, sorry) that actually have only 8-bit inputs. Most results are as good as 16-bit inputs. However, from time to time (< 5% on the dataset I tested) you may find posterization/saturation artifacts due to the lack of color depth (intensity resolution/dynamic range).

  2. Why am I getting different results everytime I run Exposure on the same image?

In the paper, you will find that the system is learning a one-to-many mapping, instead of one-to-one. The one-to-many mapping mechanism is achieved using (random) dropout (instead of noise vectors in some other GAN papers), and therefore you may get slightly different results every time.

  1. Pre-trained model?

The repository contains a submodule with the pretrained model on the MIT-Adobe Five-K dataset. Please make sure you clone the repo recursively:

git clone --recursive

We also have pre-trained model for the two artists mentioned in the paper. However, to avoid copyright issues we might not release it in public. Please email Yuanming Hu if you want these models.

  1. Why linearize the photos? I changed the Gamma parameter from 1.0 to 2.2, the results differ a lot.

A bit background: the sensor of digital cameras have almost linear activation curves. This means if one pixel receives twice photons it will give you twice as large value (activation). However, it is not the case for displays, which as a nonlinear activation, roughly x->x2.2, which means a twice as large value will result in 4.6 times brighter pixel when displayed. That's why sRGB color space has a ~1/2.2 gamma, which makes color activations stored in this color space ready-to-display on a CRT display as it inverts such nonlinearity. Though we no longer use CRT displays nowadays, modern LCD displays still follow this convention.

Such disparity leads to a process called Gamma correction. You may find that directly displaying a linear RGB image on screen will typically lead to a very dark image. A simple solution is to map pixel intensities from x to x->x1/2.2, so that the image will be roughly converted to an sRGB image that suits your display. Before you do that, make sure your image already has a reasonable exposure value. An easy way to do that is scaling the image so that the average intensity (over all pixels, R, G and B) is some value like 0.18.

Another benefit of such 1/2.2 Gamma correction for sRPG is better preservation of information for the human visual system. Human eyes have a logarithmic perception and are more sensitive to low-light regions. Storing a boosted value for low light in 1/2.2 gamma actually gives you more bits there, which alleviates quantization in low-light parts.

Google linear workflow if you are interested in more details. You may find useful information such as this.

Why linearize the image: Exposure is designed to ba an end-to-end photo-processing system. The input should be a RAW file (linear image, after demosaicing). However, the data from the dataset are in Adobe DNG formats, making reading them hard in a third-party program. That's why we export the data in ProPhoto RGB color space, which is close to sRGB while having a roughly 1/1.8 Gamma instead of 1/2.2. Then we do linearization here to make the inputs linear.

I tried to change the Gamma parameter from 1.0 to 2.2, the results differ a lot: If you do this change, make sure the training input and testing input are changed simultaneously. There is no good reason a deep learning system on linear images will work on Gamma-corrected ones, unless you do data augmentation on input image Gamma.

  1. How is human performance collected?

We developed a photo-editing UI to let humans play the same game as our RL agent, and recorded a video tutorial to teach our volunteers how to use it.


  title={Exposure: A White-Box Photo Post-Processing Framework},
  author={Hu, Yuanming and He, Hao and Xu, Chenxi and Wang, Baoyuan and Lin, Stephen},
  journal={ACM Transactions on Graphics (TOG)},

Related Research Projects and Implementations


Learning infinite-resolution image processing with GAN and RL from unpaired image datasets, using a differentiable photo editing model.








No packages published