Checked the implementation of the experiment, fixed a few things (max bounce, spp, area light placement). And trained 3 different scenes similar to paper (cornelbox with glass objects, kitchen, living room). Also improved our dataset viewer to clearly see all inputs, positional encodings, and clearly stated error images with error heatmaps. #### Sampling the Camera and Area Light Paper states that it sampled the location of camera and its view direction, and the area lights from a bounding box (randomly sampled) we followed the same procedure however paper doesn't state the location or extent of the bounding box, so we choose some place as we like. #### Rendering of Direct Illumination Paper stated it used Linearly transformed cosines (for lights) and ratio estimators for soft shadows, however we had the opportunity to render the training scenes offline so we rendered them as the first bounce of the lights and denoised the result with OptiX denoiser. #### Normalization of Values Paper doesn't state how it normalizes position values, we used the scenes bounding box to normalize the position values. For radiance channels we used log scaling with the max radiance values. #### Radiance Cues Paper states that it only changed the non-diffuse materials with different pregiven material roughness values with cook-tarrance (and one lambertian model) so we followed that with mitsuba's classifiers. Also the radiance cues are just direct illumination samples, so no indirect lighting is present in them. #### Perfect Speculars Paper states that to capture details in the mirror like objects, it changes the values of these objects with the first reflected hit from the camera, we followed exactly. #### Sample Count We followed what is stated in the paper, we picked our training images with 1024 spp and 256x256 rendered images. #### Positional Encoding on Inputs Paper increases the input channels by applying a mapping which creates multiple frequency layers with sines and cosines. (10 to be exact, sin and cos, resulting in 680 additional channels for our 34 channel input). #### UNet and Encoded Inputs Paper states that to capture effects of nearby pixels it puts all the inputs into the convolutional layer. We followed and put all of our input channels into the UNet.