Creative Process

Daytime Image

While on location, the art director captured a series of overlapping iPhone images covering the entire view. Although shooting handheld with your phone is not the most ideal method, having a significant amount of overlap from one image to the next will greatly help to reduce stitching errors in post, which was the case here. After receiving the images, we first merged them using panoramic software to create our starting point.

After identifying the position of the horizon line, we made our final crop based upon set design. This removed a decent amount of the bottom but tripled the height of the image creating a large void on top. To fix this, we dropped in a new sky from our catalog taking guidance from the art director and DP on the selection. Likewise, to help push the scale back a bit, an additional 5 foot sliver was generated in post by book-ending the left edge. Last, the encroachment of the porte cochere (left of center) was edited out of the image.

Nighttime Image

With the blessing of the art director, we then proceeded to use a combination of conventional methods with new AI tools to generate the night view.

Our first step was to crop out as much of the sky as we could without loosing any architectural information. Outputs in AI are limited in resolution and decreasing the height of the image as much as possible would allow us to work within this limitation. Plus, the sky in the night needs to match the day and it is fairly easy to edit the sky to appear as night.

Next, a series of overlapping plates were created from the this thinner panorama and processed through Adobe Firefly. This is not as straight forward as it sounds and one has to pay close attention to the prompts used not only for creative reasons but also issues regarding registration. After upscaling the outputs to match resolution, we layered and aligned them onto our thinner panorama before blending them together with masks and layer adjustments.

After taking our blended AI night plates and dropping them into the master file with the full height, we then moved onto using conventional methods to finalize the image. Our first step was to overlay the daytime sky and edit it to appear as night. Afterwards, we created a generic night from day conversion using a variety of LUTs, adjustment layers and hand painted light scallops before blending our AI night atop. Both of these not only eliminated the disjointed feel of the image but also helped set us up for the next phase.

Given the need for a perfect registration, we now scoured over the night image ensuring there were no mismatches from day to night. From a global perspective, AI can be quite good at creating a matching night version, however, at the micro level, slight differences can exist. We have found this commonly happens with text; occasionally the AI engine can not identify the exact font used and may replace it with a slightly different one. This was the case with the logo & lettering on the trailer (right of center) and a couple of the building signs. Also, in the distance, small changes occurred such as replacing a far off office building with a parking garage or reading roof vents as windows. Although, we suspect the lack of resolution and sharpness in this image probably added to these misinterpretations. Last, on the highway, AI kept adding in light streaks from cars no matter how many times we told it not to. All of this needed to be corrected for print, which we did by taking cuts from the daytime image and blending them into the night.

Although this is a tedious job, bare in mind we would have to do this anyway even with a custom day|night backing we captured ourselves. Over the course of a day, vehicles move and are swapped, leaves and branches change position, and light streaks from cars are hard to not capture at night. This may, or may not, be a little more involved with an AI Night, but the time saved from conventionally hand painting the night version is far greater.

To finalize for print, as with all backdrops, we first added a blur to help create the illusion of depth on stage. Since this backing was created from iPhone images, the blur was not as intense as with our captures, but one was stilled needed. Likewise, additional layers were added to compensate for CMYK conversions and the propensity for backings to wash out when backlit.