AI Solution for Townscape Background Art

Over the last decade, our studio has been creating 2D backgrounds for various games. Townscape backgrounds are a very common type of asset.

This is a San Francisco street scene created by us. Typically, the workflow is as follows:

1. 3D modeling, texturing, and rendering the scene.
2. Painting over the 3D rendered image.

For a picture like this, the combined 3D and 2D work would take around 5–6 days. Now, while AI can accomplish the same task, clients might be reluctant to pay as much as before.

By AI You can quickly generate many beautiful, random images of San Francisco townscapes. However, in video game production, certain game design specifications must be met, regardless of whether the artwork is produced traditionally or using AI. For this particular background, the specifications include:

1. General composition must follow the game designer’s provided draft. For example, a typical three-story San Francisco house on the right side of the frame, a street with a cable car on the left, and a crosswalk in the foreground.
2. The style of buildings and cable cars.
3. Colors and textures of specific components.
4. Overall lighting and color scheme.

Meeting these requirements can be somewhat challenging with AI. Below is our AI-assisted workflow:

Step 1: Sketch to Image

?For the first step, we use OpenArt’s Sketch-to-Image feature to create a photorealistic-style image. This tool requires an initial sketch, which can either be hand-drawn line art or a 3D image.

Currently, we prefer creating a 3D draft rather than a hand-drawn sketch, as it is faster and easier to adjust. Here is our initial draft. For OpenArt, the 3D model doesn't need to be high-resolution—basic boxy shapes for buildings are sufficient. However, it’s essential to maintain accurate proportions, such as the dimensions of key features like windows, because the AI's Sketch-to-Image process closely follows the general form without significant alterations.

On the street to the left, we place a cable car from stock 3D models. The style doesn't need to be exact—mentioning "San Francisco" in the prompt will prompt AI to adjust the style appropriately during repainting. Near the left edge of the frame, we only place simple boxes, AI will figure them out.

This stage took about 3 hours in total. Building the 3D models consumed a relatively small portion of this time; much of it was dedicated to the design phase, adjusting composition, proportions, and styles.

We would then send this 3D draft to the client for approval, render an Ambient Occlusion (AO) image, and upload it to OpenArt’s Sketch-to-Image feature.

We select a realistic style for repainting. In the prompt, we include essential keywords like "San Francisco street" and "sunny daytime." We set a low "Creativity" value (around 0.3–0.4) to preserve the composition. Below is the resulting image:

The lighting and mood meet our requirements. For specific game projects, if component colors need adjusting—for example, changing the main building walls to blue—we can easily achieve this by roughly painting the wall blue in Photoshop, then running the image again through OpenArt's Image-to-Image feature with a low "Creativity" setting to maintain key details.

Step 2: AI Filter

We now have a 3D-rendered, photorealistic image. From here, we can either repaint it into a digital painting style commonly used in interactive story apps or convert it into an anime style popular in visual novels.

In this example, we choose to convert it into an anime-style background using OpenArt’s AI filter, selecting "Anime" as the output style. Below is the result:

Alternatively, to match the unique style of your game, you can train a custom style model within OpenArt and use the Image-to-Image feature to repaint the image accordingly.

Step 3: Upscale the Image

Currently, we have a low-resolution image. We upscale it in OpenArt to up to 4K resolution using the Ultimate Upscale feature, selecting "Refined Upscale" mode and setting the enhancement level to around 0.4 to refine details.

Step 4: Digital Paint-over by Hand

Finally, traditional paint-over is necessary to correct minor issues, such as tidying up the shop interior and crosswalk details. Additionally, we add two tiny stoppers on the cable car tracks. We can further enhance the overall picture by adjusting the values and hues.

Another Sample:

This is another sample background, depicting a university facade.

Similarly, the first step involves creating a rough 3D model, taking around 5 hours, including design and adjustments. Again, simple boxy modeling is sufficient.

You may observe the difference in details between 3D model and final art, this is what always happen in repainting.

Here, you can observe the difference between the AI-enhanced workflow and the traditional approach. Traditionally, we would paint over a high-resolution 3D render to produce the final art. Building such a detailed 3D scene for this facade would typically take around two days.