Earlier this year, I used AI to finish the art for a 2D game and released its demo on Steam. The workflow was already advanced for its time: AI generated characters, backgrounds, and motion, while I cut, completed, keyed, aligned, and imported everything into the engine.
Looking back, that AI behaved more like cruise control. It worked only during the generation step. Organizing, editing, integration, and checking were still manual.
When I started a new bottom-of-screen idle driving game in July, I brought Codex into the art workflow. The important change was not simply a stronger image model. Image generation, project context, local files, traditional editing tools, and code could now participate in one connected process.
The old workflow was cruise control. The new one feels closer to L2 autonomous driving: I choose the destination and judge the result, while AI handles most of the route and asks me to take over when needed.
The real change is not that AI draws better
A standalone image model usually handles a single request: give it a prompt and it returns an image. If one part is wrong, you change the prompt and roll the dice again—often losing the parts you already liked.
Once image generation is connected to Codex, I no longer have only an output window. I have a collaborator that can participate in the full task. Four capabilities have mattered most.
1. It understands project context
I can refine an image through natural conversation: what must stay, what should change, what kind of game the image belongs to, and which visual and functional constraints it must satisfy. The discussion can include the project's direction, not just the current prompt.
2. It can review its own output
Previously, the task ended as soon as AI produced an image. Now the workflow can inspect whether the result meets the requirement and attempt a correction when it does not.
For example, an image requested with a transparent background may contain a fake checkerboard instead. I used to generate a green screen and remove it myself. The new workflow can identify the problem, use an image-processing step to remove the background properly, and deliver a usable file.
This does not mean AI always judges correctly. It means review and revision can become steps in the pipeline instead of always beginning with manual cleanup.
3. It can handle files and local tools
Generation is only the beginning. Images still need naming, filing, resizing, layer separation, variants, and engine integration. Those steps used to be scattered across the Downloads folder, Photoshop, and the game engine.
Codex can now organize results inside the project and choose between generative methods and more deterministic editing tools. Generation handles creative variation; traditional processing handles stable repetition. The two approaches complement each other instead of competing.
4. It can build a purpose-made tool
Some tasks are difficult to describe precisely and cannot be fully generated: pixel-level placement, wheel size, light range, or whether a composition simply feels right.
Instead of slowly adjusting everything in Photoshop or the engine, I can ask AI to make a small utility for that exact pipeline. It exposes the parameters I need, I adjust them by eye, and the automated process continues with the approved values.
Three real pipelines show how “autopilot” works
The new game includes interchangeable vehicles, body colors, wheels, lights, day and night driving backgrounds, parallax layers, interiors, city maps, and country maps. One universal prompt cannot reliably produce all of them, so different tasks use different pipelines.
Vehicle bodies: from design approval to a car running in the game
The vehicle pipeline first creates a design-confirmation image. After approval, it removes the wheels, completes the wheel wells, and generates the lighting and headlight layers.
Then it opens the positioning tool. I adjust body and wheel placement, the area of the lights, shadow scale, and overall composition while seeing the real background and a translucent reference vehicle.
After I approve the position, the pipeline generates 16 day and night color variants, creates UI icons, and integrates the result into the engine. I do not need to open Photoshop or the game engine manually before the new car appears in the game.
Driving scenes: from one reference image to moving day and night layers
A driving background must satisfy more than appearance: the palette, road position, building perspective, and matching night version all need to remain consistent.
It also has to move. The workflow identifies foreground, midground, and background elements and prepares continuous layers for parallax scrolling. Codex can already handle most of the separation; when placement is imperfect, I adjust each layer in a positioning tool before the workflow integrates it and configures the relative scrolling speed.
Wheels: when generative AI is the wrong tool for geometry
A wheel must rotate continuously. Generative models struggle to guarantee a perfectly regular circle or directionless lighting. Those errors may be invisible in a still image but become obvious as soon as the wheel turns.
Instead of generating more candidates, I changed the technical route. I provide the wheel style, the workflow finds structural references, and then Blender builds and renders reliable circular geometry.
Uniform lighting prevents visible seams but also makes the wheel feel flat. The solution was to generate a normal map, rotate it together with the wheel image, and let Godot's 2D lighting calculate a dynamic shadow.
I divide complex work into four categories
After building these pipelines, I summarized AI's role in the work as four categories: generation, automation, human takeover, and human decision-making.
- 01GenerationUse generative AI when the task requires genuinely new images, video, or other material.
- 02AutomationLet AI operate traditional tools when the result must be stable, deterministic, and repeatable.
- 03Human takeoverWhen the first two methods cannot solve the problem, I step in—ideally through a purpose-made tool rather than repeatedly learning a complex professional interface.
- 04Human decisionsI retain the choice of what to make, whether the direction is right, and whether the result is good enough.
These responsibilities should not collapse into the sentence “let AI do it.” Generation seeks creativity, automation seeks certainty, takeover handles exceptions that are hard to express, and decisions determine who is actually making the work.
Useful automation does not remove the person from creation. It releases human attention from repetition so it can remain with taste, selection, and responsibility.
The real benefit is parallel creation
When art is no longer a chain of operations that I must perform one by one, I can move more work forward at the same time. Programming, art, story, UI, and other parts of the project now progress through their own pipelines. I even wrote and recorded the video while directing Codex to work on project tasks.
That is why I call the workflow “autopilot.” I still choose the destination and remain responsible for the result. AI handles most of the route and asks me to take over when it reaches something it cannot solve.
It is neither cruise control that works on one short stretch nor handing over the steering wheel without judgment. For an indie developer, this middle path is how AI can grow from “able to generate an image” into “able to help me finish the work.”