Where AI Failed

In the last post, I talked about the people and decisions behind HAPTIC. This time, I want to get into the part of using AI that can get lost when all you see is a completed film: the results that weren’t useful, the work that was still left to complete, and the places where I chose a different path. And, really, where AI just made things more difficult.

I’ve said that AI helped me make this film within the resources I had, which is true. BUT, it also took a lot of prompting, adjusting, reviewing, and trying again (and again, and again).

Getting an image wasn’t the same as getting the right image

The shot references are a good place to start, so we'll start there. As I mentioned in Part 3, I used AI to create a reference for every shot. Getting them into a relatively consistent storyboard format took quite a bit of trial and error. A good chunk of the early generations were downright creepy! And these were internal planning images - they weren’t even going into the finished film. I needed them to communicate what I wanted to shoot, which meant I still had to look at each result and decide whether it was useful, along with using them to communicate with others that were collaborating with me on the film (intimacy coordinator, wardrobe, makeup, etc.).

A picture can be perfectly presentable and still tell me very little about the shot I’m trying to plan. If it feels generic or doesn’t communicate the intention, being able to generate it quickly doesn’t solve the problem - I need a reference that helps me explain what I want the end result to look like, and more importantly, FEEL like.

The images eventually gave me something valuable to work with. Getting there required time I had to be willing to spend. I felt that time was worth it, but I wouldn’t describe the process as simply typing a prompt and being handed a finished storyboard - it was far from it.

The pieces still needed to fit

The same for post production: Creating a plate gave me an element to work with. It didn’t give me a finished scene.

For the opening sequence, I worked with separate ocean, storm, and raft elements. The actor was filmed, isolated in DaVinci Resolve Fusion, and combined with those plates. I assembled the composite and worked on color, lighting, reflected light from the water, reflected light from the lighting ON the water, and the other details needed to make the elements feel like they occupied the same space.

In the forest sequence, the actor was filmed against green screen. The sled, ground, trees, and surrounding environment were separate generated plates. Making the actor appear attached to the sled required painstaking manual rotoscoping, along with more color, lighting, and blending work.

Those are examples of AI contributing useful material while leaving a substantial amount of work for me. A generated environment doesn’t automatically know how to belong around a photographed performance. Simply handing over a green-screen actor and asking for a finished scene isn’t a reliable substitute for constructing it component by component. AI kind of sucks at this and I knew that going into it.

Sometimes I used another approach

The Monster Hands dream/nightmare sequence is a useful example here. No AI was used. I created it through manual compositing because AI would not have accomplished what I originally intended and understanding that I didn't need anything here, was also important.

It’s an example of a limit on where I considered the tool useful. I had something specific I wanted to create, and I did the effects work myself.

Having access to a tool doesn’t mean every scene has to pass through it. The point was to finish HAPTIC, and there were places where a different method made sense.

What did it actually save?

Access is the operative word here vs quantifying time saved. AI gave me visual references I couldn’t otherwise afford and elements I couldn’t have created in the same way with my budget. It also required iteration and extensive finishing work, which I felt was a reasonable trade-off for the ability to accomplish something within my budget.

For me, that’s the useful lesson in looking at where it failed or reached its limits. I still needed to know what I wanted, recognize when the result wasn’t serving it, and do the work that came next. “Just generate it” leaves out quite a lot of the job and I personally feel that if you're in that headspace, you're looking at creating content, not storytelling.

In the sixth and final post, I’ll come back to the question that follows all of this: would I use AI again? The answer depends on the job, the resources, and the people I could bring into the process.