Case study · 2026
How I used Higgsfield AI to build a shot that did not exist
There is a fine line between using AI as a tool and using it to make slop. Over the past few months I have directed and edited a run of high-budget commercial work, and on nearly all of it I have folded AI into the workflow. This is the hybrid model I landed on, run on one shot from a Forza spec, and how I get away with it.
The rule I keep coming back to: AI should not replace real life. It should enhance it, the same way VFX always has. Mix AI elements into real shots and you get a better result with more control and fewer headaches than letting the model drive the whole frame.
Here is the shot I needed, why it did not exist, and how I built it.
The problem
The shot I needed was never filmed
The prior shot in the cut establishes the subject: a man standing on the bleachers, watching a car run the track. I needed a wide that put him and the track in the same frame. The trouble is, all I had was the wide of the empty track. No man in it. And reshooting was not on the table.
The car drives through the frame, the camera is a slow handheld. I wanted the bleachers in the foreground and the action happening out in front of him. So the job was not to generate a new shot. It was to place my real subject into a real plate that never had him in it.
01
Render the top image in Higgsfield
I took a still of the first and last frame of the shot, dropped them into Higgsfield Cinema Studio as the start and end frames, and uploaded a reference image of my actor. Then I described the scene and generated a handful of sample videos.
The output I am after is what I call the top image. It is the element I want to bake into the plate. In this case, the man on the bleachers. Grade your shot before you generate, because you lose color control once it is in the model. Pick a generation whose camera movement matches the original, and watch the actor's position and motion so the continuity holds against the shot before it.
02
Be picky about the take
This is the part to slow down on. Generate, alter the prompt, generate again, until you have a take that almost exactly matches the shot you put in. Little details earn their keep here. The arm resting on the fence is good, for example, because it matches the body language of the prior shot.
By now you have something that looks close. But no model is perfect. The generation is likely 1080p, it carries artifacts, and it is missing details from the original. Look closely at mine and you will catch it.
The catch
There is no car
This is usually where people stop, and it is the birthplace of slop. The model gets fine detail and movement wrong, and most editors shrug and settle. Mine built a convincing man on the bleachers and quietly deleted the car off the track. Ship that and it reads as fake instantly.
The generation's real value was never the whole frame. It was the man. So I kept him and threw the rest away, then put the real plate back underneath.
03
Mask the subject out: half AI, half Arri
The goal was never to replace the image. It was to composite one character into it while keeping the quality of the original. So I drew a simple mask around the man and pulled him out of the generation, turning the frame into half AI and half Arri footage.
AI works best when it is not the focal point.
It is a great supporting actor and a bad lead. This shot works because the character sits blurred in the foreground. Your eye is going to the car on the track, so the man does not need to hold up as razor-sharp 4K. I laid the AI image over the original and masked the subject out of it, cutting him loose from the generation and dropping him into the real world.
04
Build a Fusion clip, meet the new problem
I selected the AI clip and the real plate, made a Fusion clip out of the two, and opened it on the Fusion page. The color matched, the shadows lined up, the blur read as real. That single move saved me four hours of modeling and compositing, and the car was back in frame from the real plate.
The new problem: the car was getting clipped at the edge of my subject mask, and the camera move on the AI layer did not perfectly track the background. That is fine. That is what the next step fixes.
05
Track the AI layer to the real plate
Inside the Fusion clip I added a Planar Tracker to the real background clip, drew a track mask, hit set, and switched the motion type to translation, rotation, scale. Track forward, track back, then add that planar transform to the AI clip in the foreground.
When it finishes, the AI layer is locked to the background camera movement. The man stops floating and starts riding the same handheld motion as the plate he is standing in.
06
Clean the mask
With the layers tracked, I went back into the mask to tidy the edges. Soften where it should blend, stay rigid where the scene demands it. The fence post next to him wants a hard edge so it reads as a real object. Other areas want a softer falloff so the cutout disappears into the rest of the frame.
07
Put the car back behind him
In real life the car drives behind the actor and the bleachers, not in front of them. So I copied the real background plate up over the Fusion clip and cut it into the frames where the car is in view. I separate that section into individual frames, mask the car, then soften and clean those masks.
The result is a mask that pushes the car back into the distance, passing behind my newly composited subject the way it actually would. Foreground man, midground bleachers, background car. The depth order is correct again.
We blended AI and real life, and the seam is gone. The more I work with these tools the more obvious it gets that they are just that: tools. The control you have in the edit plus the speed you get from a good generation is where the real leverage is. It only takes a few extra steps to dodge the slop and land something that actually holds up on a high-budget cut.
Higgsfield built the wide.
I kept the man.
Elsewhere