The original sunset shot. A cyclist on a forest road, warm pink and orange light across the trees and grass.
Before · sunset original
The same shot regenerated by Higgsfield AI in overcast weather. Gray sky, fog through the trees, flat light, identical camera move.
After · Higgsfield AI + comp

Case study · 2026

How I used Higgsfield AI to fix an unfixable shot

Jacob Geeslin·Filmmaker, founder of Cut Time
Sponsored byHiggsfield

The shot I needed didn't exist. I had a take of a cyclist on a forest road at sunset. Warm light, long shadows, all of it. The next scene in my edit was overcast and gray. Cutting between the two would feel violent.

Reshooting wasn't on the table. Neither was twenty hours of frame-by-frame VFX to re-sky and re-grade the original.

What worked was Higgsfield AI. You feed it a video and a few stills of the look you want, and it generates a new version of your shot in that look. Same camera move, same framing, new world.

It worked. With one big problem. Here is what I did, in order.

01

Export the base plate

I pulled the cyclist take out of Resolve as its own self-contained file at full resolution. This is the base plate. It is the foundation Higgsfield uses to learn the camera move, the framing, and the composition.

Whatever motion you want preserved has to be in the base plate. Higgsfield respects what is in the source clip. If your camera move is locked here, the AI version will be locked too. If your source has a wobble, the AI will faithfully recreate that wobble.

The base plate. Sunset cyclist on the forest road, full frame.
The base plate I sent into Higgsfield. Full-resolution export of the original take, no grade applied.

02

Match the destination look

The next scene in my edit was the look I needed to land in. I grabbed reference stills from it: foggy redwoods, cool ambient light, the kind of overcast that you cannot fake in post.

Pick stills where the look is unambiguous. If the destination scene has a specific tone in the trees, find a frame where that tone is obvious. The AI uses these as the targets for the regeneration, so generic stills give you generic output.

Reference still. A figure crouches at the base of a giant redwood, cool flat overcast light through the canopy.
One of the reference stills I fed Higgsfield. Cool ambient light, foggy mid-tones, no warm color.

03

Generate the new-weather version

I loaded the base plate and the reference stills into Higgsfield. The output is a new version of the shot, same camera move and framing, rendered in the look of the references.

Sunset cyclist became overcast cyclist. Same road, same motion through the trees, gray sky, flat light. Cuts straight into the next scene. Nothing in the wide frame screams that the weather was rewritten in software.

Here is the exact prompt I used, pasted verbatim. No prose tricks, no engineered phrasing. Specificity beats style.

The exact prompt

Higgsfield · Kling O1

change the time of day in the video to match the reference image 10am redwood national park fog. maintain original image motion, sharpness and quality

Model

Kling O1

Feature

Edit

Output

1920 × 1080

Inputs

Video + 1 still

The Higgsfield AI output. Same composition as the original, now with overcast sky, fog, and cool flat light.
Higgsfield's output. Same camera move, same framing, completely different world.

The problem

The AI butchered the cyclist

This is where AI generation falls apart. It handles wide shapes well. Sky, road, treeline. It does not handle fine moving detail. My cyclist came out warped. Limbs in wrong positions. Helmet half-resolved. Face smudged into nothing.

Shipping that is exactly how AI work starts looking like slop. So I didn't ship it.

The AI plate in motion. Watch the cyclist on the road. Limbs distort, the helmet shifts, the figure stops feeling human.

The AI plate's background was right. The cyclist was the only thing wrong. So I kept the real cyclist from the original take and used the AI plate as the world around him.

04

Stack the layers in Resolve

Imported the AI clip into the project and dropped it on the timeline. Original on V2, AI on V1. The original sits above the AI in the layer stack.

Order matters because Fusion uses the timeline order to lay out the source clips inside the comp. Stack them how you want them ordered before you open Fusion.

05

Make a Fusion Clip

Selected both clips, right-clicked, chose New Fusion Clip. This compounds the two layers into a single clip on the timeline that you can open in the Fusion page and treat as one composite.

If you have never done this: a Fusion Clip is a container. Both source clips are still there underneath. Inside the container they show up as MediaIn nodes you can wire together however you want.

I double-clicked the Fusion clip and switched to the Fusion page tab at the bottom of Resolve.

06

Track the AI road

Inside Fusion you get two MediaIn nodes, one per source clip. I hooked a Planar Tracker to the AI clip's MediaIn output, drew a tracking region on the road, and picked an area with enough texture for the tracker to lock onto. Tracking mode set to Translation, Rotation, Scale, then let it run.

Track the AI, not the original.

You are not aligning the two clips with each other. You are capturing the AI road's motion data so you can drive a mask along it later. The patch has to follow the world it is landing in, not the world it came from.

Fusion Planar Tracker view. A green rectangular tracking region is drawn on the AI plate's road.
The Planar Tracker region on the AI road. Enough texture for the tracker to lock onto, no moving objects inside the box.

07

Cut around the cyclist and a slice of road

I added a Polygon node and traced the shape of the patch. The patch is not just the cyclist. It includes a slice of road under him.

Cutting tight to the rider is the obvious move and it is wrong. The seam where his tires meet the AI road is the most visible failure point in the comp. Tracking errors and color mismatches will scream there. Including a chunk of original road beneath his wheels means the cyclist sits on real footage, and the seam moves to a place where it is easier to hide.

I feathered the polygon edges. A hard edge reads as a cutout. A soft edge blends.

Fusion polygon mask traced around the cyclist plus a long slice of road on either side.
The polygon mask. The rider is dead-center, and a generous slice of original road wraps above and below him so the seam never sits at his tires.

08

Lock the cutout to the AI road

I added a Planar Transform node, fed the polygon into it, and drove it with the data from the Planar Tracker in step 06. The mask now follows the AI road wherever the camera moves.

Without this, the cutout would sit static while the AI background slid underneath. The cyclist would drift off the road within half a second.

The masked cutout in isolation. The cyclist and a slice of road float on a transparent background, still in the warm sunset color of the original.
The cutout, alone. Real cyclist, real slice of road, still in the original sunset grade. This is what gets composited on top of the AI plate.

09

Color-correct the patch

The patch is sunset. Warm, golden, contrasty. The background is overcast AI. Cool, flat. They do not match out of the box.

I added a ColorCorrector on the original clip's branch and used the tracked polygon as its effect mask, so the color shift only applies inside the patch. Pulled temperature, contrast, and saturation toward the AI plate. Got it close, then let my eye catch the rest.

The sunset-colored cyclist patch sitting on top of the cool overcast AI road. The orange warmth of the original reads as obviously out of place.
Patch laid on the AI road before color correction. The seam is invisible on shape, but the temperature mismatch makes the whole thing feel pasted in. This is what the ColorCorrector solves.

10

Keyframe the color across the shot

The AI plate's grade is not static. The light shifts subtly across the clip and a single fixed correction will not hold up.

I set values at the start of the clip, scrubbed forward to a point where the AI grade had drifted, and set new values. Repeated through the shot. Resolve interpolates between keyframes and the patch follows the AI's grade as it evolves.

11

Merge and finish

A Merge node with the AI plate as the background and the masked, color-matched original on top. Output to MediaOut, back to the timeline.

I went back to the Edit page. The Fusion Clip plays back the composite. Real cyclist, real road slice, AI sky, AI everything else.

The full Fusion node graph. MediaIn1 (AI) feeds a Planar Tracker. Polygon1 feeds a Planar Transform driven by that tracker. MediaIn2 (original) goes through a ColorCorrector masked by the tracked polygon. Two Merge nodes combine everything.
The full node graph. Two MediaIns, one Planar Tracker, one Polygon driven by a Planar Transform, one masked ColorCorrector, two Merges.
The finished comp. Real cyclist riding through AI weather. Cuts into the next scene without anyone in the room asking how I got the weather to match.

See it in the cut

The full film

The shot in this case study lives somewhere in the middle. See if you can pick it out.

AI is good at filling in worlds. It is not good at the thing in the middle of one. If you can isolate the part it cannot do and replace it with real footage, you get a lot of mileage out of an imperfect generation.

Higgsfield got the sky.
I kept the cyclist.

Elsewhere