AI color grading video workflow: match any reference look

The fastest way to color grade a video with AI is to hand it a picture instead of a description. This is one grade followed from start to finish, run in the Video Color Grading workflow in Picsart Flow.

The source is an eight-second clip shot in flat natural daylight. The reference is a single soft pastel still. Between them sits a chain that reads the reference image, writes out what it finds in plain language, grades one frame, and then carries that result across the entire clip. Watching a real run end to end shows more than a feature list can, because the interesting decisions all happen in the middle and every one is visible while it is made. By the end the footage has kept its subject, its framing and its motion, and wears a palette it was never shot in. Nothing about the content itself is replaced along the way, which is the whole trick.

The clip and the look it needed

The source footage is ordinary in the best way. Two figures ride a cable car through alpine scenery, and the color is exactly what the camera saw: earthy greens in the trees, cool blues in the sky, neutral daylight on both faces. Nothing is wrong with it. It simply looks like footage rather than like a film.

The reference is a still from a different world. Soft pastel throughout, low contrast, pinks against a pale cyan sky, everything flattened into something closer to illustration than photograph. The distance between those two images is the entire job.

Loading them takes a moment. Open the workflow, clone it into the Flow editor, then drop the clip into the input video node and the still into the reference image node. The reference is the choice that matters. A picture with a clear dominant palette, consistent lighting and an obvious contrast profile carries across cleanly, while a busy image with six competing colors gives the workflow nothing definite to hold.

What the AI saw in the reference

Running it produces something unexpected before any picture appears. The color analysis node returns a written breakdown of the reference under five headings, and it reads like notes from a colorist rather than machine output.

Overall tone comes first: high-key, surrealist pastel with a dreamlike softness. Dominant palette follows, naming actual colors rather than vague families, calling out a wash of rose quartz and bubblegum pink, accented by serenity blue and pale cyan, grounded by neutral taupe. Contrast levels land next, flagged as low contrast with a flat lighting profile.

Then it splits highlights and shadows apart, which is the moment it stops sounding automated. Shadows are significantly lifted with a strong magenta bias, eliminating true blacks. Highlights roll off soft and creamy, with no harsh clipping. Color harmony closes the read, naming a pastel complementary pairing of pink against cyan with saturation matched so neither hue overpowers the other.

Reading that back is the first checkpoint of the run. The reference has been understood correctly, so whatever follows will at least be aiming at the right target.

The grade written out in words

That analysis does not stay a description. It becomes the instruction, assembled into a prompt built in two labeled halves, and the split is the reason the footage survives what happens next.

The first half is an anchor describing the source rather than the destination. It records the alpine palette with its earthy greens and cool sky blues, a full dynamic range with defined shadows, neutral daylight exposure, and naturally warm skin tones. None of that is the look being applied. It is a careful statement of the starting point.

The second half is the color style transfer, describing the destination in the same specific terms: a shift to a soft pastel palette dominated by blush pinks and warm peach, with muted lavender undertones replacing the natural greens. Naming both ends converts a vague wish into a controlled move, since the model knows what it is changing as well as what it is changing into. The opening line makes the constraint explicit, asking for a transformation that preserves all original content and luminance structure.

One frame, then the whole clip

Here is the part that surprises people. The workflow never grades a video. It grades one picture.

A frame extractor pulls a single still out of the source clip, and that one still is what Nano Banana Pro grades using the assembled instruction. The output is a colored frame showing your own footage, your own composition, wearing the reference palette. Kling 3 Omni then produces the finished video from that graded frame and the original clip.

Checking the colored frame before looking at the video is the second checkpoint, and the more useful one. A grade that misses shows up in a still, which costs far less than a full render, and swapping the reference is the fastest correction available. Exporting the finished clip usually requires signing in.

Settling the look once also keeps the result stable. A model grading a whole clip from scratch makes the same decisions over and over, and small differences between frames become flicker.

What changed and what did not

The finished clip is the same eight seconds. The same two figures ride the same cable car through the same mountains, framing and motion identical, and the shot is now blush pink and pale cyan, shadows lifted, contrast flattened into the illustrated quality the reference had.

What did not move is the point. Subject and composition are untouched, and the luminance structure telling the eye where light comes from survives, which is why the result still reads as footage rather than a reimagined scene. Video color grading fails most often by quietly rewriting content along with color, and the anchor half of that prompt holds the line.

Worth being straight about the limits. This is a fast, repeatable way to put a look on a clip, and it is not a substitute for a colorist grading a finished production shot by shot. Pastel color grading makes a good first experiment, because soft palettes declare themselves clearly and the change is obvious on screen.

Start grading your videos

One image decided everything in that run, and reading the analysis before the render is what made it predictable. Choose a still whose color you already want, drop in a clip, and check the frame before the video. Open the workflow in Picsart Flow and grade your first clip, then keep the references that work as a palette library for everything you shoot next.