agentleFS
PicsArtPicsArt/gen-ai-skills/skills/realistic-color-grade/SKILL.md

realistic-color-grade

Skill7 starsChanged 20 days agoMITView on GitHub

Make a video's colours look more realistic and natural with the ai_realism look — take the over-processed, plastic, AI-generated look off footage with a gentle grade, a shine correction on skin and light colour grain, and choose its values per video. Use when asked to make video colors more realistic, natural or camera-like, to make AI-generated footage look real, to take the glossy or plastic look off skin, or to add subtle grain for realism.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

5 rules. About 0 are common, 3 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Every pass is subtle on purpose.

    line 25
  2. Distinctive:

    Not the tool for:

    line 39
  3. Distinctive:

    Put the look on the media layer itself (an image or video layer), in reference mode, with the values chosen for this video:

    line 49
  4. Very specific:

    Look at the footage first: picsart_media_probe for its resolution, and picsart_scene_contact_sheet with times: [t1, t2, …] and resolution …

    line 64
  5. Very specific:

    The layer must render its own media: If the host layer also carries effects, or is a scene_ref, its source is drawn as a nested …

    line 83

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Colour realism with ai_realism

ai_realism is the look for "make the colours more realistic". It runs the three passes an editor does by hand on footage that looks too clean, too saturated or too glossy — typically AI-generated video:

  1. A gentle grade over the whole frame — contrast, highlights and saturation pulled down slightly, plus an optional white-balance nudge.
  2. A shine correction on skin — the skin's own hue is sampled every frame inside the portrait segmenter's skin matte, only bright skin is keyed, and its saturation, lightness and contrast come down, so the plastic, over-lit look goes. White shirts and warm walls are left alone.
  3. Light colour grain over the whole frame, as a camera sensor's.

Every pass is subtle on purpose. If the change is visible at a glance, it is too strong.

The footage decides the grain, not the wording: a request that names only the colours still gets grain when the footage is perfectly smooth, as AI renders are, and none when it is already noisy or the user asks for none. Say in your reply how much grain you added and why, and offer to remove it.

When it fits

  • Generated or heavily processed footage: waxy or glossy skin, punchy saturation, crushed highlights, no noise at all.
  • A request phrased as "more realistic", "more natural", "less AI", "like it was shot on a camera", "less plastic".

Not the tool for:

  • A stylised grade ("cinematic", "vintage", "teal and orange"). Use a LUT grade effect or a look such as duotone; picsart_scene_capabilities lists them.
  • Footage that already looks natural. The grade and the grain add little and can flatten it.
  • A fix for one colour only (an over-saturated sky). Use hsl_secondary directly, keyed on that colour.

Apply it

Put the look on the media layer itself (an image or video layer), in reference mode, with the values chosen for this video:

picsart_scene_apply_preset({
  "scene": ..., "layerId": "clip", "preset": "look:ai_realism", "mode": "reference",
  "params": { "strength": 1, "skin": 1, "grain": 0.06 }
})

Anything left out takes the look's default. picsart_scene_capabilities({ sections: ["look:ai_realism"] }) returns every parameter with its range.

Choose the values for this video

Look at the footage first: picsart_media_probe for its resolution, and picsart_scene_contact_sheet with times: [t1, t2, …] and resolution set to the scene's composition.width and composition.height to see the colour cast, how much skin shine there is and whether it already has noise.

ParameterDefaultChange it when
strength1The grade: highlights and saturation down, contrast barely, so blacks stay deep. Lower (0.4–0.7) for footage that is already soft or low in contrast; 0 skips it.
temperature0Negative (to about -0.3) to cool an orange, over-warm render; positive to warm a cold, blue one. Applies even at strength 0.
skin1The shine correction: bright skin's lightness and saturation down, its texture kept. 1 for glossy AI skin, about 0.5 for a light touch, 0 when nobody is in frame or the skin is already matte. In a dark scene it reaches only the brightest skin, as it should.
skinHue20Rarely. The skin's hue is sampled every frame; this is only the fallback when too little skin is visible to sample.
skinHueRange25Widen (30–40) when part of the shine stays; narrow (15–20) when lips or a warm background change too.
grain0.080.05–0.1 is felt rather than seen; from about 0.12 it shows on flat dark areas such as a night sky. Lower for a low-resolution source that will be scaled up (grain is per pixel, so it grows with the scale) and for footage with noise of its own; 0 for none.

Everything else is fixed inside the look (which tones count as shine, how much the key is softened). For control over those, see the last section.

Traps

  • The skin pass needs a person in frame. With nobody segmented it does nothing, so footage without people gets the grade and the grain only.
  • The layer must render its own media. If the host layer also carries effects, or is a scene_ref, its source is drawn as a nested composition, which is never segmented, so the shine correction is skipped. Keep the host's own effects empty, and put any other effect on an adjustment layer above it.
  • Some targets skip the skin pass. The grade and the grain still render there.
  • It is a static look. In a layer's looks, list it before any look that moves or pads its subject, or validation refuses the order.

Check the result

Render the same moments with and without the look, each with picsart_scene_contact_sheet with times: [t1, t2, …] and resolution set to the scene's composition.width and composition.height, using a frame where a face is large. Then check:

  • skin shine is calmer but skin still reads as skin, not grey;
  • white clothes, teeth, eyes and warm backgrounds are unchanged;
  • the grain is barely visible at full size and does not shimmer as blocks;
  • the whole frame has not gone flat. If it has, lower strength.

Beyond the six parameters

The look expands to three effects on the media layer, in order: adjust (the grade), hsl_secondary (the shine key and its correction) and film_grain (with color 1). To tune what the look fixes — which brightness counts as shine, the key's blur, the grade's balance of contrast and highlights — detach it with picsart_scene_resolve_looks, or picsart_scene_apply_preset with mode: "expand", and edit those effects. Their parameters are in picsart_scene_capabilities. Each works alone too: film_grain with color 1 for grain only, hsl_secondary for any one colour.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.