How do you encode a color palette into an AI video agent for consistent visual style?

Encode the palette as a named, hex-anchored block inside the agent's persistent context — 3–5 tonal modes (e.g. "Mode A — split-toned amber and emerald, #C9852B / #2F6F5E") plus dark-to-light ratio, lighting source, and explicit negative constraints. Load it once with the treatment, then prepend it to every shot prompt so every frame is graded against the same color script.

Start by writing the palette as a structured block the invideo agent can hold across the whole project, not as a sentence you re-type per shot. invideo is an agentic video tool that loads a treatment document once and keeps every directive — camera, lens, lighting, palette — locked across every generation, so palette encoding is really about what you put in that document.

Write 3–5 named tonal modes, each with hex anchors and a usage rule. A working block looks like:

PALETTE (locked)
Mode A — Split-toned amber & emerald | #C9852B warm key, #2F6F5E shadow | use for interiors, lamp-lit
Mode B — Cold arctic mist | #A8B7C4 mid, #E6ECF0 highlight, #1B2330 black | use for exteriors, night
Mode C — Burnt sienna dusk | #B3552E key, #5A2418 shadow | use for transitions, memory beats
Dark-to-light ratio: 85:15
Lighting source: practicals only (lamps, windows) — no top-down fill
Negative: no teal-orange Hollywood grade, no neon saturation, no flat daylight

Upload that as part of the treatment alongside the script, then instruct the agent verbatim: "deeply understand this and save it into context for further generations." The same prompt language one production used to lock an Arcane-style visual block worked because it forced persistent storage, not per-shot re-reading. From then on, every shot prompt the invideo agent assembles follows a fixed 9-element order — camera, lens, lighting source, palette, composition, atmosphere, mood, film attribution, negative prompt — so the palette slot is always populated by name ("Mode A") and the hexes ride along.

For reference-driven palettes (you have stills or a film you want the look from, not hex codes yet), don't drop the images into prompts directly. Instruct the agent to read the colors and textures off the references and translate them into the named modes and hex values above. One director put it plainly: "the better move was to have the agent read the colours and textures of them and prompt for that instead" — the resulting generations "came back hyper-realistic with the exact colour temperature I was looking for." Batch the references by purpose (one batch for palette, one for lighting, one for composition) and tell the agent what to take from each and what to ignore.

Keep the block enforced shot-by-shot. Run the invideo agent in always-ask mode so it shows you the assembled prompt — palette line included — before spending credits; if a shot drifts off-palette, the fix is in the character sheet or world reference, not a re-roll. Across documented productions, the same locked palette block held for a 70-second short produced for $750, a 3-minute Arcane-style episode at ~$315 per finished minute, and a 90-second horror short that ran an 85:15 dark-to-light ratio across every scene — same encoding pattern, different palettes.

For model routing: the invideo agent has every current video and image model available (Veo, Kling, Seedance 2.0 for video; Recraft, Nano Banana, GPT-Image-2 for stills) and routes each shot to the one that respects the palette block best — Seedance 2.0 reference-to-video carries color context across clips when you also pass a locked still as anchor, which is usually how palette holds through motion. You don't pick the model per shot; the agent does, against the same locked palette.

Beyond the encoding itself: treat the palette block as a reusable style bible. Save the named modes + hexes + ratio + negatives as a template you paste into the next project's treatment — same structure, swap the values. That is what makes the palette reproducible across films, not just shots.

Watch some of these to see what works for you:

The better move was to have Agent 1 read the colours and textures of them and prompt for that instead.

— invideo's creative team, on translating reference imagery into reproducible palette prompts