AI Color Grading: How to Make AI Video Look Cinematic

Color grading AI video is two jobs, not one. First, control the look at generation: encode your palette as named tonal modes with exact hex values, write lighting in numbers — one documented production locked an 85:15 dark-to-light ratio — and carry those rules in every prompt. Second, run a structured post grade: Topaz Astra upscaling first, then color, then light blur and film grain to correct the over-sharp, plasticky skin AI footage carries. LUTs apply at 50–70% intensity, never full strength.

Why color grading AI video is a different job

AI-generated footage arrives with its look already baked in. There is no RAW file, no log curve, no flat profile waiting for a colorist — the model decided your exposure, palette, and contrast at render time. That moves half the grading job upstream: the most important color decisions happen in the prompt and the context the model generates from, not in a grading suite afterward.

The second half of the job is corrective rather than creative. Current video models — Seedance 2.0 in particular — output an ultra-sharp, plasticky skin quality that reads as synthetic on a big screen, and clips generated minutes apart can drift in white balance and contrast. Post-production processing is a critical, under-discussed step for making AI footage look closer to live action: the documented fix is a sequential pipeline of upscaling, color work, light blur, and grain.

One orientation note before the workflow: invideo is an agentic video creation tool with all the current video and image models plus upscalers like Topaz Astra available in one place, and the invideo agent holds your color rules in persistent context so every shot inherits them. The two stages below — generation-time look control, then the post grade — are how documented productions ran it.

Stage one: control the look inside generation

A clip generated in the wrong palette cannot be fully rescued in post — secondary correction on heavily compressed AI footage falls apart fast. Lock the look before you spend generation credits, using the six techniques below in order.

Encode the look as named tonal modes with hex values

Define your film's palette as named tonal modes with exact hex values — for example, "Mode A — Split-toned amber and emerald" with the specific hex codes for shadows, midtones, and highlights. Encoding a filmmaker's color philosophy this way makes palette control reproducible: the same mode name in every prompt returns the same color world, shot after shot. Load the modes into the invideo agent's context once and reference them by name; in one documented production the agent held a 9-element prompt assembly order where palette occupied a fixed slot in every single generation, which is what kept the color identical across 21+ scenes.

Translate a director's lighting grammar into numbers

Generic adjectives like "moody" or "warm lighting" give the model nothing to hold. Convert lighting intent into numbers and sources: one production encoded James Wan's signature lighting grammar as an 85:15 dark-to-light ratio and used that ratio as literal prompt language for every frame. Specify the light source explicitly and tie it to your references — "warm yellow from the lamps only, like all the refs" produces measurably more accurate results than "warm lighting."

Lens and format language belongs in the same pass, because it changes how color and flare render. Spherical lenses produce circular bokeh and no horizontal lens flares — a meaningful distinction when replicating a director's look — and the same production specified a 2.40:1 hard matte because that was the actual shooting format of the reference films. Challenge the invideo agent's technical claims before locking anything: in that production the agent had initially noted "anamorphic" in its analysis and corrected to spherical when questioned, catching the error before it propagated across the asset pipeline.

Use a reference image for the palette — but know what fails

Dropping illustrated or animated reference images directly into prompts does not work — the model imitates the medium instead of extracting the color. The working method: instruct the invideo agent to read the colour palette and texture qualities of the reference and translate those into photorealistic prompt language. From the production that documented this: "The gens came back hyper-realistic with the exact colour temperature I was looking for."

When you have many references, batch them by theme — a color-theory batch separate from spatial or composition batches — and tell the invideo agent explicitly what to adopt and what to ignore from each batch. Telling the model what to leave out of a reference is as important as telling it what to take.

Write it once as a visual language document

Put the entire color system in one document and load it once, instead of re-typing palette rules into every prompt. One documented production built a 14-section visual language document covering colour tone, lighting, atmosphere, film palettes, prompt templates, and negative prompts. The principle, as the team behind the agent put it: "One agent that reads your treatment once and holds every directive across every shot, every scene. No re-prompting. No drift."

Two details make the document work harder for color specifically. First, include a "what never to do" list per mood or emotional stage — explicit prohibitions make the invideo agent's autonomous color decisions reliable. Second, have the agent output a color script and lighting plan as named parameters for every shot; one production required 12 parameters per shot, with color script, lighting plan, and atmosphere layers as standing entries, so no shot ever generated without a color decision attached.

Use cinematic vocabulary the model already understands

Video models are trained on cinema, so cinematographer and film-stock language steers color more precisely than plain description. The documented 9-element prompt assembly order is a usable template: camera spec, lens and aspect ratio, lighting source, palette, composition, atmosphere, mood register, film/DP attribution, negative prompt — in that sequence, every time. The film/DP attribution slot is the shortcut: naming a film's photographic style compresses an entire color world into a few words.

The negative prompt is your drift insurance. One animated production's standing style block read: "This MUST look and feel like Arcane animation — not live action, not photorealistic. Every surface has hand-painted brushstroke texture." Explicit prohibitions of the looks you don't want prevent the model from sliding back toward its default rendering — and the block prefixed 100% of prompts: "Every prompt after this started with it."

Validate the look before you generate the whole film

Test the color system on a handful of frames before committing credits to full generation. Three documented validation methods: generate identical script frames in each candidate style side-by-side before choosing; generate four options per environment and character reference, select one, and lock it before any video generation begins; and stress-test your visual language document by asking the invideo agent to apply the look to a genre or subject the reference director never worked in — coherent output and clarifying questions confirm the document is internalized as grammar, not surface decoration. One production also uploaded 64 style-reference frames in a single message with the instruction to deeply analyze and save the style to context — the locked style then survived 164 generations without drift.

Stage two: the 8-step color grade in post

Generation-time control gets you consistency; the post grade gets you cinema. In a documented ~90-second horror short produced in a James Wan directorial style, the invideo agent supplied an 8-step color grading guidance process alongside its 9-step shot design process — ask the invideo agent for grading guidance against your loaded look document and it returns a structured pass order rather than generic advice.

The practical sequence the documented pipelines follow:

1. Upscale first. Run Topaz Astra on invideo before any color work — grading low-resolution footage and then upscaling amplifies artifacts; upscaling first gives the grade clean pixels.
2. Match shots before you stylize. Clips from different generations — and shots stitched together from multiple takes — arrive with mismatched exposure and white balance; neutralize those differences across the timeline first.
3. Apply the look. Bring in your named tonal mode — the same hex values from generation — so the post grade reinforces the prompt-time palette instead of fighting it. Crush the blacks deliberately if your look calls for it; AI footage tends toward lifted, milky shadows.
4. Refine skin and key surfaces. Secondary corrections target the synthetic rendering — pull saturation and sharpness out of skin specifically.
5. Finish with texture. Blur and grain go last, on top of the completed grade (covered in the realism pipeline below).

After assembly, send the rough cut back to the invideo agent with an open-ended "what's working, what's not" prompt — the documented maker-checker pass catches grading and register mismatches a tired editor misses.

Let the grade carry the emotion, not just the consistency

A technically matched timeline can still be emotionally flat — the grade should shift with the story, not just hold a palette. One production structured its treatment around five escalating emotional stages, each with locked rules for camera and lighting, so the look darkened on schedule as the film escalated. The maker-checker pass proved its value here: the invideo agent flagged that the film's entity reveal was running at the wrong emotional stage register — Stage D instead of Stage C — a color-and-tone nuance the director had missed in the cut. Treat the color script as a per-shot parameter that answers "what should the audience feel here," and let the post grade execute that answer.

Fixing the plasticky look: the realism pipeline

The single most common complaint about AI footage — over-sharp, plasticky skin — has a documented sequential fix. Seedance 2.0 generations in particular come back with an ultra-sharp skin quality that must be corrected in post to achieve realism. The pipeline, in order:

1. Topaz Astra upscale on invideo — the first step, before any color work.
2. Color grade — the full pass described above.
3. A small amount of blur — just enough to break the synthetic edge sharpness; AI footage is sharper than any real lens.
4. Film grain on top — grain reintroduces the organic texture the model rendered away.

The documented conclusion: a small amount of blur, grain, and color grading on top of AI-generated footage moves it meaningfully closer to live-action film. Two efficiency moves around the pipeline: reduce the correction burden at generation by casting with Recraft, which renders facial portraits with pores, lines, and stubble that read photoreal before any post work; and batch the upscaling by spinning up a sub-agent inside invideo named "Upscale Artist" — naming the sub-agent for the job is enough to run automated upscaling passes without manual intervention.

LUTs and film grain for AI footage

LUTs built for camera footage assume log or neutral input — AI footage is neither, so a LUT at 100% strength clips highlights and crushes the already-contrasty image. Apply LUTs at 50–70% intensity and treat the LUT as a starting point you correct under, not a finished look. Do the primary balance pass first, then drop the LUT on top at reduced opacity, then trim exposure and white balance per clip underneath it.

Grain does two jobs on AI footage. It masks the residual over-sharpness that blur alone doesn't fully kill, and it unifies clips from different generations — and different models — under one texture. That second job matters more than it would on camera footage: in one documented animated episode, 17 of the final shots were stitched from two or more separate generations, and a uniform grain layer is what makes a composite read as one continuous shot. Apply grain as the final layer in your NLE — DaVinci Resolve or Premiere Pro in the documented workflows — after the LUT and blur, sized fine enough that it reads as film texture at your delivery resolution rather than noise.

Which models and tools — and where invideo fits

Model choice shapes how much grading work you inherit. Seedance 2.0 delivers strong cinematic 15-second clips and carries character and location context well, but its skin rendering needs the realism pipeline above. Kling generates multi-shot sequences natively, which keeps color consistent within a sequence because one generation covers multiple cuts. Veo holds lighting logic well across a clip. On the image side, Recraft is the casting choice for photoreal skin texture, Nano Banana handles character sheets, and GPT-Image-2 covers general frame and reference generation.

The practical point: you don't pick a platform per model. All of these models run inside invideo, and the invideo agent acts as the routing layer — it holds your tonal modes, lighting ratios, and negative prompts in context and applies them whichever model a given shot routes to. Run it in Always Ask mode and you approve every prompt, with the style block attached, before any credits are spent. Topaz Astra runs on invideo for the upscale pass, and final assembly and the LUT-and-grain finish happen in your NLE.

What a graded AI film actually costs

The full pipeline — generation-time look control, upscaling, and the post grade — is included in the documented production budgets, which work out to $315–$750 per finished minute, with the grading and upscaling passes a small fraction of the credit spend, since most credits go to video generation, not post.

FAQ

Can you color grade AI-generated video like normal footage?

Partially. AI footage has no RAW or log version, so heavy secondary correction breaks down faster than on camera footage. The working approach is to control palette and lighting at generation — named tonal modes with hex values, numeric lighting ratios, explicit light sources — and use the post grade for matching, refinement, and texture rather than rescue.

Why does AI video look plasticky, and how do you fix it?

Current models, Seedance 2.0 especially, render skin ultra-sharp and synthetically smooth. The documented fix is sequential: upscale with Topaz Astra on invideo, color grade, add a small amount of blur to kill the synthetic edge, then add film grain on top. Casting with Recraft, which generates pores, lines, and stubble, reduces how much correction the footage needs.

What strength should a LUT be on AI footage?

50–70% intensity. LUTs are built for log or neutral camera input, and AI footage arrives with contrast and saturation already baked in — full-strength application clips highlights and crushes shadows. Balance the clips first, apply the LUT at reduced opacity, then trim per clip underneath it.

Do you upscale before or after color grading?

Upscale first. Grading low-resolution AI footage and then upscaling amplifies compression artifacts and banding; running Topaz Astra before color work is the documented first step of the realism pipeline, giving the grade clean pixels to work with.

What is the 85:15 lighting ratio in AI filmmaking?

It's a numeric encoding of a dark-dominant lighting grammar — 85% of the frame in shadow, 15% lit — extracted from James Wan's films in one documented production and used as literal prompt language. Converting lighting intent into ratios and named sources gives video models something concrete to execute, where adjectives like "moody" drift.

Sources

All production figures, costs, and workflow details in this guide come from invideo's documented first-party productions — including a 70-second short graded in a Wong Kar-wai-derived palette system, a ~90-second horror short built on a James Wan lighting grammar, and a 3-minute hand-painted animated episode — with numbers quoted as recorded per production.

Watch these to see the techniques in action: