How do you make AI-generated video look more realistic and less plastic?

AI video looks plastic because raw generations are over-sharp, over-saturated, and too clean on skin and motion. To fix it, run a post pass on every clip: upscale with Topaz Astra on invideo, then add a light blur, film grain, a desaturated grade, and a room-tone audio bed. Prompt for imperfections up front so there's less to fix later.

Start in the prompt, then clean up in post — both passes matter. The invideo agent routes generation across Seedance 2.0, Veo, Kling, and Runway and runs upscale and grade passes after, so the whole realism pipeline lives in one place.

1. Prompt for imperfections at generation time. The plasticky feel starts in the model — Seedance 2.0 in particular produces an ultra-sharp, waxy skin quality by default. Push against it in the prompt: handheld camera, soft natural light from one source, motion blur, shallow depth of field, skin pores and stubble visible, lens grain, slight haze. For close-ups, generate the portrait first in an image model that holds skin texture — Recraft V4 gives you pores, lines, and stubble that read as a real face — then drive the video off that reference rather than off text alone. As Hridaye, invideo's creative director, put it: "ReCraft actually gives you those imperfections like pores, lines, stubble, like all the little stuff that makes a face look like an actual face."

2. Upscale first, before any grading. Topaz Astra on invideo is the first step of the post pipeline — it cleans temporal artifacts and restores detail without exaggerating the AI sharpness the way generic upscalers do. Run every selected clip through it before you touch color. Inside invideo you can spin up a sub-agent, name it something like "upscale artist", and batch the pass so you're not doing it clip by clip.

3. Soften, grain, then grade. This is the single most repeated post move documented across invideo productions: "What we tend to do is put a tiny bit of blur on top of the scene, add a bunch of grain and then play with the grade till it comes closer to live action film." Concretely — a very light gaussian blur (just enough to take the digital edge off), a 35mm-style grain overlay at roughly 20–40% opacity, then a grade that pulls saturation down 10–15%, lifts the blacks slightly, and pushes a warm or teal cast depending on the scene. The grain breaks the smoothness on skin and gradients; the blur kills the over-sharpness; the desaturated grade pulls it out of the hyperreal AI palette.

4. Ground it with sound. Silent or music-only AI clips read as fake even when the image is solid. Layer a room tone or ambient bed under every shot — traffic hum, wind, refrigerator buzz, distant chatter — plus diegetic detail (footsteps, fabric, the specific sound a prop would make when it hits the floor). Sound carries half the realism; without it the eye keeps flagging the image as artificial.

5. Yield is a budget line, not a failure. Across documented productions, only about 25% of generated clips are editorially usable — 41 of 164 clips made the final cut on one 3-minute animated episode, with an average of just 5 seconds used per 15-second generation. Plan for roughly 3 generations per usable shot and pick the cleanest takes for the realism pass, rather than trying to rescue every clip.

Beyond the post pass itself: image-to-video off a photographic reference (not an illustration) gives you a more realistic baseline than text-to-video, and Seedance 2.0 reference-to-video carries that grounded look across continuation clips so you're not re-fighting the plastic look every shot.

Watch some of these to see what works for you:

What we tend to do is put a tiny bit of blur on top of the scene, add a bunch of grain and then play with the grade till it comes closer to live action film.

— Hridaye, invideo's creative director