Correction, 2026-10-02: this page stated as fact several things we never measured: that people rarely spot AI video from picture quality and that four of the five tells are fixable without a better model; that hands are "improving fast"; that sustained eye contact is "the strongest uncanny trigger in the format"; that asking for a worse camera operator produces a more believable shot; that models default to sourceless light and that committed lighting is "the single highest-leverage change available in a prompt"; that a light grade "moves it a long way" and grain is "the cheapest correction there is"; that one real shot raises the perceived reality of the shots around it; that a different model "changes how often it fails, not whether"; that models "cluster closer than marketing suggests"; that more resolution shows more of the artifacts; and that social and advertising audiences increasingly do not care. Those are now written as likely causes, things to check and things to try on your own shot, and the FAQ no longer points to a model ranking for realism, because we have not measured one. The page recommended a grade and grain without saying what 8frame's own editor can do: the Video Editor mode has filter presets and a brightness slider per clip, and no curves, colour wheels or LUTs. We also added a frame-rate check (every video model we measured on 8frame returned 24 fps), the output resolution per model and the fact that there is no video upscaler on 8frame, a note that the Montage node's merged file has no sound, test costs from our price registry, and the Sources section the page was missing.
Generated video tends to give itself away in a handful of recognisable ways. This page covers five: physics, hands, eye contact, motion that is too smooth, and lighting that never commits. For each it gives what it looks like, the likely cause, what to check and what to try on your own shot. It does not rank models for realism, because we have not measured which one looks most real. For morphing, flicker, warping and the other frame-level faults, see AI video artifacts explained.
TL;DR
- Physics: pouring, falling, contact. If the shot is about a physical interaction, consider filming that shot and generating around it
- Hands: frame them out, keep them still or small, and fix them in the start image before you animate
- Eye contact: prompt for the glance away and the blink, or frame in profile
- Motion that is too smooth: describe handheld drift in the prompt, and check the downloaded file at its own 24 fps before you blame the model
- Lighting: name the source and the direction
- In the edit: 8frame's Video Editor has filter presets and a brightness slider per clip, but no curves, colour wheels or LUTs, so a full grade or a grain layer belongs in your own editor
- Resolution sets sharpness, not realism. Set it before you generate: there is no video upscaler on 8frame
1. Physics that almost works
What it looks like: liquids that pour slightly wrong, fabric that settles too fast, objects that touch without pushing each other. A viewer may not be able to say what is wrong and still feel that something is off.
Likely cause: the model generates frames that follow plausibly from one another; it does not appear to simulate weight, force or contact. We cannot see inside the models: treat this, like every cause on this page, as the explanation that fits the failure, not a documented mechanism.
What to try: keep the physical interaction out of the generated shot. Generate the room, the mood, the product sitting still, and film the pouring, the clicking, the sealing. The same split applies to product demos.
What to check: step through the moment of contact frame by frame. If it still fails there after a prompt change and a second model, film that shot.
2. Hands and small articulated things
What it looks like: extra or merged fingers, joints that bend the wrong way, a grip that changes between frames. Look hardest when the hands do something specific, such as typing, tying a knot or pouring.
Likely cause: hands have many joints and constantly hide parts of themselves, and in video the hand has to be right in every frame, not just one.
What to try: frame them out, give them a simple object to hold, keep them still, or keep them small in frame. In image-to-video, fix the hand in the start image before you animate it. If the hands are the point of the shot, film them. See which AI model handles hands best.
3. Eye contact that never breaks
What it looks like: a generated person holds the lens with a steady, slightly glassy stare. Real people blink at irregular intervals, glance away and refocus.
What to try: prompt for the glance away, the blink, the look down and back. Or frame in profile or three-quarter, so the eyes are not the centre of the shot. Run it against your plain prompt on the same model and keep whichever reads better on your shot.
What to check: watch the eyes at normal speed, then step through. If the stare survives a prompt change, recompose rather than retry.
4. Motion that is too smooth
What it looks like: camera moves that glide without weight. A real operator overshoots slightly at the end of a pan and settles after a push.
What to check first: whether the motion looks wrong in the file you downloaded or only after export. Every video model we measured on 8frame returned files at 24 fps. If your timeline or export runs at 25 or 30 fps, the editor has to repeat, blend or interpolate frames to fill the gap, and that can add stutter or doubling that makes the motion look wrong. Watch the file at its own 24 fps first.
What to try: describe the imperfection: "handheld, slight drift, a small overshoot at the end of the pan." Compare it with your plain prompt on the same model and keep whichever reads better on your shot. UGC-style ads run into the same thing, where a polished camera can make the ad read as a commercial: see how to storyboard a UGC ad.
5. Lighting that never commits
What it looks like: pleasant, even light with no visible source and no shadow that means anything. Real light comes from somewhere and leaves shadows.
Likely cause: if the prompt does not say where the light comes from, the model is free to choose, and you can end up with light from nowhere.
What to try: name the source and the direction: "single window, camera left, hard afternoon light, deep shadow on the right side of the face." Compare it on the same model with a version that leans on quality words ("cinematic, high quality") and keep whichever reads more real on your shot. Drafts are cheap on Wan 3.0 at 480p: 24 credits for 5 seconds.
The finishing pass
These work on clips you already have, without regenerating. We have not measured how much any of them changes how real a clip reads, so judge each one on your clip.
Grade it. 8frame's Video Editor mode has filter presets (Cinematic, Film Noir, Kodak and Super 8 among them) and a brightness slider per clip. It has no curves, colour wheels or LUTs, so a full grade belongs in your own editor.
Add grain, or ask for it. A grain overlay also belongs in your own editor; add it across the whole timeline rather than only to the generated clips. You can also try "visible film grain" in the prompt and compare it with your plain prompt on the same model.
Cut it short. Step through the clip and keep its cleanest stretch: in a short cut between other shots, a flaw has less time on screen. Cut on the timeline of the Video Editor mode, which has video, text, caption and sound layers, or in your own editor. The Montage node on the canvas joins the picture only: its merged file has no sound, so add dialogue, music or sound effects on the Video Editor timeline.
Mix real and generated. If one shot has to hold up (a physical action, hands doing precise work), film it and generate the shots around it. Then grade the whole sequence together so the generated clips do not stand apart from the filmed ones; see how to match AI clips to real footage.
Resolution is not realism
Resolution sets how sharp a clip is at full screen. It does not change where the light comes from, how the camera moves or where the eyes look. Set it before you generate, because there is no video upscaler on 8frame: the upscale tool works on images only.
On 8frame, Seedance 2.5 and 2.0 offer 480p or 720p; Kling v3 Standard returns 720p; Veo 3.1 and Veo 3.1 Fast are 1080p and Veo 3.1 Lite is 720p; Wan 3.0 offers 480p, 720p or 1080p; Hailuo H3 outputs 2K. No video model on 8frame outputs 4K. More in why is my AI video blurry.
When to change the shot instead
A clip that completes is billed even if it looks fake; only failed generations are refunded automatically (see do failed AI generations cost credits). So change one thing per run: the prompt first, then the model. To compare models, keep the start image and prompt fixed and change only the model: at 480p, Wan 3.0 is 24 credits for 5 seconds and Seedance 2.5 is 70. A higher price is not a measured guarantee of a more convincing shot, so judge on yours.
If a shot depends on physical cause and effect, hands doing precise work, or a recognisable person staying the same through a long take, and it still fails after a prompt change and a second model, change the shot: film it or cut around it, and generate everything else.
FAQ
Which model looks the most realistic? We have not measured it, so we do not rank models for realism. Run the same prompt or start image on two models and compare the five tells on your shot: at 480p, Wan 3.0 is 24 credits for 5 seconds and Seedance 2.5 is 70.
Does higher resolution make it look more real? We have not measured that. Resolution changes sharpness; it does not change the lighting, the motion or the eye contact, which are the tells on this page. Pick it before you generate: there is no video upscaler on 8frame, and no video model on 8frame outputs 4K.
Can I grade AI video on 8frame? Partly. The Video Editor mode has filter presets and a brightness slider per clip, but no curves, colour wheels or LUTs. A full grade and a grain overlay belong in your own editor.
Why does the motion look choppy or doubled after export? Check the frame rate. Every video model we measured on 8frame returned files at 24 fps. If your timeline or export runs at 25 or 30 fps, the editor fills the gap by repeating, blending or interpolating frames. Watch the downloaded file at its own 24 fps first.
Will viewers care if they can tell? It depends on the format, and we have not measured audience reactions. For documentary, testimonial and anything presented as evidence, generated footage changes what you are claiming: check your client's and the platform's rules on labelling it before you publish.
Sources
- 8frame prices: 8frame tool registry credit functions calculateWan3Cost (Wan 3.0 at 480p, 24 credits for 5 s) and calculateSeedance2Cost (Seedance 2.5 at 480p, 70 for 5 s), read 2026-10-02
- Output resolutions per model; no video model on 8frame outputs 4K: 8frame video generation code and front tool registry, via the 8frame content fact sheet, read 2026-10-02
- No video upscaler; the upscale tool is image-only: 8frame front tool registry, read 2026-10-02
- Video Editor mode (filter presets, brightness slider per clip, video, text, caption and sound layers; no curves, colour wheels or LUTs): 8frame front end, via the 8frame content fact sheet, read 2026-10-02
- The Montage node's merged file has no sound: 8frame back-end render path, via the 8frame content fact sheet, read 2026-10-02
- 24 fps: generated video files on 8frame over the 120 days to 2026-09-30, measured per model (every measured model returned 24 fps)
- Completed clips are billed, failed generations refunded automatically: 8frame billing, see do failed AI generations cost credits
- Causes of each tell: our reading of the failures, not a measurement or a vendor statement
Commit the lighting, break the eye contact, and film the physics.
Try it without signing up: the 8frame canvas opens on a real board. The canvas is free and unlimited; generation is paid from $19/month.