Generated video fails in recognisable ways, and each failure has a cause. Knowing which artifact you are looking at tells you whether to change the prompt, change the model, or change the shot, and the third answer is more often correct than people expect.
TL;DR
- Morphing: objects transform into other objects. Usually too much motion or too long a clip
- Flicker: brightness or texture pulsing between frames. Temporal instability, worse on cheap tiers
- Warping: straight lines bending as the camera moves. The model has no 3D model of the scene
- Ghosting: faint duplicates trailing motion. Fast movement beyond the tier
- The plastic look: too smooth, too even. Often the model's default, fixable in the prompt
- Most are shot-selection problems, not prompt problems
Morphing
What it looks like: an object gradually becomes a different object. A hand becomes part of a cup, a chair leg merges into a table, a face shifts identity mid-shot.
Cause: the model generates a coherent sequence without a persistent model of what objects are. Over enough frames, its idea of what something is drifts.
Fixable by: shorter clips (morphing scales with duration), less motion, a higher tier, and fewer objects in frame. Seedance 2.5 at 157 credits holds object identity better than cheaper tiers.
Not fixable when: the shot has many interacting objects or a long duration. That is a shot-selection problem.
Flicker
What it looks like: brightness, colour or texture pulsing frame to frame. Most visible on flat surfaces and in shadows.
Cause: temporal instability. Each frame is generated with some independence and small variations become visible as pulsing when played.
Fixable by: a higher tier, simpler lighting, and avoiding large flat areas of uniform colour where variation is most visible. Also: some flicker disappears under a grade and a grain overlay in the edit.
Warping
What it looks like: straight lines bending, architecture flexing, floors undulating as the camera moves.
Cause: the model has no 3D representation of the scene. It generates plausible next frames rather than moving a camera through a space, so parallax is approximated rather than computed.
Fixable by: less camera movement, or a locked-off camera. Straight lines with a static camera hold fine.
Not fixable when: the shot requires the camera to move through architecture. Generated video is weakest exactly there, and this is a shot to cut rather than to retry.
Ghosting and trailing
What it looks like: faint duplicates behind a moving subject, or a smear that is not motion blur.
Cause: fast motion beyond what the tier resolves.
Fixable by: slowing the motion in the prompt, or moving up a tier. Ghosting is one of the clearest signals that you are asking a cheap model for something it cannot do.
The plastic look
What it looks like: skin without pores, surfaces without texture, lighting that is perfectly even, everything slightly too clean.
Cause: partly the model's default aesthetic, partly the absence of the imperfection real footage has.
Fixable by: asking for imperfection explicitly. "Slight handheld movement, minor focus breathing, visible film grain, uneven practical lighting." Real footage has lens imperfection and this is the single most effective prompt addition for realism.
Also fixable in the edit: a grain overlay and a grade do more than most people expect.
Text garbling
What it looks like: signs, labels and titles rendering as plausible-looking nonsense, changing between frames.
Cause: models learn the visual texture of writing rather than that a specific glyph sequence is the only correct one.
Fixable by: not generating text. Add type in the edit. This is not a limitation to work around, it is a workflow rule.
Hands
What it looks like: extra fingers, merged fingers, impossible joints, and hands that change count between frames.
Cause: extreme articulation, constant self-occlusion, and no strong prior on finger count.
Fixable by: composition. Frame them out, give them a simple object to hold, keep them still, or keep them distant. See which AI model handles hands best.
The triage that saves credits
Before retrying, ask which category you are in:
Prompt problem: the plastic look, some morphing, motion that is too fast. Change the prompt, 24 credits to test on Wan 3.0 at 480p.
Tier problem: flicker, ghosting, adherence. Move up. Kling v3 Standard at 57 silent, Seedance 2.5 at 157 for physical detail.
Shot problem: warping under camera movement through architecture, hands doing something specific, text in frame, two people interacting, crowds. Change the shot. Retrying these is how people spend a month's plan learning what the failure modes are.
And since there is no video upscaler on the canvas, none of these are fixable after generation. The decision happens before you submit.
FAQ
Why does my AI video morph halfway through? Duration and motion. Shorter clips morph less, and a higher tier holds object identity longer.
Can I fix flicker in post? Partly. A grade and a grain overlay hide a lot of it.
Why do straight lines bend when the camera moves? The model has no 3D scene. Use a static camera, or cut the shot.
Which artifacts mean I should change the shot? Warping through architecture, hands doing something specific, text, crowds, and two people interacting.
Diagnose the artifact, then decide: prompt, tier, or shot. The 8frame canvas is free and unlimited, and generation is paid from $19/month.