1. Consistency

Consistency—also known as temporal consistency—remains one of the most difficult problems in AI video generation. AI must maintain the same character, objects, environment, lighting, and visual style across multiple frames and scenes.

Common issues include:

• Characters changing appearance between shots
• Clothing changing unexpectedly
• Hairstyles or facial features shifting
• Objects appearing or disappearing
• Background details changing between scenes

This is why image-to-video workflows are often preferred for narrative projects, as they provide stronger visual references for the AI to follow.

We cover image to video in all our AI video tools in the course and you can read more about these tools here.

Often called morphing, drift, or identity drift, this occurs when characters, objects, or environments slowly change throughout a clip.

Examples include:

• Faces transforming during a scene
• Hands changing shape
• Objects merging together
• Buildings changing structure
• Clothing evolving mid-shot

These errors typically become more noticeable in longer generations as small mistakes accumulate over time.

These have become less of an issue as the years have moved on with AI video and the key tools we use, Veo, Seedance, Kling and such like have managed to limited these - not always and completely, but mostly. 

Even the best models occasionally produce unusable results.

Common artifacts include:

• Distorted faces
• Extra fingers or limbs
• Floating objects
• Flickering details
• Broken anatomy
• Warped environments

This is completely normal. AI video generation is often an iterative process where multiple generations are created before achieving the desired result.

Many AI video models can now generate speech and lip movements, but synchronization can still break down, particularly during:

• Fast speech
• Emotional dialogue
• Close-up facial shots
• Longer conversations

You may notice:

• Mouth movements that don't match words
• Delayed lip movement
• Unnatural facial expressions
• Robotic-looking speech delivery

For important dialogue scenes, dedicated lip-sync tools or additional editing may still be required.

Text prompting for your speech - for example inside Veo that we show in the course, often produces the best lip-syncing results, as opposed to dedicated lip-syncing like in Runway. However the consistency of the voice may be a slight issue (which can be resolved exporting the audio and using voice changer in Elevenlabs for example and using that audio inside your edit). There are ways around this, although it's getting better every month. 

AI often struggles with complex human behaviour.

Problem areas include:

• Handshakes
• Hugging
• Dancing
• Sports movements
• Physical interactions between multiple people

While a single person walking can look highly realistic, multiple people interacting naturally remains challenging.

Largely an older issue that tools like Seedance and Google Omni seem to mave limited, even able to produce great fight sequences and difficult movement. 

Most AI video models perform best with short generations. As clips become longer, consistency and quality can begin to degrade.

Common issues include:

• Visual drift
• Character changes
• Motion instability
• Loss of detail
• Scene incoherence

This is why many creators generate shorter clips and edit them together during post-production rather than relying on a single long generation.

The length of clip is often limited to 8 - 15 seconds, although this can extend in most tools - like we show in Scenes in Flow - where consistency remains. However, for timing and flow, you may still want shorter shots and edit these in your chosen editor to get the timing perfect. 

AI often struggles to render:

• Signs
• Product labels
• Logos
• Documents
• On-screen text

Letters may become distorted, unreadable, or change between frames. If accurate text is required, it is usually added during editing rather than generated directly by the AI.

We find this still in a lot of tools, with the exception of Seedance which we often find gives the best results for background text. See our examples in the seedance lecture or when using Seedance in the Runway lecture comparing tools. 

Although AI can create visually stunning scenes, it doesn't actually understand physics.

This can result in:

• Unrealistic object movement
• Strange gravity effects
• Incorrect reflections
• Water behaving unnaturally
• Hair and clothing moving oddly

Most viewers won't notice minor issues, but complex scenes can reveal these limitations.

Less of an issue than it used to me, but you will no doubt come across these in your generations at some point. 

One of the realities of AI video creation is that results can vary dramatically, even when using the same prompt.

You may generate:

• One incredible clip
• Several average clips
• A few unusable clips

Learning how to prompt effectively, use references, and iterate on generations is often more important than the tool itself.

Also, some tools just produce better quality that others in certain styles, that's why we show you an array of tools and you can decide which is best for you and your needs. See all video tools here

While these limitations are important to understand, AI video quality continues to improve at an extraordinary rate. Problems that seemed impossible to solve just a few years ago—such as realistic motion, sound generation, and character references—are already becoming significantly better.

The most successful AI creators understand both the strengths and weaknesses of current tools. Rather than fighting the technology, they build workflows that work with it, using the right generation method, references, editing techniques, and creative decisions to achieve professional results.

Example: Creating your video project with voice over narration instead of lipsync dialogue.