Getting past the uncanny valley
Realistic humans come from photography language, not adjectives. Specify the lens and light — '35mm, shallow depth of field, soft overcast light' — and add the imperfection cues that break synthetic smoothness: 'natural skin texture', 'candid', 'unposed'. In motion, small actions read truest: a subject turning, breathing, shifting weight, adjusting a sleeve. The stillness between movements is what makes a generated person feel present.
Know where the technology is weakest and stage around it: complex hand choreography, fast overlapping actions, and crowds of interacting people are the failure modes. A clip of someone pouring coffee reads flawlessly; a clip of someone shuffling cards mid-conversation is asking for trouble. Frame hands loosely or keep them occupied with one simple task.
Consistent characters across shots
One-off clips are easy; a recurring person across a whole campaign is the real problem. Nidhogg solves it three ways. Character training builds a LoRA from 3–20 photos of a person, so the same face renders reliably across prompts. Kling 3's Elements feature accepts reference images inside a generation, and Seedance 2 takes reference images for its multi-shot sequences.
The practical workflow: train the character once, lock a wardrobe and lighting description you reuse verbatim, and vary only the action and setting per clip. Consistency comes from disciplined prompts as much as from the tooling.
Making a person speak
For talking footage, skip text-to-video roulette and use the dedicated tools. InfiniTalk turns a single portrait photo into a talking avatar synced to audio, and OmniHuman drives a fuller performance — gesture and expression — from an image plus a voice track. For voice, Nidhogg's TTS engines offer a library of preset voices; there is no voice cloning, so you either pick a preset voice or record your own audio and use it as the driving track.
If you already have footage and just need different words, LatentSync and Sync Lipsync redub an existing video to a new audio track, matching the mouth to the replacement voice.
Use it responsibly
Generating people carries obligations that generating landscapes doesn't. Only animate or train on photos of real people with their consent, and don't generate footage that presents a real person saying or doing things they didn't. For client and brand work, disclose that footage is AI-generated where the context implies otherwise.
The safest creative territory is also the most flexible: invented people. A generated spokesperson has no scheduling conflicts, no likeness-rights negotiations, and can be regenerated in next season's wardrobe with one prompt edit.