AI video models struggle with POV shots because their training data is dominated by eye-level, third-person framing, they have no true 3D spatial reasoning, and text prompts give them no reliable way to lock first-person camera position, head motion, or hand placement. POV requires camera intent the model was never taught to infer from words alone.
The breakdown sits in four specific places. First, training-data bias: most film footage models learn from is shot from an observer position, so when you prompt 'POV' the model regresses to a low-angle or close third-person shot that approximates first-person without committing to it. Second, no camera-parameter understanding: video models don't represent lens height, head pitch, or rig position as separate variables — they treat 'POV' as a style word, not a camera coordinate, so framing drifts every generation. Third, no persistent 3D scene: the model can't reason about where the viewer's body, arms, or feet should enter frame, which is why POV outputs often show floating hands, missing legs, or a phantom torso. Fourth, motion coherence: a real POV shot carries head sway, footstep cadence, and breath rhythm — motion priors the model has very little of in its dataset.
The practical consequence on a production is iteration cost. In one documented short film, the team flagged POV shots as one of the hardest categories they hit — "multiple iterations, multiple different prompting techniques" before a usable take landed. Across documented productions, the baseline is roughly 3 generations per usable shot and a ~25% selection rate from raw generations to final cut; POV shots routinely run worse than that baseline.
What actually unblocks them, in order of reliability:
Give the model a real POV reference video. Act the shot out on your phone — walk the path, hold the object, move your head the way the character would — and upload that clip as a reference. The model now has explicit camera height, motion cadence, and hand position to anchor to, instead of guessing from the word 'POV'. invideo is an agentic video creation tool with all current generation models available, and the invideo agent routes that phone-shot reference into a reference-to-video model (Seedance 2.0 reference-to-video is the strongest current option because it accepts character and location references alongside the motion reference) so character identity and world don't drift while the camera is locked.
Generate a POV still first, then animate. Lock a single first-person frame with an image model (Nano Banana or GPT-Image-2) — hands in frame, correct lens height, correct object placement — and feed that as the start frame for video. You're solving the spatial composition problem in image space, where models are much stronger, then asking the video model only to add motion.
Route the model choice deliberately. Kling and Veo handle controlled camera motion differently than Seedance 2.0; for a slow, locked POV with subtle head movement, a still-to-video pass on Kling often holds framing better, while Seedance 2.0 reference-to-video wins when you have the phone-mock reference and need motion fidelity. The invideo agent picks the model per shot rather than forcing one model across the whole sequence.
Specify the camera literally, not stylistically. Prompt in coordinates the model can act on — "camera at 165cm height, slight downward pitch, hands visible in lower third of frame, walking pace head bob" — instead of "POV shot". The word 'POV' alone is too ambiguous for current models.
These are the techniques that consistently move POV shots from failed generations to usable footage — what works on any specific shot depends on the action, the framing, and how much of the actor's body needs to be in frame.
Watch some of these to see what works for you:
A good POV shot is really hard. Multiple iterations, multiple different prompting techniques and that's when we experienced a magic moment with agent one.
— invideo's creative team, from a documented AI short film production