How to Make the AI Hug Video (and Why Most Attempts Fall Apart)
The AI hug trend puts two people in one embrace who were never photographed together. Here is the two-step pipeline that actually holds identity, what each step costs, and the failure mode nobody warns you about.
The AI hug trend is short clips of two people embracing who were never in the same photograph — a couple apart, friends in different cities, someone and a relative who has died. It carries emotionally, which is why it keeps coming back around.
Most tutorials point you at a one-click generator. Those work when both faces are frontal and well lit, and produce melted features when they are not. The reason is structural, and knowing it tells you how to fix the result.
Why one-click hug generators break
A hug is the hardest possible case for a video model. Two identities have to stay stable, and they have to occlude each other — arms cross faces, one head turns into the other's shoulder. Every frame where a face is partly hidden is a frame the model has to guess at, and identity errors compound from there.
Asking a video model to invent the embrace from two separate photos means it is solving composition and animation at once. Split those into two steps and both get easier.
The two-step pipeline
Step 1 — compose the still. Use an image-editing model to put both people in one frame, in the pose you want, before any motion exists. On fal, fal-ai/nano-banana-2/edit costs $0.08 per image and fal-ai/nano-banana-pro/edit costs $0.15 (read 2 August 2026). Iterate here, not in video — you are paying cents per attempt instead of dollars.
Ask for a specific arrangement: who is on the left, whose arm goes over whose shoulder, both faces visible to camera. Vague prompts give you vague anatomy.
Step 2 — animate the still you approved. Feed that composed frame to an image-to-video model. alibaba/happy-horse/v1.1/image-to-video is $0.14 per second; fal-ai/kling-video/v3/pro/image-to-video is also $0.14. A five-second clip is $0.70.
Total for a finished hug clip: well under a dollar, most of which is the video step you only run once.
The prompt that matters
For step 2, restraint wins. "The two people hold the embrace, slight natural sway, both faces remain visible to camera" produces a stable clip. Asking for a full hug motion — approach, embrace, squeeze — asks the model to move two faces through heavy occlusion, and that is exactly where identity slips.
Five seconds. Small movement. If you want a longer piece, generate two short clips and cut between them.
The "hug my younger self" variant
The same pipeline, with a childhood photo as one of the two subjects. It is currently the most-shared version of the trend. One extra note: old photos are usually low resolution and differently lit, so the composition step needs more iterations — which is precisely why you want that step to be the eight-cent one.
Before you post
Two people in a frame who were never together is exactly the kind of content platforms now expect to be labelled. Meta applies AI labels on Instagram and Facebook, and disclosure is an enforced requirement for sponsored content. If the clip involves someone who has died, the ethical question is worth more thought than the technical one — many people find it moving, and some find it a violation. Ask the family, not the model.