How do you fix AI video scenes where two characters need to physically touch or interact?

Fix two-character contact shots at the reference level, not with longer prompts — give the model a visual anchor of the exact physical arrangement:

1. Hand-sketch the contact pose and upload it as a reference
2. Build a fused character sheet of both characters in contact
3. Overgenerate and stitch the best seconds (Frankenstein shot) In one documented production, 75% of the film featured a two-character carry shot solved this way.

Start by replacing the text description of the pose with a visual reference of it — contact arrangements (bodies in contact, ropes, props between characters) break AI video models faster than almost any other scenario, and the documented fixes all work by conditioning the model on an image of the arrangement rather than a sentence about it. invideo is an agentic video creation tool with all the current video and image models available, and the invideo agent routes your references to the right model per shot.

1. Hand-sketch the exact contact pose and upload it as a reference. Draw the arrangement you need — who holds whom, where limbs connect, relative scale, where any prop sits between the bodies. A rough sketch on paper, photographed on your phone, is enough; spatial relationships matter, draftsmanship doesn't. Upload the drawing to the invideo agent, which attaches it to an image model such as Nano Banana and prompts against it until it returns an accurate character sheet of the configuration. In one documented production, the image model could not produce a two-character carry configuration from text prompts at all; a hand sketch of how the smaller character attaches to the larger one got the correct sheet on the next pass, and 75% of that finished film features the resulting contact setup. As the team put it: "The lesson for the day truly is that when the models get stuck you draw, you shoot, you bring your hands in and you get it done. And that's when agent one meets you there and takes it over the line."

2. Build a fused character sheet on top of locked individual sheets. Lock each character's own reference first — multi-angle turnarounds with face and detail close-ups, generated with objects removed from their hands, since held props cause inconsistency across turnaround angles. Then generate a fused sheet showing both characters together in the contact pose, and attach the fused sheet plus both individual sheets to every video generation of the interaction. For scale: one production covered 4 characters and 1 prop with just 11 reference images total. If the touching character's appearance evolves across the sequence — costume changes, accumulated carried objects — make a separate fused sheet per beat so each segment generates against the current look.

3. Overgenerate the interaction and stitch the best seconds — the Frankenstein shot. Contact physics rarely hold for a full clip, so plan for it: budget an average of 3 generations per usable shot, then cut together the seconds from each take where the interaction reads correctly. In one 3-minute production, 17 of the final shots were stitched from 2 or more generations — more than 40% of the finished episode is composited this way. One adjacent lever: Seedance 2.0 reference-to-video accepts character references alongside each clip, so your fused sheet carries into every segment of the interaction — and since all current models run inside invideo, the invideo agent routes each contact shot to the model that handles it best.

These are some of the ways to problem-solve this — what works depends on your shot.

Watch some of these to see what works for you:

The lesson for the day truly is that when the models get stuck you draw, you shoot, you bring your hands in and you get it done. And that's when agent one meets you there and takes it over the line.

— invideo's creative team