Reference-to-video routes a character anchor into the video provider so the face survives the motion. Plus: which video model fits which kind of clip.
Video is the format that breaks identity for most AI tools. A beautiful still and then a moving clip that looks like a different person. SocialAF’s video router solves that by sending the character’s reference images straight into the video model’s reference-to-video endpoint when one exists, and using the closest equivalent fallback when it doesn’t.
Every clip in the “In motion” section on the landing page is routed this way. The examples below show the actual prompt + model used for each.
Step 01
Same Generate page, top-right tab from Image to Video. Pick the character you already built. The reference array is already attached; the studio detects char-refs are present and starts routing through the reference-to-video pipeline when you submit.
Step 02
Video prompts are different from image prompts. Add motion vocabulary: how the camera moves, how the subject moves, what beats happen across the clip. Don’t over-script . Five seconds is two beats max.
Notice the example below: one subject move (rowing), one camera move (dolly-in), one light cue (warm window). That’s the right density for a 5s clip.
Prompt
the same woman performing a single dumbbell row, slow camera dolly-in, gym mirror catching warm window light, sweat detail, cinematic 35mm
Step 03
The catalog has six families with different personalities. Quick guide:
Prompt
the same man jumping from a coastal cliff into turquoise ocean, slow motion drop, splash on entry, photoreal adventure sports
Step 04
3-5 seconds for vertical social (TikTok, Reels). 8-10 for editorial b-roll. 15 for a longer hero shot. Per-second models charge proportionally; per-generation models like Grok ignore the duration field and run at their fixed length.
If you’re drafting, render at 5s. Once the framing works, regenerate at 10-15s for the keeper.
Prompt
the same woman walking through a sunlit narrow stone alley in a cream blazer and black trousers, slow dolly-in, photoreal fashion editorial
Step 05
When a character has refs AND you supply a start image, the studio prefers the reference-to-video route (refs are a stronger identity signal than a single frame). When you don’t supply refs, the start frame becomes the first rendered frame via the model’s image-to-video slug.
Use the start-frame option when you have a still you love and want to animate it specifically. Otherwise let the refs drive.
Prompt
the same man performing a high leap on a black stage with a magenta spotlight, slow-motion, dust particles in beam, photoreal stage cinematography
Step 06
Video generation is slower than image. Plan on 60-180 seconds for most providers, longer for Veo or Kling Pro at 1080p. The studio polls until the clip lands and uploads to your gallery.
Same diagnosis loop as image: face drift means swap to a provider with a dedicated r2v sibling (Kling, Wan, Seedance, Happy Horse). Motion drift means tighten the motion vocabulary in the prompt or change the model.
Prompt
the same man pulling a pizza from a wood-fired oven with a long peel, flames visible behind, slight slow motion on the pull, photoreal restaurant cinematography
Video carries the face. Now the next layer: making the clip actually speak in your character’s voice.
The whole pipeline is in your dashboard. Start with a character, ship every format from there.