📱 Want the full experience & exclusive pop-ups? Download the iOS App · Android coming soon
Video Studio
Generate short vertical clips for social. Upload a reference still to keep a character consistent between posts, or go text-only.
Admin access required. Sign in
Add a start frame to keep a character looking consistent between posts. Add an end frame as well and the model animates the journey between the two poses. Any size works — images are resized and cropped to your chosen aspect in the browser before upload.
Scene Builder — put two characters in one frame
The model takes one image per frame and has no way to be handed two separate characters, so naming a second character in the prompt just invents one. Instead: cut each character out, stage them here, and use the result as a frame. Build two scenes — the before and after of the action — and set them as start and end frames to define the movement instead of hoping for it.
Workflow for an action shot: add both characters, Auto-arrange, nudge them into the "before" pose, Harmonise, then Use as START frame. Save the scene, move the layers into the "after" pose, Harmonise again, then Use as END frame. Generate. Harmonise runs a light image-to-image pass so everything shares one lighting environment — keep strength low, above about 0.4 it starts rewriting faces.
Prompt builder — tap to append
Starter templates
Saved character & style snippets
Save a character description once, then prepend it to any prompt. Stored in this browser only.
How to get the best results
One action per clip. Five seconds holds a single beat. "Waves, then trips, then laughs" comes out as mush; "waves slowly" comes out clean.
With a start frame, describe motion only. The appearance already comes from the image. Re-describing the character makes the model fight the reference and the face drifts.
Without a start frame, describe appearance in detail — subject, materials, colours, setting, art style — because nothing is anchoring it.
Name the camera explicitly. "Camera holds still" or "slow push in" beats leaving it unsaid; unspecified camera work tends to drift.
Say "slow" more than feels natural. Fast motion is where warping and extra limbs appear.
Expect to generate three or four per keeper. Draft at 540p, then re-run the winner at 720p using the Re-run button.
A 5-second clip takes about 60–90 seconds; 9 seconds takes longer. The model renders one clip at a time, so extra requests are queued automatically — line up as many as you like. Keep this tab open and the queue keeps moving. Draft at 540p and only re-run keepers at 720p — it halves the bill.