Troubleshooting Rumpelstiltskin AI prompts for role and motion errors
Last updated:
When a Rumpelstiltskin AI prompt fails, first check that the video and photos were attached as references and mapped to the intended roles. Then inspect whether the failure concerns identity, motion or audio. Change one relevant instruction at a time; a longer prompt cannot repair a missing reference or guarantee a successful generation.
Verify references before rewriting the prompt
A model cannot follow an attachment that was not supplied through the appropriate input field. Start by checking that the source clip is a video reference, rather than a link pasted only into prose. Confirm that the intended identity photograph is attached as an image reference. Reference names and accepted input controls vary between interfaces. The Seedance 2.0 model page describes support for image, audio and video references, but that does not mean every hosting interface exposes identical fields or token names.
For two people, check the image order as carefully as the text. On our site, photo one belongs to the little dancer and photo two belongs to the seated woman. A reversed upload can create a role error that looks like a weak prompt. For a standalone tool, use its actual reference labels and verify their previews before submitting. Never guess that an unsupported parameter such as an identity-strength slider exists just because a prompt template mentions it. Keep a record of the inputs and settings used for each test.
Describe the target by actions and costume
An instruction to replace “the man” is less useful when several shots, partial silhouettes or reactions make the role ambiguous. Describe the unique performer who approaches the seated woman, wears a dark tailcoat with a white shirt and black bow tie, raises a finger to his lips and dances on tiptoe. Map that performer to the uploaded identity. For a second replacement, separately identify the woman seated in straw who covers her mouth in surprise. Those descriptions remain meaningful even when the camera changes position.
Separate identity from the details that should stay fixed. Ask for the uploaded face, hairstyle and visible skin identity while keeping the reference costumes, character scale, setting and performance. If a full-body photograph is intended only as identity guidance, say so; otherwise the prompt can accidentally request its casual clothes in the barn. In a solo edit, explicitly retain the seated woman rather than adding a second replacement instruction. Do not combine conflicting requests to preserve the original outfits and copy the photographs' outfits.
Diagnose motion and audio as separate problems
If the face changes but the dance becomes unrelated, inspect the motion reference and the preservation instructions. Ask for the visible action order, camera movement, framing and cuts to follow the supplied clip. A source containing several shots should not simultaneously be described as one uninterrupted shot. A short invented dance description can also compete with the reference choreography. Use the reference as the authority for timing rather than requesting a different routine and hoping the model will combine both.
Sound needs its own check. A request for matching audio can lead to newly generated music rather than the exact original recording. Our delivery process restores the reference soundtrack after generation, which is why an identical soundtrack should be evaluated separately from the image model's audio capabilities. In another tool, check what its audio option actually promises before paying. Do not stretch a shorter generated picture simply to fit a longer soundtrack; doing so changes the visible timing you were trying to preserve.
Change one variable and keep the comparison
Write down the failure before submitting another render: unchanged dancer, swapped roles, face drift during a turn, incorrect footwork or missing sound. Keep the same inputs while testing a targeted correction so you can tell whether that correction helped. Compare complete clips and close-ups, not only the opening frame. A single favorable run is evidence for those inputs, not a guarantee for every face. Set a spending limit and stop testing when you cannot explain what the next attempt will change.
For a hands-on starting point, the existing copyable-prompts page supplies templates; this guide explains how to diagnose their limitations. You do not need to edit prompts to use our generator. Upload one photo for Just me or two for Two of us, in the labeled order, and select a one-time pack. Watching examples and reading templates is free; personalized generation starts at $6.99. A confirmed technical failure restores credits, while a completed video that looks wrong should be reported with the video and a specific reason so support can review it.
For a starting template, use our copyable Rumpelstiltskin AI prompts.