I Spent 159 Credits Trying to Clone Myself on Video — Here's What Broke
Six generations, two models, no usable clip — and four findings that transfer to every video model.
What this covers
Building a talking-head video clone: reference photos, a voice clone, and two different video models. Six generations across an evening.
No usable finished clip. That’s the honest result, and the findings are worth more than a working demo would have been.
What broke and why
The newer model was worse. A controlled A/B — same prompt, same reference, same audio, same length — put Seedance 2.5 against 2.0. 2.5 produced a cleaner image and a worse likeness, and cost 43% more. Newest is not automatically best.
A reference outranks your text. A shirt logo kept rendering despite the prompt saying “no logo” three separate times. It wasn’t invented — it came from the reference photo. You cannot write away what your reference shows.
But you can transform what you can’t negate. Asking the model to simplify branded elements into a universal version removed the logo instantly, where three negations had failed. Describe what you want it to become, never what you want removed.
Camera-lock language freezes the subject. After writing “static,” “locked” and “fixed” five times in one prompt, the head never moved for the entire clip. The model reads camera-lock as subject-lock. State the lock once, then say the opposite loudly: the subject is not locked.
Gaps get invented, not left blank. A tattoo appeared on the back of an arm that no reference photo covered. Plausible-but-wrong is more dangerous than obviously-wrong — it passes inspection.
And the voice was never the video model’s job. Several generations were spent trying to make a video model preserve a cloned voice. It doesn’t, and it isn’t supposed to. The voice gets lip-synced on afterwards as a separate step.
What it cost
159 credits — about $7.82 at the plan rate, in one evening, for zero usable clips.
Two subscriptions were bought to get there: $59/month for the video platform and $6/month for voice cloning.
The number that matters: the log said 102 credits when the account said 159. Cancelled jobs still bill, and a test run you forget about is real money. Track the balance, not the sum of what you remember generating.