@MikuBTC the middle frame is the interesting part. from the backend side, every model wants frames in a different shape, so start/middle/end is a fun new one. it already showed up in ella's model list today
the content filter vs the uncensored guy
we blocked the fight out in grey boxes first, then seedance 2.5 turned it into this. it kept the villain's grey cube head, so that "sensitive content" card is covering a literal cube lol
footage made in ella, edit made with claude
@codyschneider we ended up letting people tag their own brand assets right in the prompt with @, so the model gets the actual image, not a description of it. a checker loop on top of that sounds like the missing half. how many rounds does it take before the vision model stops flagging stuff?
@devinteerfilms agree, and from the engineering side the render got fast but the rerolls didn't. picking the right model and redoing the shot that ignored half the prompt is still where the time goes
@CalixAVelarde 1/10th the price is hard to argue with lol. do you mix them, minimax for most shots and seedance for the ones that have to look perfect?
@SinroxP same, the waiting is the worst part. that's why in ella the job keeps running when you leave the tab, so you can edit the last clip while the next one cooks. what are you using for the ugc clips right now?
friday 4:59pm. "just one small deploy"
one prompt in ella (seedance 2.5). claude wrote the prompt and cut the edit, i just supplied the trauma
worst thing you ever shipped on a friday?