In June, Runway debuted a brand new text-to-video synthesis mannequin known as Gen-3 Alpha. It converts written descriptions known as “prompts” into HD video clips with out sound. We have since had an opportunity to make use of it and needed to share our outcomes. Our exams present that cautious prompting is not as vital as matching ideas probably discovered within the coaching information, and that reaching amusing outcomes probably requires many generations and selective cherry-picking.
A permanent theme of all generative AI fashions we have seen since 2022 is that they are often glorious at mixing ideas present in coaching information however are usually very poor at generalizing (making use of realized “data” to new conditions the mannequin has not explicitly been educated on). Which means they’ll excel at stylistic and thematic novelty however wrestle at basic structural novelty that goes past the coaching information.
What does all that imply? Within the case of Runway Gen-3, lack of generalization means you may ask for a crusing ship in a swirling cup of espresso, and offered that Gen-3’s coaching information consists of video examples of crusing ships and swirling espresso, that is an “simple” novel mixture for the mannequin to make pretty convincingly. However if you happen to ask for a cat ingesting a can of beer (in a beer industrial), it is going to typically fail as a result of there aren’t probably many movies of photorealistic cats ingesting human drinks within the coaching information. As an alternative, the mannequin will pull from what it has realized about movies of cats and movies of beer commercials and mix them. The result’s a cat with human fingers pounding again a brewsky.
A couple of fundamental prompts
Through the Gen-3 Alpha testing part, we signed up for Runway’s Commonplace plan, which offers 625 credit for $15 a month, plus some bonus free trial credit. Every era prices 10 credit per one second of video, and we created 10-second movies for 100 credit a chunk. So the amount of generations we may make had been restricted.
We first tried a couple of requirements from our picture synthesis exams prior to now, like cats ingesting beer, barbarians with CRT TV units, and queens of the universe. We additionally dipped into Ars Technica lore with the “moonshark,” our mascot. You may see all these outcomes and extra beneath.
We had so few credit that we could not afford to rerun them and cherry-pick, so what you see for every immediate is strictly the one era we obtained from Runway.
“A highly-intelligent individual studying “Ars Technica” on their pc when the display screen explodes”
“industrial for a brand new flaming cheeseburger from McDonald’s”
“The moonshark leaping out of a pc display screen and attacking an individual”
“A cat in a automobile ingesting a can of beer, beer industrial”
“Will Smith consuming spaghetti” triggered a filter, so we tried “a black man consuming spaghetti.” (Watch till the tip.)
“Robotic humanoid animals with vaudeville costumes roam the streets accumulating safety cash in tokens”
“A basketball participant in a haunted passenger prepare automobile with a basketball courtroom, and he’s enjoying in opposition to a group of ghosts”
“A herd of 1 million cats working on a hillside, aerial view”
“online game footage of a dynamic Nineties third-person 3D platform sport starring an anthropomorphic shark boy”
