MAR 2025
This feels like a major inflection point for AI video generation.
krea.ai has quietly released an accessible, TRAINABLE, video model from Wan 2.1.
Product fidelity through video generation has always been a problem: You upload a pixel perfect image of a specific make/model/color of a car and watch as the AI video generator hallucinates what the car should look like in motion. And as a camera orbits around it, the original car melts away into some generic vehicle that looks nothing like the real thing.
With trainable video, you are able upload a dataset and train the video model on all angles of the car. This has been achievable with still images for quite some time (Flux) and I feel like I have been waiting for video LoRAs forever. The time has come 🤗 .
Although still early days, the outputs from simple text prompts are consistent through multiple angles: the design, features and paint color of the example vehicle (2025 Audi E-tron Quattro) persist across video generations. (Yes the spinning rims are a problem).
I included a few images from the dataset (still frame CG renders of the car) in the first comment in case anyone wants to compare the fidelity and consistency across the text-to-video generations in the edit.
AND this is only testing with a still frame dataset- Wan 2.1 also takes video training data which will be a whole other animal... Fun times! 🙌