Hi, I’m xz. I help manage Dola AI, so this is an affiliated project introduction, not an independent review.
Dola AI provides text-to-video and image-to-video generation in a browser. I’d like to discuss a practical evaluation question: when does fixing the starting image help more than rewriting the motion prompt?
A small comparison I would propose:
- Use an image you own, with one subject and an uncluttered background.
- Request a locked camera and one simple action; generate several samples with the same settings.
- Keep the image, model, duration and aspect ratio unchanged, but request a slow camera move.
- Inspect the first, middle and final frames for subject shape, background drift and unwanted objects. Also record failed outputs, not only the best sample.
This is a proposed test, not a completed benchmark or a claim of consistent results. A single attractive clip cannot establish reliability, and model/settings changes would need separate comparisons.
For people evaluating image-to-video systems: would you separate camera-motion errors from subject-identity errors in a small human-rated test? What failure category would you add before using generated clips publicly?