Experimenting with AI-assisted 3D animation – Gemini to Leonardo AI workflow

Hi everyone,

I’ve been experimenting with an AI-assisted workflow for creating short 3D-style animated sequences, and I’m interested in getting some feedback from people who work with animation in Unreal Engine.

For this test, I used Gemini to create the initial visual/image, then used Leonardo AI to turn the image into an animated sequence. The goal was mainly to see how far this kind of image-to-video workflow can go before the result starts to look noticeably artificial.

Here’s the test:

I’m particularly interested in feedback on a few things:

  • Does the movement feel convincing, or does it still look obviously AI-generated?
  • How noticeable are problems with consistency between frames?
  • For a game-production workflow, where do you think this type of process could actually be useful?
  • Would you use AI-generated motion mainly for quick concept/prototype work, or do you see potential for final animation after cleanup?
  • What would be the biggest thing you’d improve if you were taking this result into Unreal?

I’m still experimenting with the workflow, so I’d especially appreciate technical or artistic criticism rather than just whether the video looks good.

If anyone here has tried a similar image-to-video workflow and then brought the result into Unreal, I’d be interested to hear how you handled the cleanup and consistency issues.

Interesting experiment. For me, the biggest potential of this workflow is probably in the early concept and previs stage rather than directly replacing a traditional animation pipeline.

The main things I would watch before bringing AI-generated footage into Unreal are temporal consistency, character/object identity, and camera continuity. A sequence can look convincing at first glance, but small changes in proportions, lighting, or object placement between frames become much more noticeable when you try to integrate it with actual gameplay assets.

For prototype work, though, I can see this being very useful. You could quickly test an idea, mood, camera movement, or even a gameplay concept before investing time in modeling and rigging everything. I could also imagine similar workflows being useful for architectural visualization—for example, when presenting concepts related to Bausanierung Frankfurt, where quick animated visual studies could help explore lighting, materials, or the transformation of an existing space before creating a fully detailed Unreal scene.

If I were taking the result further into Unreal, I would probably use the AI output mainly as a visual reference and then rebuild the important assets or animation in a more controllable pipeline. That way, you keep the speed of AI experimentation without being limited by inconsistencies in the generated frames.

It would also be interesting to see a comparison between the original AI-generated sequence and a cleaned-up Unreal version of the same shot. That would make it easier to judge where the workflow saves time and where manual work starts taking over again.