Hey folks, I've been running the Luma Dream Machine API through its paces for a few days, trying to integrate its video generation into a content pipeline. Overall, I'm impressed with the motion quality, but I keep hitting a consistent snag.
About 30% of my generated videos have these bizarre text artifacts. I'm not talking about the *intentional* text in the prompt, but garbled, random letters or symbols that look like a corrupted font, often flickering in the corner or superimposed on objects. It happens even with simple prompts like "a cat sleeping on a windowsill."
I'm using the standard API call, nothing fancy:
```json
{
"prompt": "a serene lake at sunrise",
"duration": 5
}
```
My initial thought was prompt contamination, but I've stripped it back to basics. I'm also batching jobs, and it doesn't seem tied to any specific aspect ratio or duration setting.
Has anyone else encountered this? I'm trying to figure out if it's:
* Something in the training data bleeding through.
* A decoding issue on my end when handling the MP4 stream.
* A common pitfall with certain descriptive words.
If you've found a workaround or a pattern to avoid, sharing would be a huge help. I'm currently adding a post-processing step to flag these, but it's not ideal.
it's always an API issue