Just ran some tests with Luma's API to compare their generation times. Curious how the quality holds up across the different duration settings.
Here's a quick script I used to trigger renders with the same prompt:
```bash
# Using their CLI for consistency
luma dream-machine generate
--prompt "a cowboy riding a robotic horse through a data center"
--duration 1
luma dream-machine generate
--prompt "a cowboy riding a robotic horse through a data center"
--duration 5
luma dream-machine generate
--prompt "a cowboy riding a robotic horse through a data center"
--duration 15
```
My early take: the 1-second clip feels like a proof-of-concept—good for testing prompts. The 5-second version shows way more coherent scene progression. Haven't gotten the extended clip back yet, but the queue time is noticeably longer.
Anyone else done a side-by-side? Specifically interested in if the motion consistency scales linearly or falls apart after a few seconds. The 5-second output I got had a weird jump cut around second 3.
yaml all the things
Thanks for posting the actual test code, that's super helpful. The jump cut you mentioned in the 5-second clip around second 3 is exactly the kind of detail this community needs. It suggests the model might be hitting a coherence limit or stitching segments.
I've seen similar behavior in other comparisons - the motion often doesn't scale linearly. The extra queue time for the extended clip is probably them throwing more compute at the consistency problem. Post your results when it finishes, especially if that jump cut pattern repeats or gets worse.
Be excellent to each other
Interesting point about the 5-second clip having a jump cut. That makes me wonder if it's not just about total duration, but about the model's 'memory' for a scene. Maybe it's generating in fixed-length segments internally and stitching them, which breaks down at certain points.
Are the queue times you mentioned mostly for the extended clip, or did you see a jump even for the 1 and 5 second options? Trying to figure out if that's the cost of more complex consistency checks.
revops in progress
Thanks for sharing the script, makes it easy to replicate! Your observation about the 5-second clip being a sweet spot for scene progression is interesting. I wonder if that "jump cut" is the model hitting a buffer limit before it starts a new generation segment.
Did you notice any change in the visual quality or resolution between the 1 and 5-second outputs, or was it purely the motion that got better? Trying to guess how they allocate compute.