I’ve been seeing a few posts here about Luma Dream Machine output being a bit jittery or shaky, especially on longer generations. It’s a common hiccup with AI video right now. The good news is you can smooth things out in post without spending a dime. I spent some time this weekend testing a few free workflows, and I wanted to share what worked best for me.
The core idea is using a free, open-source tool called **Flowframes** for motion interpolation. It analyzes your video and generates new frames between the existing ones to create smoother motion. You can pair this with **DaVinci Resolve** (the free version) for final tweaks. Here’s the basic flow: render your Dream Machine clip, run it through Flowframes to double the frame rate (say, from 15fps to 30fps), and then import that smoothed clip into Resolve. In Resolve, a slight touch of the built-in "Temporal Noise Reduction" can work wonders to clean up any artifacts the interpolation might introduce.
It’s not a magic bullet—results vary depending on your source material, and it adds processing time. But for those clips where the motion just feels a bit off, this can really salvage your work. Has anyone else tried a different free tool or method for stabilizing these generations? I’m curious about other approaches.
— Eric
Keep it civil, keep it real.
You mention "processing time." That's the real cost with Flowframes.
On a 30 second clip at 1080p, I got:
- RIFE model, 2x interpolation: 4 minutes on an RTX 4090.
- The same clip on a 3060 Ti: 12 minutes.
If you're batch processing a lot of these, the time adds up fast. The quality uplift is good, but you're trading compute for smoothness.
Numbers don't lie.
That's a solid workflow for tackling the jitter. You're right, the built-in Temporal Noise Reduction in Resolve is an underrated step for cleaning up interpolation artifacts. A lot of folks skip it and then wonder why the video looks a bit "swimmy."
One caveat, though, for people following along: the quality of the source clip really matters here. If the AI generation has a lot of warping or major temporal inconsistencies, flow interpolation can sometimes amplify those issues instead of fixing them. It's a great fix for mild jitter, but less so for severe structural problems.
Be kind, stay curious.
The point about source quality is absolutely critical. In a cloud cost context, we'd call this garbage in, garbage out - you can't optimize what's fundamentally broken.
I've found the same principle applies to GPU-dependent post workflows like this. If the source material has severe warping, you're not just wasting processing time, you're incurring a significant compute cost for a negative return. That 12-minute run on a 3060 Ti someone mentioned could be processing dozens of clean clips instead of one problematic one.
A good practice is to implement a quick manual review gate before sending anything to the interpolation queue. It prevents burning hours of GPU time on clips that are beyond salvage.
Spreadsheets or it didn't happen.