Has anyone done a test running a clip through both the cloud-based Project Starlight and Starlight Mini, and comparing the results?
I boxed myself into the position of only being able to use the cloud version, because I gambled on an AMD GPU before Starlight Mini came along, and my card is currently not capable of running it. (And I have little hope that it will be supported anytime soon. Lesson learned, the hard way…)
I’ve used my initial 100 credits and several weeks’ worth of free 10 second evaluation clips, and have seen some very promising results from the Cloud version. But I’m wondering about the term ‘Mini’. Beyond it’s being scaled so that it runs locally on NVIDIA GPUs, are there significant differences in the output quality due to a technologically-limited implementation on less capable hardware?
I’m at the point of considering eating the $800 that I spent last year on the AMD card and spending a bit more on top of that to get a halfway-decent NVIDIA card that will run Mini, thinking maybe an RTX5070 ti with 16GB, as a minimum configuration.
I can deal with long render times as long as I’m not taking a quality from as a result of running locally vs paying for the cloud version.
I did two or three tests with the very first initial Cloud version of Starlight and wasn’t really impressed that much as it kinda looked like a smoother version of RheaXL.
In contrary SLmini is really impressive for the largest part (except for the only very occasional monster face and distorted text). What’s especially nice is that it doesn’t “overdo” things like many of those other regenerative models.
But I don’t know how the current iteration of the full cloud Starlight performs - so it could be that maybe SL has generally gotten much better - or it’s an advantage that SL mini is a stripped down version and does less…
Thanks for your observations. Because I’m AMD-bound, the only tests I’ve performed thus far have been in the cloud. As it happens, I’ve repeated a couple of the 10-second previews two or three times, and have seen different results, most notably between the early version of the cloud processing - when it took 30 - 40 minutes to process - and the later version, which now takes 3 - 5 minutes. And I’ve gotten better results with the newer, faster version.
I did one longer clip, which used up most of my 100 credits. It was the introduction of a TV program from the 1980s and started out its digital life on a TV station’s digital DVD recorder. The introduction was nearly 40 seconds long, and the cloud did the entire thing, taking up about 72 of my credits. But it came out as a single file, of course.
I had to make a minor change and re-did the whole thing, breaking it up into 3 10-second free previews and the rest of my credits for the 4th, leftover segment. I’ve noticed just the smallest of differences at the transition points, almost like Starlight has a region of overlap before and after each frame to figure out how to process it. Given there’s nothing before the first frame of each file, it lacks this overlap and doesn’t seem to be able to balance the result like it would have if there were more frames before this first one.
All of the previews I’ve been sending to Starlight have originated from bad, low-res source footage like this, and have turned out surprisingly well. I’ve been working with another user on restoring these programs for DVD sales and special broadcasts, and some of source videos we’ve worked with have a lot of tape dropouts from the masters. We have to go in frame-by-frame and correct them, where possible, by judicious copy/paste from adjacent frames. Tedious and time-consuming, to say the least.
But I sent an example of a dropout-heavy clip to the Starlight, and it did a great job of removing them all by its lonesome. Doesn’t work quite that well on film scratches, though, as they’re vertical and pretty continuous for many frames at a time.
Is it ok if it’s a pro-res .MOV file at 59.94fps about 45 seconds long? I had to change the framerate to 29.97 be able to do it with 10 second free previews, but it would be preferable if we could leave it at 59.94
Remember, I can’t run Mini due to the lack of an nVidia GPU. So my only tests have been 10-second free previews with the cloud version.
As I recall, the only upscale option in the preview seems to be 2X, which would be HDV 1440 X 1080 for this file.
The filename is somewhat misleading. My colleague, who wrote this file, has some secret sauce involving avisynth scripts and miscellaneous voodoo that he performs on the front end. He did run it through Topaz/Iris, but only to deinterlace it. If you look at this screenshot, you’ll see that all the enhancement paramaters are set to 0.
This is pretty far from the original source file, which was from mpeg-2 VOBs from the DVD recorder. But it’s the only file that I have that’s been processed by online Starlight and can be used to compare what eventually comes out of Starlight Mini.
