Sorry I failed to mention that. I had already tried updating the dependencies, but forgot to mention it. That didn’t help either.
Here is a link (DV AVI Clip for Testing) to download the file and test it yourself.
Yes, I have the same problem! Tried several times.
Similar problem as well, I had the program install all the dependencies but would end up failing during export.
Gave Video Denoise Studio a go, no errors on the export at all, and while the end result appears to be very subtle, if I zoom in close I can see that the image was infact cleaned up, you can really notice it on solid backgrounds and certain skin tones.
However, when running it through SLP 2.6 2x (480p to 960p), I started getting the dreaded blured out tile issue, only now it is one big tile that takes up most of the screen using the 1.0.0 launcher. Now if I go back and I run the original file prior to running the denoiser, no tiling issues at all. Not sure what to make of that. I really like how it cleaned up. My work flow as follows:
- Capture analog at 720x480 via lags lossless avi
- Deinterlace using Hybrid QT/Vapor and maintain 29.97 fps
- Used the Denoiser Studio at this stage, exported to ProRes 4444
- Color Corrected in Premiere Pro, exported 29.97 to ProRes 422 STD
- Laucher 1.0.0 ==> SLP 2x (480p to 960p) Sharp set to 3
So again, if I remove step 3, no issues at step 5, but with step 3, I get the random blurred out tile issue showing up.
Based on my extensive experience in video restoration, I can definitively say that the best approach is to feed the source video into SeedVR2.5 or SLP without any modifications. While pre-processing might seem good, AI models don’t understand it well and it gets corrupted. Therefore, even in the worst-case scenario, the video should be processed as is, without any alterations.
6 Likes
The only thing I am doing in Premiere is white balance correction and in certain cases, adjusting the brightness and contrast on vhs footage that in somecase gets too dark, as I read that making sure the brightness is adjusted correctly and basic color correction would be good to do before Topaz and that actual color grading be done afterwards. Otherwise, Topaz would struggle on say a reeally overly dark video for example. Curious your thoughts on this?
EDIT: I have also ready that if possible, try to do any and all adjustments on the front end prior to capturing with a decent proc-amp. And my first choice would be a Signvideo PA-100, but they are near impossible to find anymore, so I am considering a BVP-4 Plus to add to my capturing chain so I can make those adjustments before digitizing. So it would be:
- Panasonic AG-1980
- DataVideo TBC-1000
- BVP-4 Plus
- Magewell Pro Capture AIO or GV-USB2 (depending on the source, but normally the Magewell).
I owned a BVP-4 Plus before and I am well aware that you need to use it carefully, as their false claims of adding resolution back in the day just resulted in unwanted video noise. So I would only be using it for dealing with overly dark or bright videos and adjusting the color tone if something is way out of wack, basically just the proc-amp features.
I’ve experimented with many different combinations, most of them theoretical and not practical. You add light to the video, but details are lost, so AI has difficulty understanding it and produces a low-quality output. The best thing is to leave it as is; then the details are well preserved, and AI perceives everything correctly. Other processes should be done after the AI upscales. That’s how I do it. If the video needs cropping or there are size issues, I only correct those and don’t do anything else. I process the video as is with AI. This is the only way to best preserve details. Apply subsequent processes to the upscaled video. Don’t even denoise, don’t add lighting, don’t do anything at all; that would do more harm than good. AI does everything itself, including denoising and stabilization; it’s all built from there.
If you are cropping or resizing, use FFV1 or ProRES HQ codecs for output to better preserve pixels.
2 Likes
I will recapture and give your suggestion a test, I am assuming no adjusts to the proc-amp at all, not even the built in ones on the capturing devices? The only resizing I do is at the time of deinterlacing in VirtualDub, and I adjust for square pixels. But one question, is it best to downscale to the 640x480 route or to preserve the horizonal and upscale the verticle to 720x540. There seems to be people who swear by both methods, but I think most prefer the 640x480 method.
1 Like
Don’t apply any filters; the original raw image will give the best output quality, believe me. It preserves details very well. Regarding downscaling, only reduce it by 50% if absolutely necessary; other reductions still distort the pixels. I’ve tried different reductions and found that they distorted pixels and details. Only a 50% reduction preserves the pixels best. For this, I use Shutter Encoder, a free app, with the FFV1 codec and a 50% scale. It processes it in about 10 seconds.
Also, I made an error, I use Hybrid/QTGMC for deinterlacing and cropping to 640x480, not VirtualDub.
Responded in my Deinterlace Studio thread. The app is now updated. Thanks for reporting issues with the necessary information that I can actually use!
1 Like
In my Video Denoise Studio app, there is a preview that works exactly the same as Topaz’s video preview. Click on it to show the source frame and release button to see the denoised frame. The preview actually processes the neighboring temporal frames so you have an accurate representation of how the specific preview frame will actually look like. I set the default denoise strength to be very mild and conservative and the default number of temporal frames is also not set very high because in fast moving scenes, it can result in more smearing. So you should manually scroll to fast motion scenes in the preview to compare with the source image and increase the denoise strength or temporal frames as needed per your liking.
For me I use denoise on very noisy videos because SLP/SEEDVR2, while it has an internal denoise filter, can mistaken heavy noise as detail and generate details based on the noise. Topaz says SLP is intended primarily for medium- to high-quality content and specifically warns that it is “not intended for very low/poor condition quality media”; they recommend Starlight Mini instead for such material. That is because strong random noise can become part of the information the diffusion model tries to reconstruct. I have seen the noise baked into the detail on some of my test samples but I was able to almost entirely eliminate the noise derived details by pre-denoising the video. I also denoise to reduce dirty faces, which works to a certain extent but you need to increase the denoise strength. You need to run short test clips prior to a full run because every video is different.
I never seen or heard of the “dreaded blurred out tile issue” so I don’t know what it is. I have an ambitious plan to implement the denoising ability into my SLP Tuning Launcher to include a preview feature that takes into account the actual denoise settings so you can see how post-SLP processed image looks like with and without denoise pre-processing. This will make it MUCH easier to decide what denoise settings to use and whether or not to even pre denoise.
Further discussions of my Video Denoise Studio app can continue in this thread:
https://community.topazlabs.com/t/video-denoise-studio/104757
Edit: I notice you output with prores, try FFV1 Intra to see if that makes a difference. SLP will take much longer to load but you can just clip a 20s section for testing so it won’t load that long. Also have you tried to Deinterlace with my Deinterlace Studio prior to Denoise? On some of my interlaced videos with strong noise and artifacts, I deinterlace with my Deinterlace Studio and then denoise prior to feeding into SLP and the output looks better vs without denoising. Faces are also less dirty and less imperfect.
I don’t know why some sneaky individual is keep flagging all my posts and causing them to be hidden. If you have an issue with what I’m saying, please have the courage to explain yourself openly instead of hiding behind anonymous reports.
1 Like
I’ve observed the same thing as well. For example, if you preprocess the source with hqdn3d for denoising/smoothing, the result is generally worse than feeding the original source directly into SLP. Likewise, if you run Proteus pass 1x and then feed that into SLP, artifacts can sometimes appear.
So be careful with preprocessing. Only “gentle,” non-invasive processing might be okay. QTGMC deinterlacing, however, doesn’t seem to be a problem.
The SLM model also seems to be less sensitive to this than SLP.
1 Like
Just posted in the general section, added many comments and created some “design documentation” to help whomever wants to modify/improve the scripts.
There are 2 scripts:
- zmon is the monitoring script, that takes care of predicting when a video will be processed, also does a backup of frames in real-time, so you don’t lose your work when Video AI crashes or freezes
- zbuild builds the video from the PNG frames.
Hope this helps, comments are ideas for improvements are always welcome!
Dom
Posted in: Predicting Starlight Processing / Real Time Backup
How long should the “loading model” portion of the process take? Because it seems like I’m having to wait for over 5 minutes before anything starts happening.
I don’t know why some sneaky individual is keep flagging all my posts and causing them to be hidden. If you have an issue with what I’m saying, please have the courage to explain yourself openly instead of hiding behind anonymous reports. If there genuinely is a problem with something I’ve said, I’m completely willing to listen and will gladly modify or remove the message as appropriate. I just don’t understand the point of repeatedly flagging posts without explaining what the actual issue is.
9 Likes
It’s a false info. It is actually not loading model at the very beginning, but ffprobe is scanning the whole source video file. How long it will take depends on how long your video is. I almost take 22+ minutes when I try to upscale one hour video with SLP.
1 Like
any issue with the cloud processing in 1.7 tonight ? i have a hell to make it working. it erased a file i sent yesterday night, trying to send some files again this night and they are erased, not even an error message or something, the file is sent, i agree with the trim and credit, and then itdisapears from the list.
any idea what’s going on ?
Are you overclocking your GPU? That looks a lot like VRAM errors. GDDR7 has built in error correction, but it is easy to overwhelm it at heavy loads (such as any Starlight process).
