Flowframes Early Access Download: 1.38.0
Added 2022-07-05 22:37:21 +0000 UTCSo here it is, the first Flowframes version with VapourSynth-based RIFE, eliminating the need to store all the frames on disk.
Do keep in mind that some things don't work with the new implementation yet, like FPS limiting/downsampling.
Full changelog on Discord and in Flowframes.
Comments
sadly this version is not woking for me at all. fail to load any type of video file it always gives an error and thats it.
HakuryuuFilms
2024-03-10 16:54:14 +0000 UTCTime to get an update I guess. Thanks.
Indulge Me
2022-07-20 13:33:29 +0000 UTCCorrect, I was trying to trim. Thanks! Looking forward to future updates.
Rici
2022-07-16 21:50:19 +0000 UTCThere are a few problems I found. 1. I saw a message saying "Set multi gpu threads to 1 or 2" when using multiple gpu. I am using amd and nvdia graphics card together, but amd must be deactivated before it works. 2. When running a specific video codec in ncnn/vs (the video was recording with obs amd-h.264), the whole process was skipped and it was completed in 2 seconds.
jo
2022-07-15 05:39:57 +0000 UTCTrimming won't work for RIFE-NCNN-VS if that's what you tried. Should be implemented in the next update.
LTSM
2022-07-14 06:37:19 +0000 UTCHello, when trimming start / end times, it just keeps going. E.G. start of 00:20:00 and end of 00:21:00 just for a small sample clip. But it keeps on encoding far past the number of frames that should be expected.
Rici
2022-07-14 06:19:01 +0000 UTCQuality should be the same
LTSM
2022-07-11 15:27:56 +0000 UTCThanks, and lastly if you have the time, is NCNN/VS higher quality than CUDA?
Jake
2022-07-10 20:39:14 +0000 UTCIn the installer, select "Vulkan + CUDA" on the framework selection screen
LTSM
2022-07-10 14:00:19 +0000 UTCI still don't know what it happens, but it should not be a problem.
LTSM
2022-07-10 13:59:33 +0000 UTC1) Fill fix soon 2) That happens when your GPU runs out of memory. 3) You can resume by dragging the temp folder into Flowframes, no the video 4) Will look into this 5) Probably related to GPU ID, working on a fix
LTSM
2022-07-10 13:59:21 +0000 UTCis there a version of this with the embedded python? Or do you have a link to an idiots guide to installing the relevant things under windows? I'm new to python and not having much luck getting anything to work :(
Tom Bass
2022-07-10 12:29:22 +0000 UTCThis is brilliant, no more bloat during processing! You do exemplary work!
Kezo
2022-07-09 08:33:31 +0000 UTCWhat an improvement. Thanks! It's come a long way since the first version! Love not needing to reserve space for PNGs...
Andy Zou
2022-07-08 13:50:26 +0000 UTCThought about that but it turned out it would only have around 30% more performance than my overclocked 3060 TI (connected via USB-C in an eGPU-case). However, it would be really awesome if I could use the integrated NVIDIA T600 (4 GB VRAM) for nvEnc-encoding while using the 3060 TI for the AI. That would speed it up a lot. Usually with the AI working 1,6 GB VRAM are free. When it is encoding though, free VRAM shrinks to around 100 MB, so I can't do a lot while it is encoding. Would there be a way to configure that somewhere?
Joel Wypior
2022-07-08 09:11:18 +0000 UTCThanks!!
Power AI
2022-07-08 03:41:18 +0000 UTCPerhaps for such heavy videos you should buy an RTX 3090 video card with 24 GB of video memory and a powerful GPU
Power AI
2022-07-08 03:39:45 +0000 UTCThe solution is: use the spatial media metadata injector YouTube provides. It just takes a few seconds and the tags are back.
Joel Wypior
2022-07-07 22:00:43 +0000 UTCI believe that's just the 360 degree tags missing in the file. If you open both files in Windows Media Player which doesn't support 360-degree vids and they look exactly the same, that's the case.
Joel Wypior
2022-07-07 21:59:28 +0000 UTCA question, when using dedepulication it always says; "Error: Index was outside the bounds of the array." Is that a concern at all?
Jake
2022-07-07 21:16:23 +0000 UTCThe problem is that it makes the video flat and after that it has to be recoded again to 360 degree format via Adobe Premier, which takes extra time and resources. There is also an unpleasant docking line that is visible at high speed, for example, when the video is shot on the roof of a car.
Power AI
2022-07-07 17:25:13 +0000 UTCI discovered several bugs in this version: - When trying to start the interpolation while it is still reading the meta data, it of course doesn't start, but after that it continues to load the whole input file endlessly and has to be stopped in task manager. - When doing other things while the AI is running and the VRAM fills up, it continues the job, even continues to create all the interpolated pictures, it even continues encoding, but from that point on the video ouput file doesn't grow anymore, so from that point on the work is lost. - Doing a step-by-step interpolation, when the extraction of images has finished, it's not possible to close flowframes, load the same video and continue with the 2nd part, the interpolation. - Also when cutting the frames folder, pasting it outside the temp folder and loading that instead of the input video, there is an error message almost instantly that no frames could be extracted, and there are already extracted frames. Even after manually running the last step (reset and clear cache). - VaporSynth doesn't seem to run on Intel UHD graphics (i9, 11th generation), even though it has 32 GB of shared VRAM left. Error message: "thread count has to be between" I believe 1 and -1, can't look it up right now. NCNN-Vulkan runs on it though.
Joel Wypior
2022-07-07 17:07:47 +0000 UTCI have already converted a lot of 3D-360-degree-videos in the top/bottom format and it worked flawlessly. Doesn't have to be explicitly supported I believe.
Joel Wypior
2022-07-07 08:39:03 +0000 UTCHi, is it possible to add support for spherical 360 video so that after recoding it also remains spherical and does not become ordinary flat?
Power AI
2022-07-06 23:06:17 +0000 UTCNo, frame count is needed and there's no way to get it without reading the entire file once
LTSM
2022-07-06 21:07:48 +0000 UTCWhen FlowFrames is reading the metadata, it reads the whole file, which in case of a 200 GB file takes a while. Would there be a way of reducing that?
Joel Wypior
2022-07-06 21:06:58 +0000 UTCWould there be a way of making VaporSynth use less VRAM? I'm in the process of interpolating my 5898x5898px stereoscopic 360 degree-videos from 30 to 60 fps. My RTX 3060 TI has 8 GB of VRAM. The input video format is H264 with around 1500 mbps. Using RIFE - CUDA doesn't work because there's not enough VRAM. Using Half-Precision mode uses only around 6 GB of VRAM but the resulting video flickers. NCNN/VS also tries to occupy more than 8 GB of VRAM and crashes. RIFE - NCNN works without half-precision mode though and only uses 6 GB of VRAM, works really well. But VS would be so much better speedwise and to not loose any detail in JPG compression. PNG-files would take up about 1 TB for each video...
Joel Wypior
2022-07-06 20:53:27 +0000 UTCThe guy that runs TVP keeps kicking me because I said imo Rife is better than TVP can you do something with him. he cant handle other peoples opinions
Neff
2022-07-06 20:04:48 +0000 UTCShould be fixed in the next update (~1 week), it will use the NCNN GPU ID set in the settings. But yes in 1.38.0 it will use the default GPU.
LTSM
2022-07-06 13:05:19 +0000 UTCThank you! VaporSynth doesn't honour the GPU ID though and always uses the 1st GPU. Would there be a quick fix? Unfortunately my first GPU doesn't have enough VRAM, so I can't use it.
Joel Wypior
2022-07-06 12:24:17 +0000 UTCThanks very much.
EvilMadness
2022-07-06 08:42:39 +0000 UTCCurrently, trying to work with 4k upscale of my selected anime collection. The current video has it seems combination of animating on 2s, 3s, and I think sometimes 4s... I've noticed in the past that there is a difference how dedup behaves if you use 2k vs 4k, where same settting seems to have been more sensitive on 2K. I pretty much turned it down do 0.5 for 2K animation and seemed OK to me to keep it like that. But I wonder, if you would do it, assuming everything else is ideal, no noise, compression artifacts... I took care of that in prep work, and usually use clean source.... What would you choose. Silly question maybe, but I can get OCD like that.... So would appreciate your opinion.
Milica Zivanovic
2022-07-05 23:08:13 +0000 UTCDeduplication works, but frame extraction is necessary for that, however the output frames are encoded in RAM so still better than the regular implementation. Sensitivity completely depends on your content. What are you trying to interpolate?
LTSM
2022-07-05 22:59:13 +0000 UTCEpic! Thanks man, this will save up some HDD space. You did not mention it in changelog, but based on your comment from yesterday, I suppose that accurate frame dedup is still not going to work with this new model? I'm sure that there will be many factors in source, that can influence this, but humor me if you don't mind... Which dedup sensitivity setting for 2k and 4k would you recommend?
Milica Zivanovic
2022-07-05 22:57:44 +0000 UTCThank you for all your hard work 😎
F2ngz
2022-07-05 22:38:03 +0000 UTC
