Replies: 8 comments 10 replies
-
|
Hi @naxci1 - is there any difference between doing that and using a ResizeAndPadImage node before going into SeedVR2 ? ResizeAndPadImage is a native ComfyUI node to resize images |
Beta Was this translation helpful? Give feedback.
-
|
I've just started using ComfyUI, actually I started using it with SeedVR2, so I'm not very knowledgeable and I'm continuing to do these things the old way, but it can also be added as a node. |
Beta Was this translation helpful? Give feedback.
-
|
Sounds good - it's just nice to do everything in the same workflow and avoid having to deal with temp file / potential quality loss recompression, etc... |
Beta Was this translation helpful? Give feedback.
-
I asked ChatGPT to create this app for me. After countless iterations, it finally turned into exactly what I need. This app can simply break videos into equal number of segments as specified by the user. There is also option to downsize resolution to 1/4, 1/3, 1/2, 2/3, 3/4, and 1 (no scaling). Since SEEDVR2 works better by upscaling, I would downsize a 1080p video by 1/3 to 640x360 and then upscale in SEEDBR2 by 3x back to 1080p or for 720p, I would downscale by 1/2 to 640x360 and upscale in SEEDVR2 2x back to 1080p. Becareful though, sometimes downscaling the video too much causes too much original detail/info to be loss such as downscaling a 1080p to 1/4 resolution and upscaling 4x back to 1080p, the resulting video is actually worst than processing SEEDVR2 at 1x due to too much loss of original details but there is less flickering the more you upscale. Another example is sometimes downscaling a 1080p video to 960x540 and upscaling it with SEEDVR2 back to 1080p yield better results than downscaling to 640x360 and upscaling it with SEEDVR2 back to 1920x1080 though this is usually not the case. You need to test each individual video.
What this tool has that Lossless Cut doesn't is the ability to have the end frame of a preceding segment terminate at the closest key frame and the start frame of the succeeding segment begin at that "exact" key frame for seamless transition allowing for lossless direct-stream cuts when using 1x (no resolution scaling). If you don't check this option or if you choose a downscale less than 1.0, then the entire segments will be reprocessed, in which case it doesn't even matter if the termination points land on key-frames. With Lossless Cut, if you split the video into multiple equal segments and choose the "align segment times to keyframes" function, the problem is the end key frame of a preceding segment might not land on the same key frame as the start key frame of the subsequent segment causing gaps or overlaps resulting in non-seamless transitions when you rejoin the segments. What Lossless Cut can do that this tool can't is the "smart cut (experimental)" function where it can have the segments terminate at exact frames between key-frames and only process the few frames at the ends of the segments between the key frames while keeping the rest of the segments unprocessed and lossless. This function also only works for no resolution scaling (obviously) and only for some videos. It often produces errors from my experience so you must check each file to see they open even if they output successfully. This is why I used ChatGPT to create this tool. 
I was about to share this tool and just learned in a post above that there exist the ResizeAndPadImage node so I incorporated that into the workflow by LHKJacky This way you can use this tool to split the file at exact matching keyframes without resizing and have the downscale done within ComfyUI. Just drag and drop this png file into your ComfyUI to load the workflow. |
Beta Was this translation helpful? Give feedback.
-
|
Hi @skv89 Great work, congratulations. I've done a simpler version before, and I shared it on the forum. I ran the file you added, but it was a different application, and it didn't work. It would have been better if you'd added the Python version. It's easy to make it executable in Python. That way, if there are problems, the code can be fixed immediately. When I expand the workflow, it gives an error. Can you attach it as a file? I couldn't see the details. It's too small on the screen. |
Beta Was this translation helpful? Give feedback.
-
|
Sorry I just realized I uploaded an outdated version of the tool. I edited my post above with the latest version and included the .py file also. |
Beta Was this translation helpful? Give feedback.
-
Thx |
Beta Was this translation helpful? Give feedback.
-
|
Hi @skv89 There were some errors in the code, so I fixed them. I resized it correctly and added a settings button. When you save the settings, it saves them in a file and reads them from that file when you reopen it, eliminating the need to re-enter the settings each time. I created a pyw file so it works directly. I made some adjustments to make it easier. I hope you like it. A tip: ChatGPT isn't good for coding. Claude is the best, followed by Gemini. I use Gemini because it's free. 
|
Beta Was this translation helpful? Give feedback.
-
|
I reviewed the workflow, and 3x upscaling actually yields better quality. A batch size higher than 101 would be even better. So, I lowered the fps to 15 and reduced the video size. Enabling cache_model speeds things up even more. Since preserved_vram doesn't work, I disabled it. Instead of downscaling with Resize and Pad Image, downscaling with the code you wrote without losing quality and then adding it gives better quality. Set frame_load_cap = 0 so it can process the entire video. I cut the video into 3-minute segments and set the workflow to import all the videos in the directory sequentially. I add the same number of videos in qeury, and it sequentially upscales all the videos in the directory. I've shared this workflow before. 
|
Beta Was this translation helpful? Give feedback.
-
I'm using this right now. I disabled the automatic size because it creates different sizes, but I need 720p so YouTube can accept it in HD. So I manually set it to 720. To avoid OOM errors, you need to split the videos into 1-3 minute segments. I reduce the fps to 15, and when finished, increase it to 30 or 60 fps. |
Beta Was this translation helpful? Give feedback.
-
|
When you say "automatic zoom", do you mean the "ResizeAndPadImage" node? From my testing it works just fine and the scale remains intact as long as you calculate the exact downsize for both the width and height. Now with this modified workflow, I don't even need to use that AI created splitter app to downscale. I just use that app to split videos at 1x resolution with the key-frames option checked to avoid re-encoding. |
Beta Was this translation helpful? Give feedback.
-
I wrote something wrong there. Instead of doing this with Node, I think it would be better to do it yourself by reducing the video size as much as you want. I think I do it this way. Because I choose the codec I want. Of course, everyone's ease is different. |
Beta Was this translation helpful? Give feedback.
-
|
You mean if you do it yourself you can choose a codec that can compress the file size better? But how does that affect the overall process because the final output size after SEEDVR2 processing is based on the bitrate you choose in the VHS Video Combine node. Does having a smaller input file size affect the speed of the SEEDVR2 rendering or the quality of the final output? From what I've experimented with so far, it seems like using the node to downsize is the fastest approach. |
Beta Was this translation helpful? Give feedback.
-
Each node added to the workflow means additional work and more lost time. That's why I don't add anything to the workflow. Because my videos are long and numerous, the key is faster processing. I already do other tasks separately, so I want them faster without wasting time. The most time-consuming part is SeedVR2 video upscaling, which I want to do as quickly as possible. I use H265 or AV1 codecs for video output, which offer the best quality and smaller file size. |
Beta Was this translation helpful? Give feedback.
-
|
Hi. @skv89 I spent 5-6 hours today working on this app. It had a lot of issues and bugs, so I fixed them. I think it turned out to be a good thing in the end. Let me know if there are any bugs, it's still in beta. 
|
Beta Was this translation helpful? Give feedback.
-
|
Wow very impressive! Though I think you should still keep the Help button for the Quality, which is helpful for beginners. |
Beta Was this translation helpful? Give feedback.
-
I didn't realize that the AI deleted it itself, it can be easily added again. |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Before the SeedVR2 process, I reduce the size of my videos by 50% to achieve greater enlargement, because the more upscaling, the higher the quality of the output video. I've found that upscaling 3x yields better results. I use the free Shutter Encoder app to do this, and once the process is complete, I combine the audio with the video. You can easily perform many other operations.
After adding the video to the app, select the output codec (e.g., H265) and select Image Scale on the right. You can reduce the size of the video by 50% by opening it and selecting 50%.
https://www.shutterencoder.com/
Beta Was this translation helpful? Give feedback.
All reactions