I was looking around to find how to get rid of duplicated frames and found this. I have to say, you got the idea I need but implemented in an unnecessary complicated way. First of all, you don't need a complicated diagram to load 2 adjacent frames. Just set increment for them seperately starting with 0 and 1, and max value at last frame - 1 and last frame. That will solve the trick. Secondly, to get rid of duplicated frame, just split the result batchs and throw away last frame. After the whole process was done, manually add last frame. That's is alot easier and faster. Too much step will cost more computing power, which is a big deal when we only have a low-end machine.
This has been very useful in the past couple of months. I realised that if you have a video as input or made on the fly (like CogVideo5) you can just feed it into the interpolater and videcombine the output, and that just works without trouble. But I was making a flow that did something to the images and then wanted to interpolate that live, but couldn't work it out. So of course, I can make it save images out and then go through another stage like the above to make the interpolation. What happens is I get loads of tiny videos which are just each batch run of a single frame, what I wanted it to do was interpolate the image sequence and then save the video at the end without it creating the intermediate images.
@@neuron_ai Thanks for the reply. After I wrote that, I tried something with "meta batch". This combined with loadvideo, to which it connects, had to be set to the right number of input frames in the video, and bingo, it worked!
Great tutorial, do you happen to know of a way to use this concept/method to continue a previously generated clip output to another one? I'm trying to create a much longer animation out of many shorter ones.
Yeah, there's a ton of on-line services that seem to do it like ltx studio but they basically do transition, I can do that in Adobe, your method seems far more less likely to have issues and as long as I understand the nodes you use properly, your method seems to have less chances of hallucinating while making the transition from one clip to another.
This is so slick!! I've been playing with SD hyper 1step gen and animate diff now I need to interpolate, and this is exactly what I've been searching for!! Thanks a bunch!
If you use animatediff or svd or something like this and you want to make interpolation it can be hard on the vram because it gets done in one step. with this workflow you can interpolate between frames step by step. so you dont need large vram.
tooncrafter seems more like a photo to comic transfer. this here is a special workflow to use if you have many single images and want to interpolate between them.
@@nholmes86 yes thats what i mean. yes it might also been integrated in tooncrafter. if you want to learn how to use it in comfyui better watch this video: ruclips.net/video/twbrAKkyJ4s/видео.html This video here is made for big folders of images and how to batch process them.
This is great, and almost works. I found the final save image box is broken, but reverted to the simple save image. Also, and not your fault of course, my idiocy, I spent ages trying to get the pattern matching , thinking say you have image_00055.png etc, you would have a start seed of 55.. and puzzling how you'd specify 5 digits - then realised it's merely taking any files presumably in sort order, and so seed 0 for that reason.
Sup bro , i wanted to ask you when doing animatdiff is there any way to process the 16 frame batch size 1 by 1 and then combine them cause my 4gb gpu can't render 16 image batch at once
Hi Neuron. Really cool you made this video! Very impressive and massive thanks!! Nevertheless I found two things rather curious. Maybe you want to have a look into that: - With this workflow, it seems that it doesn't save more then the fist 3 images.. - The Load Batch Node loads ALL images in your folder.. twice.. which might take up a few minutes when you have a lot of images in your folder..
Thanks! :) I tried the workflow again, and for me it works and goes through all of my frames. Be sure that the max_value of the base incrementor node is bigger then your amount of frames. The double loading of the images is not the best solution, you are right. Maybe I come up with a different solution for this.
@@neuron_ai Hi Neuron. I re-tried a few more times with other setting and with other datasets: and it seems to work now! I have no idea why it didn't work on my initial testing. But I wanted to let you know your workflow is working well. I'm currently stress-testing it on a huge batch, to see if it would run out of memory eventually. (which in theory it should not)(I will keep you informed!) Is there a way we can load parts of a batch / specific images only, in order to speed up the process? So let's say I have 100k images, the workflow will load ALL 100k images, which takes likes a few minutes, and seems unnecessary process time. Looking forward to hear from you. Many thanks!
@@ARTificialDreams Hey, I am glad it works for you now. Its a tricky workflow, and I had some strange things happening when buidling it. Might be some buggy nodes. You can divide your images in parts by simply adjustinge the seed of the incrementer node to the starting frame -1 and the second one to the starting frame and set the max value to starting frame + amount of frames you want to process. that way you can render everything in parts.
@@neuron_ai Hi Neuron. I tried out a small batch, and went well. I tried a small batch with multiplier 10 and also went well. Alright, the big test: I retried and let it run loading a folder of 190k images, having the first frame starting at 15000. multiplier at 10... ...the animation would go well as intended until frame 120. But then the exported frame 120 is a random frame from my folder. It seems like that the 'Load Image Batch'-node 001 has have some issues loading the right image. Node 'Load Image Batch 002' seems to still be correct loading the next one. I can make a video and upload it, if you want to see the results. Anyway, maybe you know what the issue might be. I suspect that the node itself might have issues handling bigger numbers. Or maybe we can flush the memory inbetween? Alternative I was thinking about, we might get the same results with loading a folder into animate-diff? So far I believe that has limitations as it couldn't handle SDXL in the past, but maybe now it's possible and open opportunities to explore new workflow to solve the issue? Let me know what you think! And would be awesome if you have a solution. Many thanks!!!
@@ARTificialDreams i know theres a ksampler with restarts perhaps a way of "restarting" at current frame before it gets to 120? Thanks for your testing also!
![Shoot EPIC MOVIES with this FREE AI Tool! [ComfyUI Tutorial + Free Workflow]](http://i.ytimg.com/vi/gHI6PjTkBF4/mqdefault.jpg)
I was looking around to find how to get rid of duplicated frames and found this. I have to say, you got the idea I need but implemented in an unnecessary complicated way.
First of all, you don't need a complicated diagram to load 2 adjacent frames. Just set increment for them seperately starting with 0 and 1, and max value at last frame - 1 and last frame. That will solve the trick.
Secondly, to get rid of duplicated frame, just split the result batchs and throw away last frame. After the whole process was done, manually add last frame.
That's is alot easier and faster. Too much step will cost more computing power, which is a big deal when we only have a low-end machine.
I might do it different nowadays.
This has been very useful in the past couple of months. I realised that if you have a video as input or made on the fly (like CogVideo5) you can just feed it into the interpolater and videcombine the output, and that just works without trouble. But I was making a flow that did something to the images and then wanted to interpolate that live, but couldn't work it out. So of course, I can make it save images out and then go through another stage like the above to make the interpolation. What happens is I get loads of tiny videos which are just each batch run of a single frame, what I wanted it to do was interpolate the image sequence and then save the video at the end without it creating the intermediate images.
this is a step by step approach. unfortunatly this is not working with video output. you have to combine the frames into a video afterwards
@@neuron_ai Thanks for the reply. After I wrote that, I tried something with "meta batch". This combined with loadvideo, to which it connects, had to be set to the right number of input frames in the video, and bingo, it worked!
@@geoffphillips5293 oh nice one. didnt know this node. will give it a try as well.
Great tutorial, do you happen to know of a way to use this concept/method to continue a previously generated clip output to another one? I'm trying to create a much longer animation out of many shorter ones.
like conecting several sequences?
Yeah, there's a ton of on-line services that seem to do it like ltx studio but they basically do transition, I can do that in Adobe, your method seems far more less likely to have issues and as long as I understand the nodes you use properly, your method seems to have less chances of hallucinating while making the transition from one clip to another.
This is so slick!! I've been playing with SD hyper 1step gen and animate diff now I need to interpolate, and this is exactly what I've been searching for!! Thanks a bunch!
happy to hear that :)
Hi, I have a PC with a low spec. I use comfy UI Ok. for image generation. What's the purpose of this workflow? Is it mainly for video generation only?
If you use animatediff or svd or something like this and you want to make interpolation it can be hard on the vram because it gets done in one step. with this workflow you can interpolate between frames step by step. so you dont need large vram.
Thanks!! Exelent work.
A question is this similar interpolation like tooncrafter?
tooncrafter seems more like a photo to comic transfer. this here is a special workflow to use if you have many single images and want to interpolate between them.
@@neuron_ai you mean to create smooth transitions, in-between images? if its then is actually similar workflow with toon crafter
@@nholmes86 yes thats what i mean. yes it might also been integrated in tooncrafter. if you want to learn how to use it in comfyui better watch this video: ruclips.net/video/twbrAKkyJ4s/видео.html
This video here is made for big folders of images and how to batch process them.
hi. when i select allow RGBA output, it doesn't output a transparent png. any idea how i can solve this?
thats strange. check all the settings. this should be possible.
This is great, and almost works. I found the final save image box is broken, but reverted to the simple save image. Also, and not your fault of course, my idiocy, I spent ages trying to get the pattern matching , thinking say you have image_00055.png etc, you would have a start seed of 55.. and puzzling how you'd specify 5 digits - then realised it's merely taking any files presumably in sort order, and so seed 0 for that reason.
Hello, good to know that the advanced save image is broken. So many nodes a broken since the latest comfyui updates. Yes, zero is file 1. :)
Very clever, thanks 4 sharing this
awesome tuto thanks !! can you make one but this time to create image one by one with a path of pose image already done ?
Not planed at the moment. But I keep it in mind.
Sup bro , i wanted to ask you when doing animatdiff is there any way to process the 16 frame batch size 1 by 1 and then combine them cause my 4gb gpu can't render 16 image batch at once
yes, this should be possible. you have to devide the batch. I can tell you the nodes later.
amazing! love your chan
Hi Neuron. Really cool you made this video! Very impressive and massive thanks!!
Nevertheless I found two things rather curious. Maybe you want to have a look into that:
- With this workflow, it seems that it doesn't save more then the fist 3 images..
- The Load Batch Node loads ALL images in your folder.. twice.. which might take up a few minutes when you have a lot of images in your folder..
Thanks! :) I tried the workflow again, and for me it works and goes through all of my frames. Be sure that the max_value of the base incrementor node is bigger then your amount of frames. The double loading of the images is not the best solution, you are right. Maybe I come up with a different solution for this.
@@neuron_ai Hi Neuron. I re-tried a few more times with other setting and with other datasets: and it seems to work now!
I have no idea why it didn't work on my initial testing. But I wanted to let you know your workflow is working well.
I'm currently stress-testing it on a huge batch, to see if it would run out of memory eventually. (which in theory it should not)(I will keep you informed!)
Is there a way we can load parts of a batch / specific images only, in order to speed up the process?
So let's say I have 100k images, the workflow will load ALL 100k images, which takes likes a few minutes, and seems unnecessary process time.
Looking forward to hear from you. Many thanks!
@@ARTificialDreams Hey, I am glad it works for you now. Its a tricky workflow, and I had some strange things happening when buidling it. Might be some buggy nodes. You can divide your images in parts by simply adjustinge the seed of the incrementer node to the starting frame -1 and the second one to the starting frame and set the max value to starting frame + amount of frames you want to process. that way you can render everything in parts.
@@neuron_ai Hi Neuron. I tried out a small batch, and went well. I tried a small batch with multiplier 10 and also went well.
Alright, the big test: I retried and let it run loading a folder of 190k images, having the first frame starting at 15000. multiplier at 10...
...the animation would go well as intended until frame 120. But then the exported frame 120 is a random frame from my folder.
It seems like that the 'Load Image Batch'-node 001 has have some issues loading the right image. Node 'Load Image Batch 002' seems to still be correct loading the next one.
I can make a video and upload it, if you want to see the results.
Anyway, maybe you know what the issue might be. I suspect that the node itself might have issues handling bigger numbers.
Or maybe we can flush the memory inbetween? Alternative I was thinking about, we might get the same results with loading a folder into animate-diff?
So far I believe that has limitations as it couldn't handle SDXL in the past, but maybe now it's possible and open opportunities to explore new workflow to solve the issue?
Let me know what you think! And would be awesome if you have a solution. Many thanks!!!
@@ARTificialDreams i know theres a ksampler with restarts perhaps a way of "restarting" at current frame before it gets to 120? Thanks for your testing also!
Thanks 4 sharing
Isn't flowframes more easy and better?
This video is about doing this in ComfyUI. There a thousand ways of doing this in a different way.