Im really really happy with the stuff you show us here, this is the kind of content me and many are looking for! In this case with this new lama node, would be interesting to compare instead of NNCleanup, it is so powerful and have solved me some things that otherwise would take days and weeks, even on moving footage instead of single frame. Thanks Alex, this is great!
Thanks for the kind words, I'm happy you are finding value in my videos so far! I had never heard of NNCleanup before. I downloaded the demo version and it seems to be very good in most instances but because I don't have a license it adds very heavy grain/watermark over the images so it makes it hard to judge properly, let alone in motion. If I get a license or a proper trial I'll do a video where I compare results. Thanks for watching and for pointing me in the direction of this tool!
Oh man, thats unfortunate. It probably hit the memory limit of your gpu. It’s strange that it crashed your whole computer though. Maybe try again or send a report to the foundry.
Compared to solutions like fooocus, generative fill and etc. it looks relatively outdated and very rough, almost useless in terms of quality as for image inpaint (as expected this tool was built on 2021 research). Does it have some benefits in terms of sequence inpaint or you haven't tested yet?
You are right, this is by no means the most advanced solution out there... not even close. The advantage here is the fact that you don't have to leave nuke and it is just one more tool/option in a comper's arsenal. I worked at one of the large studios for almost a decade in both film and commercials and I know for a fact that you dont get access to photoshop most times, let alone stuff like comfyui / stable diffusion like tools. As for temporal consistency, the foundry's website states: "LaMa is not temporally consistent, in the example video smart vectors were used to propagate the in-painted area."
@@alexvillabon Yeah I also tested it as soon as I wrote the comment, and it seems to not respect the contents of the mask, only the surrounding areas. It could still be usefull to export the mask and video to comfyUI and test it with the inpainting there with crop and stich nodes. Get a better quality maybe for large scale matte paintings and cleanups. And maybe AnimateDiff could help with the temporal consistency even.
Im really really happy with the stuff you show us here, this is the kind of content me and many are looking for! In this case with this new lama node, would be interesting to compare instead of NNCleanup, it is so powerful and have solved me some things that otherwise would take days and weeks, even on moving footage instead of single frame.
Thanks Alex, this is great!
Thanks for the kind words, I'm happy you are finding value in my videos so far!
I had never heard of NNCleanup before. I downloaded the demo version and it seems to be very good in most instances but because I don't have a license it adds very heavy grain/watermark over the images so it makes it hard to judge properly, let alone in motion. If I get a license or a proper trial I'll do a video where I compare results. Thanks for watching and for pointing me in the direction of this tool!
Great vid Alex! Pretty soon I’m not going to be able to keep all these names straight. LaMa, ABME, MiDaS, RAFT, TecoGAN…
Ha! Absolutely, add to that the endless amount of comfyui models and loras… its tough to keep track!
That LaMa result is very useable as a first pass. Much better than photoshop generative fill.
another great tutorial Senor Villabon!
Thanks Sam! :)
i see this in beauty retouches
RTX 4070 TI, LaMa completely crashed my computer when I looked through the node with mask added.
Oh man, thats unfortunate. It probably hit the memory limit of your gpu. It’s strange that it crashed your whole computer though. Maybe try again or send a report to the foundry.
Compared to solutions like fooocus, generative fill and etc. it looks relatively outdated and very rough, almost useless in terms of quality as for image inpaint (as expected this tool was built on 2021 research). Does it have some benefits in terms of sequence inpaint or you haven't tested yet?
You are right, this is by no means the most advanced solution out there... not even close. The advantage here is the fact that you don't have to leave nuke and it is just one more tool/option in a comper's arsenal. I worked at one of the large studios for almost a decade in both film and commercials and I know for a fact that you dont get access to photoshop most times, let alone stuff like comfyui / stable diffusion like tools. As for temporal consistency, the foundry's website states: "LaMa is not temporally consistent, in the example video smart vectors were used to propagate the in-painted area."
For something like the runner it would work better if it could work temporally.
That’s the natural evolution. So far no model does that but im sure its a matter of time.
Have you tried using the inpaint first and then add the Lama? In ComfyUI we use it for inpainting that way. Maybe it makes a cleaner plate
Interesting thought. Unfortunately, I just tested it and it gives the exact same result.
@@alexvillabon Yeah I also tested it as soon as I wrote the comment, and it seems to not respect the contents of the mask, only the surrounding areas. It could still be usefull to export the mask and video to comfyUI and test it with the inpainting there with crop and stich nodes. Get a better quality maybe for large scale matte paintings and cleanups. And maybe AnimateDiff could help with the temporal consistency even.
Is Apple Silicon processor needed as well? I see it mentioned on the download page but it's not mentioned here in the video. Thanks!
Nevermind - answered my own question - CRASH! haha
:(
Awesome video Alex!
Glenn! Thanks :)
I was realy wonder this ai model.I think seems a lot not practical.Maybe I will use some easy shots.Thaks for the video.