U can install on an external machine or your flame
Or if you mean does this upload to the cloud or whatever I’m pretty sure it’s all local
2 Likes
Yeah i cant upload anything to clouds really.
Thanks for that
1 Like
Conda is your friend @Quinn. Sandbox all the different environments… otherwise it becomes a shit show fast.
4 Likes
The included install script handled all dependencies for me on RL 9.5. I got running in like 2 minutes, no fuckery.
4 Likes
Cheers @joelosis. Just rolled it out on 5 shots that came at me from out of nowhere and it did great. Had to resize down to 2k width and upscale to avoid oom but it worked a treat all the same.
2 Likes
@Xteve are you going to commit your changes to the GitHub?
Wow, just tried this out via Pinokio(issue w/Samiroto), on my Mac.
It’s very good and pretty fast it seems.
2 Likes
tried to commit the standard comfy MatAnyone version to github today… without success. I can pull, but not commit. I think it’s my proxy config that allows my VSCode to pull but not commit. GitHub is not easy to access from China. Btw, there only 2 files with a slight change of adding mps. I did’nt touch any other code. Only added mps support for apple silicon that I can use it. If you have a mac on hand and wanna try it out, cool. let me know. I can send you the changed files, you can make a diff comparison and try it out.
1 Like
Just remember you can’t track backwards with matanyone, will only matte from the first frame forward you pick from. Hence why I used it in sammieroto as it goes forwards and back. Hopefully they fix it for the next release
3 Likes
Joel, I understand. First and last frame are using perfect mattes. Propagating forwards or backwards is very easy to do. What about we can add in between some key matte frames? This could work even better on details
1 Like
Please send them over. That would be great 
1 Like
I’m honestly ok with using it through sammieroto for the time being, in a perfect world it would be a flame inference node that I just load up, either way im stoked at the results I’m getting so I can’t really complain
3 Likes
Interesting, it splits, generate and propagate between every 10 frames back and forth. thats why the matte results are stunning. I installed sammiroto to it to run with python3.13(bit faster) and torch2.8 on apple silicon. nice to have. Internally it writes all masks (segmentation and final matte) as greyscale png in 8bit to the temp folder inside sammiroto.
3 Likes
Ok I did play around today my whole sunday with sammiroto. It’s working ok. But it works until 1080p max res and about 270 frames before get choked up. Kind of slow if you have more selectives within the same prompt. The limitation of matte generation to 1080p in apple silicon and fixed 8bit internal processing makes it more of an experimental workaround for me. I will not bet on it. MatAnyone in Comfy is faster, can be text prompted OR Sam2 pointed and set & forget delivers exactly the same results in 16bit/32bit EXR. Just in 4k. That’s my personal opinion, so please test it by yourself on your machines.
5 Likes
Damn. Up and running on Rocky 9 in like 3 minutes.
7.5 seconds of 5.6K didn’t take long on a Threadripper 7965wx / RTX 5000 Ada. Maybe 3 minutes?
4 Likes
I am so trying this!! Thanks @joelosis!!!
2 Likes
Very interesting, Joel. I’ve always been a fan of your videos. Sammie-Roto works pretty well for me in HD mode on a 4K footage. The quality is similar to what Inspyrenet with a RemBG node would give (in ComfyUI), but with the advantage of being able to choose your subject. But as soon as I switch to “Full” mode, it becomes super slow (like 90s/frame) and eventually dies throwing a “CUDA out of memory” error. On pretty decent PC (i9-10920X, 12 cores, RTX 4090, 64GB RAM). The test footage is a 136 frames long 4K mp4
HD it is then 
, again the gains I got from full res when it did work was negligible, biggest gains for matte quality I have found to be the source that you feed it
Joel this is amazing
After install, it’s telling me it’s missing a module.. “gradio==5.25.2”
Thoughts anyone…? Do I need to install this from GitHub too?
The batch shell thing should install all dependancies for you no?