[QOL] suggestions

[hexw1n2]

I love sammie roto but the workflow is really annoying. I think these few small changes could make it a lot easier to use
1: ability to choose where it downloads to automatically and with a name (or auto name such as maybe the name of the source file + ROTO)
2: Ability to toggle auto download after export
3 (this is harder to implement than the others) : Batch processing. For example, the user can input multiple videos that he wants to roto, then he can set the points for each video. Then he can choose to track all videos, then export, then download which will be done automatically without user input. I am not sure exactly how it works but the most viable way i could imagine this working is maybe it would use the prexisting point files system and it will store them somewhere, then maybe there will be a temp processing folder to avoid running out of ram. I don't really know if this is possible but it would be very cool

Thank you for continuous development, this is my favorite repository right next to copyparty. <3

[Danny-707]

It would also be great to add a zoom feature and the ability to play the sequence as a video :)

[ncognico]

Couldn't agree more on the tool and the need to evolve the workflow. The naming should be editable and/or ideally could use a token system that automates it. "SourceFileName_SammieRoto_LayerName_Date_Time" would simplify so much version tracking we do. I find it easier to do single runs rather than building layers within the tool because it create so much extra variables to track.

DogRunning_SammieRoto_All_20250810_1300
DogRunning_SammieRoto_Dog1_20250810_1300
DogRunning_SammieRoto_Dog2_20250810_1300

Additionally if we could save out as other formats or as a zip with image sequences rather than solely an mp4 since the container seems to break once you cross the 4K resolution territory. Comparing this to a web based app like slapshot.ai you can see how much easier it would be to keep track of and organize the files you get back (EXR). For a non web-based comparison, ComfyUI allows a lot more control and workflows to achieve the result but is far more computationally heavy by comparison so any way to queue, save selections or frames/shots/projects would also really push this tool into another level.

One other thought I had while batch processing was would there be a way to automate a basic run with a sensitivity selection? If we could add shots into a watch folder and receive back 3-5+ mattes I would run entire dailies through this system. It would be a quick way to get basic mattes for all footage slowly but without an operator and if the selections still weren't good it would at least inform you what the system is picking up well or not and where to focus.

Maybe it's just my hardware but I've also run into issues where the tool hangs on random frames during processing and the only way to get it to continue is to hit enter on the command prompt. This is consistent between MacOS and PC. Any way to avoid needing to babysit would be much appreciated. Thank you

[Zarxrax]

I am currently evaluating rewriting the gui with QT to allow for more flexibility and features. No promises on when or if that will be completed, but it would allow for more flexibility with file saving as well as zooming. Furthermore, it seems that SAM3 might be released later this year, so maybe a new interface will be able to coincide with integrating sam3 into the application.

@ncognico could you elaborate a bit more on what types of projects and workflows you are using this tool for?

[ncognico]

@Zarxrax I am a VFX artist for feature film, episodic TV and commercial advertising. I use it in tandem and as an alternative to Resolve's Magic Mask v2, Adobe's rotobrush, and other MatAnyone/Segement Anything toolsets. I want speed and accuracy especially with fine details like hair, and SammieRoto provides both but has room to grow. It feels like a true hybrid of Chromakeying and rotoscoping. If you want more details I'd want to connect privately.

[Zarxrax]

@ncognico @hexw1n2
The newest release adds some improvements related to file naming, kindly added by @Mark-Shun. There are some limitations imposed by the Gradio user interface that is currently being used, so we limited with what can be done related to filenames.

The other suggestions will be kept in mind for a future release.

[ncognico]

Thank you very much @Zarxrax & @Mark-Shun I will be testing this out moving forward.

Side note - can you think of a way/or add a way in the future to circumvent the frame extraction process? If I could pre-extract on my end and provide a path + auto download it would save 30-40% of the current workflow time. I understand there are limitations but appreciate you working in more functionality.

If it is insightful for anyone, below is my current workflow. Any optimization that helps this process would be amazing.

1. Pre-process my footage into mp4 h264 25 mbps CBR Same as source resolution - this saves tons of time on the extract process and for b&w mattes the extra data the raw footage provides is negligible. Sometimes I will trim the clips but on a 3090/4090/M2 Max the speed is great so I process entire clips. If I shorten clips I do find it harder to realign later if they have different frame counts. (hence why naming was so important)
2. load to sammieroto
3. wait for extraction
4. make selection, single frame, usually single matte to avoid letting multiple mattes run without being able to review/tweak between.
5. matting tab -> run matting
6. let it cook
7. export (wait)
8. download
9. rename based on source file and add "roto_content_v1" so "DogRunning_roto_head_v1"

Being able to queue or do the above in fewer steps will continue to push this tool into production.

This update and any future ones are very much appreciated!

[Zarxrax]

@ncognico You mentioned EXR format a little while ago, and I'm a bit curious what you would expect that output to look like. I have never used the EXR format before, so I am only vaguely familiar with it. Would each object from sammie-roto get stored as a different alpha channel, or something like that? Would you want the original image stored in the file as well, or only the alpha channels?
If I am able to add support for it, it probably would not be very flexible with a lot of options available, so I just want to make sure I nail down the requirements correctly.

[ncognico]

@Zarxrax Yes Multi-channel OpenEXR or DPX would allow storing individual objects as separate channels within one image file, whether or not the individual channels have alpha included would depend on intent but they each can contain a separate alpha. It is the standard in VFX due to the 32 bit support and because it allows floating point images without the need to convert or compress into a log format. Learn more here. This is this same case for AOVs in Maya and render passes in 3D CG as well, where you can create a single render that has your passes (beauty/reflection/roughness/etc) embedded within to avoid rendering each one for control later. Every compositing tool out there has a way to cycle through the channels like After Effects you can use an effect called "EXtractoR" to isolate which passes you view. Nuke calls it shuffle. It is demonstrated ~2:50 in this basic Resolve tutorial. Having the option to save out an image sequence (JPG/PNG/TIFF/EXR) would generally be amazing as this is the standard practice for rotoscoping houses that perform their work inside Nuke or Silhouette.

On a basic level you could save each object to its own channel, thus simplifying the number of files needed to export all of the layers on the back end. We could grab all of the work from a shot as a zip(?) or image stringout and do the rest of the work inside our compositor of choice. The current workflow with mp4s gets the job done in the right circumstance though, so that's why I lean towards it being a togglable option.

On a more complex level you could embed any number of useful data in the channels similarly to how a VFX vendor provides roto back. Usually a studio will submit a single frame red pen and the source clip with some handles on either end. The vendor provides back a package of assorted data including black and white image string outs (EXR, TIF, PNG, JPG), nuke shapes, silhouette splines, hard mattes, soft mattes, motion blur and no blur mattes, hair passes, core mattes, etc. It can get deep quickly but having the ability to export multiple passes within one file would be incredibly powerful for the long term development. To answer your question on the original image being stored within, yes, it would be great to have the ability to save the the source plate within but I do question how much bloat it would create. I usually will convert video files to a same as source mp4 at 25 mbps CBR h264 just so the image extraction process doesn't balloon a camera clip into hundreds of gigabytes. (which I could also avoid if I could point SR to an image stringout rather than just a single clip or image). I don't fully get how this system works so you'd have to speak to what other types of data you could potentially extract to be used in VFX but it feels like the sky is the limit. If you're able to save the original image back into the EXR you might even be able to get fancy and have the alpha from the object apply to the original image as an additional channel. Not necessary at all but it would bring composite functionality to this similarly to how Mocha became more than just a tracking software,

Side note, latest release is working great. Love the naming features and performance seems improved. A few small thoughts:

1. If you do go for image sequence support I highly recommend following standard VFX naming so four digit frame count, start at 1001 for frame 1 and follow the template: "FILENAME_LAYERNAME_####.EXR" with PIZ compression (if relevant). "SammieRoto_DogVideo_DogRunning_1001.exr"
2. I would love a field for custom naming, or the ability to rename the objects/layers so that we can be as specific as possible. I prefer to specify what I am isolating and sometimes will make multiple passes of each person/object using different frames as a starting point so it helps when I can add "Actress1_v1". Perhaps another checkbox that when desired allows quick tweaking of the generated name or so that the final renaming happens within the webUI. I know this is picky but I find myself passing versions to others and will still usually append a "_v#" to make it easy to refer to. I normally apply this after I download the file but it would save a few seconds off the workflow to have another means of doing it.
3. The date and time is wonderful, but it would be a bit more intuitive if the time didn't include the seconds (or could be turned on and off). There are certainly situations where I would want that level of accuracy with my renders but every shot I do usually takes a few minutes anyways so the time could just be "Hour_Minute" / "2330". The extra info gained with those 2 digits seems less than the potential confusion when looking for sequences of numbers.

Let me know you have any other questions, thanks again!