Matanyone using sammiroto

[Bryanrosenblum]

So the results out of Sammieroto using the matte alpha not the segmentation matte would be the same as if you used matteanyone? is sammieroto using matteanyone 2 yet?

[joelosis]

Correct, sammie 2 isn’t even out yet officially and the person behind sammi would probably need time to implement into Sammie.

[Stefan]

@Zarxrax

…port for running on mac or even just a cpu mode. It has weird dependency issues making installation difficult. Also, the main focus with sam3 is on masking via text, which is not something I am particularly interested in. I am more concerned with seeing how it performs against sam2 in standard masking and tracking via points, vram requirements, etc. I haven’t seem many people talking about that yet.

Makes a lot of sense, thanks, @Zarxrax !
Are you planning on integrating matAnyone2?
Sammie-roto is so much better than straight matAnyone in comfyui …
When I was asked to try in comfyui, all the shots had the person not in frame at the beginning and end of the shots …
It looks (from the video above) that there might be better temporal stability in semi transparent areas (hairs …)?

[incognico]

@Bryanrosenblum It’s brand new and unreleased, but the plan (per @Zarxrax on discord) is to try to implement it. Seems more likely to work than SAM3, but it’s all speculation.

[Zarxrax]

MatAnyone2 is not released yet, but when it is, I will definitely be looking at implementing it into Sammie-Roto. Just don’t expect it immediately, as it will probably take me some time.

Regarding SAM3, I am currently not planning to support it, because the results from testing with point prompts was not promising. The key feature of SAM3 is just being able to segment with text prompts, and I don’t really see that as a useful feature for Sammie-Roto.

[randy]

Installation instructions for Sammie Roto 2 on ADSK Rocky 9.3 with DKU 20.2.1 and RTX 6000 Pro Blackwell / Cuda 12.8.

## Sammie Roto 2 installation

# Update and install Python 3.11 and development headers

sudo dnf install python3.11 python3.11-devel -y

# Install dependencies for OpenCV, Qt, and Video processing

sudo dnf install mesa-libGL mesa-libGL-devel libX11-devel ffmpeg -y

mkdir -p /scratch/workspace/GitHub

cd /scratch/workspace/GitHub

git clone https://github.com/Zarxrax/Sammie-Roto-2.git

cd Sammie-Roto-2

/usr/bin/python3.11 -m venv venv

# Activate your new environment

source venv/bin/activate

# Upgrade pip inside the venv

pip install --upgrade pip

# Use the provided script but force it to use your active environment

# We do this by temporarily aliasing 'python3' to your venv's python

alias python3=python

bash install_dependencies.sh

sudo dnf install xcb-util-cursor -y

cd /scratch/workspace/GitHub/Sammie-Roto-2

source venv/bin/activate

python sammie_main.py

[Jerem]

@wiltonmts

…Hey everyone! I just wanted to share a quick update I made to the latest version of the script. I added a simple user interface so you can adjust the round-trip settings directly from Flame.

Hi @wiltonmts thanks for this. The script is not available anymore, can you have a look ?

Thank you !

[wiltonmts]

Hey @Jerem , I took a look and it turns out I accidentally removed that file while organizing some things on my google drive — I’ve just put it back up. Recently, I made a nice update to this script to work as a file scan system to improve the round-trip of results from Sammie and other matte exports. If you want to grab the updated version, here’s the file:

https://drive.google.com/file/d/1MaI1_6doysXVM0oK305NZK57q6Wqd1di/view?usp=drive_link

Main updates on SammieRoto RoundTrip:

→ Path history and more token options in Settings — save and quickly switch between configurations for different projects, with standardized paths using Flame’s native tokens

→ Smart Import — scans the result folder, shows info for each file and lets you choose what to import

→ Group Mattes — automatically chains mattes through Comp nodes with your chosen blend mode

→ Multiple Groups with Compass — create several independent matte groups in the same Batch, each visually organized in its own Compass in the schematic

→ Persistent Groups — close and reopen the dialog, groups are still there. Add new mattes without rebuilding anything

As a bonus, there’s also another file included which is a matchbox shader I made and have been testing — it’s pretty cool (a version similar to JB_ocean).

🔥 SammieRoto RoundTrip for Autodesk Flame - Update

[kily]

Tested and working 100%. Really great job.

[cristhiancordoba]

@cristhiancordoba

If you guys were happy with MatAnyone…prepare your heart for MayAnyone 2 which is just announced.

For Anyone interested.
MatAnyone 2 just released.

You can try and use it on SammieRoto by just downloading the model and replace it at the checkpoints folder and rename it as MatAnyone.pth (not MatAnyone2.pth) as the original code from SammyRoto is looking for MatAnyone.pth file.

Just did a comparison with these two models using this method, and tested with a horse shot and the result is dramatically improved. Way better.

Release MatAnyone 2 · pq-yang/MatAnyone2

[theoman]

Hi,

i have been following the install suggestions but getting stuck on the python version. Ive used Venv and Anaconda but when i try the bash install_dependencies.sh it doesnt see my later versions of Python installed.

sudo dnf install python3.11 python3.11-devel -y

i have python 3.11 installed but following @randy install instructions using alias python3=python

(venv) [tmaniatis@vxfhost Sammie-Roto-2]$ bash install_dependencies.sh
Python 3.10 or higher is required. Detected version: 3.6

Why is it only looking at my system Python version on 3.6 and not the higher 3.11 or 3.13 i have

[tmaniatis@vxfhost Sammie-Roto-2]$ /usr/bin/python3.11 -m venv venv
[tmaniatis@vxfhost Sammie-Roto-2]$ source venv/bin/activate

When i use Anaconda i dont get this error but i get a CV2 module missing and cant seem to install that by: pip install opencv-python

so i get stuck there also.

i have activated the 3.11 in venv so now im a little lost here on how to fix. Im on rocky 8.7

both methods for me is failing currently, anyone have an insight orshould i install a later version of Rocky here?

Theo

[theoman]

to follow up and close out my issues, after 3 days of trying to install Sammiroto including getting an AI agent to help problem solve, just install Sammiroto 2 instead. it worked in 3 minutes flat.

[marios]

How would you rate this against Flame’s Automatte in 2026.?

[wiltonmts]

Lately, I’ve been working on some possible updates to contribute to the SammieRoto project, and I just submitted them to Zarxrax’s original repository for review. One of the additions is the VideoMaMa model; although it’s heavier than MatAnyone 2, it offers better temporal consistency. I also added the Corridor Key model, which runs on smaller GPUs—I used a pipeline that runs on approximately 8GB of VRAM. I haven’t had the time to test it extensively or with very high resolutions yet, but it’s been running well on my end.

Here is the link to the pull request I made:
https://github.com/Zarxrax/Sammie-Roto-2/issues/42

[philm]

@wiltonmts - it’s like one of those jurassic something movies: it’s all quiet while you’re taking a leak and then the trees explode and there are giant monsters all around!. Good work brother.

[kyleobley]

That’s a great addition @wiltonmts! Out of curiosity, does your implementation of Corridor Key include MLX support for MacOS like this one does?

Having access to these models/approaches in their own little environment is super handy but I’m starting to wonder if it’s better to keep models in a single location and then link to them within each environment. Having the VideoMaMa which weighs in at ~37GB in 3 different places, for example, starts feeling a bit silly. I suppose the answer would be to just do everything in Comfy but that takes away the easy of these stand alone approaches.

[cnoellert]

@kyleobley

That’s a great addition @wiltonmts! Out of curiosity, does your implementation of Corridor Key include MLX support for MacOS like this one does?

I’ve been working on an MLX port. My thought was to go Pybox. It’s “fairly” performant.. 3sec/frame at 2k on an M4 Max.

[wiltonmts]

Hey @kyleobley, good point!

I ended up tweaking the scripts just for my Linux environment and didn’t get around to Mac adjustments, which is exactly why I made a pull request to the official repo so they can review and make the necessary tweaks for other platforms like Windows and Mac.
I can take a look at the EZ-CorridorKey pipeline and suggest MLX support, though I don’t have a Mac here to test it.

On the model downloading question, it’s definitely relevant but also depends on each user’s pipeline – good point to keep in mind.

The VideoMaMa version I’m using is around 19 GB because I use “allow_patterns” to avoid downloading extra unnecessary files from the repo.

[carloscampos]

Hi everyone,

Is there a way to use a custom mask with Sammie-Roto for object removal?

Sometimes the internally generated alpha doesn’t isolate the object I want to remove. It would be great to be able to feed a custom matte or mask to guide the erase.

Any workaround for this?

Thanks!

[Stefan]

@kyleobley

…their own little environment is super handy but I’m starting to wonder if it’s better to keep models in a single location and then link to them within each environment. Having the VideoMaMa which weighs in at ~37GB in 3 different places, for example, starts feeling a bit silly. I suppose the answer would be to just do everything in Comfy but that takes away the easy of these stand alone approaches.

100% this makes a lot of sense.
Duplicated models on flame production machines is quite ugly. For Comfy we use server centralized (maintained) collections and it seems perfectly fine and fast. An option in sammie-roto would be great Alan @Zarxrax!
Btw Alan, out of about 30+ in house flame hooks, Sammie-roto was the most used in the last 2 weeks or so on our 20 ish flames.
Are you planning on implementing MatAnyone v2 soon? No pressure man