Videochat Extension
patreon
Why is the gender filter taking soooo long????
🕑 Added 2025-04-05 18:02:59 +0000 UTCHi everyone,
Over the past few months, many users requested the return of the gender filter from the original extension.
In this post, I'm going to share some thoughts about it—why it takes so long to implement, why it was never a priority, and why it will still most likely be released before the pricing changes take effect for active members.
What was the feature about?
'skip men' has been a feature since v0.1.
This feature used local "AI" to guess the gender of strangers and automatically skip them based on your preferences.
It didn't really change or get much attention after that. While gender-based stats started being counted and the ability to add skipped IPs to the blacklist was added, there haven't been any major changes since v0.1.
ML usage was part of a bigger project that got canceled. The original plan was to use TensorFlow models to detect inappropriate content in video chats to make live streaming safer, but this approach proved to be unreliable and was removed. There was an attempt to reintroduce it later in the extension as a part of the streamer mode, but it again proved unreliable, broke at some point, and was removed altogether.
After the original project got canceled, it was open-sourced, and Videochat Extension was born as a lighter, more consumer-friendly version, where gender detection was introduced as a more practical application of this technology.
However, we never really focused on gender filtering, so its implementation wasn't the best. Still, it seems people didn't mind as long as it worked in some way for them.
Was it a good feature? Not really—it was not very reliable. But since it was optional and didn't load if disabled, running actual AI models in the browser wasn't a performance issue for those who didn't use it.
So where are we now?
Gender detection isn't a big issue.
It's still a tough decision because the models for such functionality will add anywhere from 1MB to 20MB to the extension's size, resulting in A LOT of cancelled installations. What if less than 1% of users actually will use it, yet it will take 90% of the extension's size?
The main problem is that it makes no sense to release it now, as it is dependent on other things that are not ready yet, and not in a very good shape on its own.
No stable monitoring? It cant detect when new calls starting to activate detection.
No good platform for building features? It cant be implemented in a way that won't cause maintenance pain afterwards.
[WE ARE HERE AS OF v3.5.3]
No stable Video Recorder? It cant see.
No blacklist and local database? It cant add skipped IPs to the blacklist to speed up things.
No ability to skip? Guess what, it can't skip.
And while all the necessary elements exist as prototypes, that doesn't mean they're ready for public release.
The Video Recorder captures the stream in a way that requires real-time cropping, and we're unsure of the performance implications for users with low-end hardware attempting to run TensorFlow models on, say, a 2011 laptop. Everything needs to be optimized.
The database is basically ready to be integrated, but designing a robust data structure that can handle all future features alongside hundreds of thousands of entries without scaling issues is a challenge.
The Skipper Windows software is functional, but we still need to develop a lightweight user-friendly macro builder. This is a tough task—no one in the AutoHotkey community has managed to achive that in the last 22 years.
Everyone used different features, and since we haven’t used any kind of product analytics, we don't know the exact numbers. Because we don't get much feedback from you, we will keep refining and updating the extension based on our own vision for the future of the product.
Are there any issues with the gender detection?
There are plenty of adequate solutions available, but tuning them is a challenge.
Outdated models have poor results, while good models are very large in size. Training a custom model is an option but might not be the best route. There are workarounds to improve the accuracy of existing models, so we'll explore those to see if we can make them work effectively.
With these improvements, this feature should be much better than before—more accurate, performant, and stable—although false positives won't be completely eliminated.
A major problem remains: the model only sees what you see, not the full video feed as in the original extension. For example, with ome.tv, mobile users are cropped in half because of design choices in the web version that we are no longer allowed to fix. One potential solution could be to add an option to skip when no face is present for a given amount of time. Idk. It's hard.
We could take the feed directly from the page, but it's much harder to do across all alternatives, and I have been advised against such approach. Overall, it's much better to build solutions that can be applied anywhere rather than focusing on specific sites—even if that might not be the most effective solution for the given problem.
Why it was not a priority?
Because most users will never use it, and it depends on other things that needed to be done first.
VEXT already has a steep learning curve just to get started. Our goal is to avoid webpage modifications, and the workarounds need effort that many users won't put in, either due to laziness or a desire for simplicity. Even learning how to click on the extension icon is a problem for a lot of people.
Additionally, this feature requires an extra, optional permission to function. Google labels the ability to capture the active tab (on which the extension was explicitly activated) as a "Read and change all your data on all websites" permission, which is a very strange way to describe it.
How many people will go through all required steps? Install the extension, figure out how to activate it on the site, accept an additional permission with a suspicious description, press the 'capture' button, crop the video, and finally activate detection? And how many will actually install the separate software to automate skips, register it so that the extension can connect, and then figure out how to use it?
It's likely that only a small fraction of users—probably well below 1%—would manage to cover all those hurdles. The complexity sucks. The UX is terrible.
In recent months, our focus has been on essential, more widely used features and building a solid foundation—without which advanced functions wouldn't work at all. There was a time when I didn't want to invest any effort into the skipping mechanism, but since users kept asking, a workaround was eventually developed.
Overall, though, it would be a cool feature to implement. There are Omegle alternatives that don't use P2P connections, and these new features could make the extension more useful on such sites.
So when will it be released?
If you're still reading, you'll likely be able to figure out how to use it.
We've developed these features as prototypes, but integrating them will take time since the extension didn't have a platform for them until now.
The aim is to release these features during April 2025, before pricing changes take effect for active patrons. We weren't prepared for a surprise attack in December, so we now need to balance new features with contingency projects to prevent similar downtime in the future.
If we had done the same in July 2024, we could have had a fully functional independent product by December 2024. But back then VEXT was an open-source mess, with about 160k active users who collectively contributed around $40 per month in donations, with no vision for monetization since no non-commercial omegle alternative surfaced and we wanted to avoid any conflict of interest with the commercial one.
Omegle's shutdown crushed motivation after I spent months rebuilding the extension from scratch to support it. Improving the user experience on a commercial platform for free felt like an extremely misguided way to express creativity, so the project entered maintenance mode.
After that platform started attacking us using our own users, you can guess how little motivation there were to put any effort into 500-hours project to rebuilt the 4yo codebase for hundreds of thousands of active users who collectively contributed only $40 per month at the time. But that is now in the past. We are now searching for a model that makes more sense, and your recent interest has helped a lot. So thank you. But let tim cook.
All the features we're working on now will be available for free on neutral sites. Patron-only Browser Mode should work perfectly fine with the new features (at your own risk, in theory a human mod can simply notice that you skip without hands and ban you) on all sites, including those blacklisted from Website Mode either due to hostile behavior or at the request of their verified owners.
Right now, Patreon remains the big motivator for working on the extension. That's why your support makes a difference.
The wheels are turning, but it's not a race. Life is a marathon, not a sprint. What? Support our mission!
Writing fewer posts like these (which nobody reads) might speed things up a bit. But it still helps clear my head a little :)
-qrlk
