Shockingly good tool even at 0.1.0
I tried to find a video that would really confuse it, Eporner Age Verification, and it still did a surprisingly good job on a very large percentage of the video.
1 Like
Ikr, it’s crazy that people are sleeping on this.
I feel like I should just run my whole collection through it 
1 Like
I uh, sorta did that. I have ran around 250GB of video at this point. About 70-75% turned out at least 85% right or better. Tested a lot of them on SR6 of course, the output is more varied than some normal scripts and multifun player actually outputs excellence on connected axises thanks to variation. Nothing has beaten any of my go to handscripted benchmarks but it is damn impressive nonetheless. Some genres (conventional hentai) obviously don’t work, but 2.5D works great and 3D works great. Weirdest thing I have seen is two motion identical videos, just a MINOR costume change on one character, and one outputs great every time… the other is half wrong, every time.
My 5070 is putting in work lol
Also I cannot get
goblinscript.exe --help
to work in terminal, I wish I knew what the last few parameters really do
1 Like
i noticed the same thing with the costume changes. and its weird too. one video the option with more clothes had more exaggerated movement where the other set of videos it was the opposite.
Oh extra whack, because that result was flipped for me lol
Adding a bikini halved the movement
I am still working on this. I have a 0.2.0 model which can handle fast strokes which the current 0.1.0 doesn’t even “see” but otherwise I find it a little underwhelming given that it requires TWICE the processing so a video taking 2 minutes with 0.1.0 would be 4 minutes 0.2.0.
It’s not universally better which is frustrating.
No ETA when a new version will drop training one model takes like 4 hours at this point so it’s very time intensive.
5 Likes
From a user perspective, I would personally take higher accuracy at twice the time. It started way faster than I expected tbh.
4 Likes
also, could i request a feature? i noticed in VR videos i get to pick where i want the model to focus, it would be nice to have that as an option for flat videos as well, i have a hand full of videos where there is obvious motion but it seems like the model just ignores it? im not really sure whats going on behind the scenes but i feel like if we could select where to focus in a clip that would fix it? i dont know i could just be dumb and not have any idea what im talking about hahaha
This is also a good feature idea for videos with more than one male/dildo where one can steal the show incorrectly.
This is amazing, can’t wait to see where this project is going! 
I’ve already added an auto-focus feature for the next version which tries to crop footage on shot bounds. But if the model is ignoring the motion I don’t think that will help. The way it works is by letting the model skim over the video and check if there are regions in the frame the model is more confident about than others. This works pretty well in my experience the benefits are hard to measure but it does improve results on average.
@Its-a-me-A-pathy I don’t want to turn this into a video editor so that sort of thing you’ll have to use a real video editor for. The VR thing is different because that’s not a very standard thing to do but cropping and zooming in a video editor I imagine is pretty is standard.
To be honest I already tried letting claude implement this sort of thing but the user experience is just ass and I don’t want to spend so much time on this.
2 Likes
Incredible tool and so easy to use. Thx a million @gagax123
2 Likes
I added a new version.
This version has one gimmick where if you install winget install -e --id yt-dlp.yt-dlp you can fetch videos through goblinscript using yt-dlp. For the uninitiated yt-dlp can download from a lot of sites not just youtube. This is entirely optional
4 Likes
I’ll give this a go tomorrow on some of the videos I scripted with the previous version. Looking forward to comparing 
I’m not very happy with it to be honest. There’s a defect that got more frequent which is that during slow sections seemingly random phantom fast strokes are inserted. Other times the new model is clearly better. It’s like it got better at fast movement but worse at slow ones…
Don’t expect miracles
5 Likes
I have to say, overall quality of the scripts appears way up. That it does take more time doesn’t bother me at all.
Overall, great work and keep it up.
massive improvements my friend. this down right awesome software please keep up the good work.
if i could ask what improvements/ features are you planning for future updates? i am excited to give them a shot.
well it bothers me because I have to run this constantly to A/B test so these sort of changes slow me down a lot. that being said increasing the processing rate was the only way to recover fast motion.
I think I managed to fix this. I will probably create another v0.3.0 really soon.
I have no list of improvements/features I’m planning
1 Like
Fair enough, i can imagin what a pain it would be having a slow process while iterating/bug-fixing.
It was just my take as a “user/tester” since i value output-quality over speed.
And after having done 6 more or less long videos and a bit of tweaking the quality (esp. love the filler-function in the current version) is outstanding.
2 Likes
Don’t underestimate when someone puts a smile on people’s faces, this UI is a 12 of 10. 
Even if this software would produce nothing at all, it would have been worth the download just for that, thank you!!!
Now, I’ve finally found some time to test it, for now just on 2 hj scenes, and I am pretty impressed, like in holy-shit-awesome impressed.
One of them compared to the manually scripted one - a very close to 100% script, and yeah, the Goblins work gets really close, actually the first tracker/AI-generator result that I consider ‘good enough’ - of course not completely, but with such an result, it saves sooooo much manual work towards a ‘good’ script.
Probably I just didn’t recognize the needed settings, anyway, maybe as a little feedback:
The even parts are a bit problematic, I’ve wrote a super-simple filter, I guess your goal is to teach the model to do it correctly right from the start, but maybe it’s an idea to add a few lines of post-processing code on top in cases where the model follows a pattern in doing something not so well:
This is really simple logic, treat up- and downwards lines a little differently, consider intervals, …
Just an idea.
More important and the main point of this posting: Amazing job! 
3 Likes