• Content

  • AI Scripts updates and requests

Rakly3 Skipped strokes is usually not the issue in the AI script (except for one part, I'll mention it later). It's the stroke range and I can say the cap is definitely too restrictive and as the handy can handle much more than that. Just compare some of the funscript heatmaps below of some good SLR scripts that feature similar intense parts as back in to your arms. They are way more intense and play perfectly fine. Also Notice how the range is often maxed to the bottom (which makes sense if a girl slaps into you).

For reference, back in your arms:

Screwing out of school (https://www.sexlikereal.com/scenes/screwing-out-of-school-project-32655) (especially notice the difference in the fast parts of the standing position)

Resting roommates (https://www.sexlikereal.com/scenes/resting-roommates-22501)

So I took another look and made some notes for you and added some semi rough timestamps. This is not everything but I think it brings the point across:

  • 06:04 - The AI only seems to track up and down movements, blowjob segment was good overall though
  • 18:12 - The guy is fucking her hard and the intensity (stroke range) is way too limited
  • 24:24 - Maximum pounding but barely any stimulation in the script is (it's actually yellow)
  • 25:16 - No visible up and down movement but griding so too little intensity (might also be a perspective thing)
  • 25:35 - Visible movement here with low intensity so it's okay
  • 32:37 - Doggy works well since it's a slow part and good visibility
  • 38:46 - Skipped strokes, intensity way too low (perspective issue for the AI?)
  • 39:32 - Even better example as the above. One of the worst parts.
  • 42:56 - Jerk of finish, stroke range too low and at the bottom instead of top

So, to summarize some parts are good. Mostly the parts that are slow and clearly visible and in the right perspective. Fast parts and especially those that are less visible and in different perspective struggle to get a good result. Now I get that perspective and the obstruction of view can be difficult but the range should at least be limited way less.

Also, I don't know if you already do this and don't want to tell you how to do things but if I might make a suggestion: It might be a good idea to make a movie that has various common positions in different intensities (especially with the parts the AI struggles with) from movies that are scripted by someone like Realcumber. Than each time you iterate the AI you could just analyze this movie and compare the AI output with the manual script. Like with some test automation script that checks stroke range and intensity and visually using the heatmaps. The reason for me to mention this is that I think it is pretty obvious that there is something going on with the intensity by just by looking at the heatmap. For instance it's very rare for a funscript to only have 3 slightly orange parts and I think it has never been the case for a manual script made for an SLR Original.

Anyways, hope that this will help you crack the AI funscript thing!

      fenderwq Also Notice how the range is often maxed to the bottom (which makes sense if a girl slaps into you).

      This one is something we have been working on with varying results. The issue often is that, while trying to correct one thing, it 'breaks' another. The 'not reaching bottom' I think is a major problem point, partially causing the strokes being too shallow. - We will certainly take your detailed feedback in to our evaluation!

      I do want to mention that we can not use the scripts hand made by scripters just like that, to train the AI. I don't think it's ethical and may be legally questionable.

      The feedback is much appreciated and keep 'em coming! 👍️

          Rakly3 Glad to help and yeah I know how one thing can effect the other thing all to well as a tech lead of a dev team myself 🙂. The main reason I mentioned the bottom range maxing out is that it will effectively mean a larger average stroke range. In other words if you cannot max out the bottom You'll have to increase the range to get a similar sensation which might be difficult when the range is already large (I'm sure you are aware of this though).

          As for the test movie: I get that you can't just use any script for training but you wouldn't train an AI model to generate Anime images without Anime in the training set either right? In other words scripters like Realcumber set a benchmark for what users want and come to expect. Now I think Realcumber is a level of quality we won't see in AI scripts for quite a long time but a script AI needs to at least understood some of the basics that make these scripts good and I think the stroke range (and thus overall intensity) is one of these essential things. Maybe Realcumber would be OK to provide a few parts of his scripts for a semi automated test that has a very varied amount of positions and intensities? This way you wouldn't even train your AI directly on the script, you would just have a tool to automate some of your quality control for the fundamentals of your script AI. This would also allow you to plot your progress for various versions on different key factors such as intensity for a varied amount of positions and intensities. I think this will take a lot of the guesswork out of the process. Just a thought though.

              fenderwq Hi, thank you for detailed feedback. I checked problems at timestamps you listed. It was helpfull. I prepared version of "Back in Your Arms" with better tuned larger range. Large range "Back in Your Arms"

              We actually have automatic testing of our ai-scripting through comparing ai with Realcumber, we measure metrics of quality (like f-score, precission, recall, L2) as usual in Machine Learning projects The problem is umbiguity of video: there are several possible movements for single video interval, especially when penis is not visible. And detailed movements like at 06:04 you mentioned is hard problem due to lack of data of such kind.

                  fenderwq It is pleasure getting feedback from technical guy) We actually train our models on manually created scripts.
                  Keep in touch.

                      siarhei Thanks, will check out the script.

                      Seems you already have a nice automated test setup. However in this case I was referring more to the averages of a segment. If you have a movie with various segments of a somewhat consistent action / context (like a certain position and intensity) you can compare aggregated metrics for these parts of the movie. Things like: total strokes, average stroke range, average stroke position and stroke variability. I noticed big overall differences in these metrics between AI scripts and human made scripts and the difference varies greatly by position and intensity (that's why I pointed out the suggestion).

                      Now, I've already pointed out some of the specifics (like the overall intensity of mostly fast parts, some of the "perspective" issues and the bottom range issue) but another important one is the stroke variability during fast parts. I first noticed this when trying out the script for long night. It felt all over the place in the fast parts and when you analyze the script it becomes pretty obvious why. I haven't compared this in detail (as in strokes per minute etc.) but it wouldn't surprise me if it is this high variability and not the actual speed that made these parts uncomfortable. I just can't imagine that such an irregular pattern could feel good at high speeds.

                      AI generated fast part (long night)

                      For comparison: a fast part of a manual script (screwing out of school)

                      Anyways, I think it's these kind of things an automated test like I described (in combination with comparing the heatmaps) could catch early on and will allow you to better measure progress on between iterations. But I will shut up about it now 🙂

                      Good luck with the project!I

                          Rakly3

                          Rakly3 I do want to mention that we can not use the scripts hand made by scripters just like that, to train the AI. I don't think it's ethical and may be legally questionable.

                          I figured that is how you were teaching the AI with human scripts. Isn't that how AI learning works?
                          If you don't use human scripts as a teacher then how does the AI learn?

                                fenderwq The graft of the AI script is exactly how it feels. It's not smooth, it's big stroke, little stroke, big stroke, little stroke. It doesn't feel natural.

                                    siarhei I've checked out the new script for a bit (mostly the faster parts) and it is definitely an improvement. Having said that a lot of my feedback from my previous post still remains in various degrees (the one with the timestamps).

                                    Some additional notes:

                                    • 23:46 - The script is better here but I think still quite a lot more limited compared to a manual scripts (it should be clear purple in the heatmap here). The movement is also way to high up (like in "blowjob position") and since she is on top of him slamming into him it should be at the bottom (is this due to the limiter not knowing the position? Btw, I was thinking it might be a cool idea to have a refiner AI that does a second pass to determine things like if the script should be limited to the bottom or top and maybe smooth out the motion on fast parts intelligently / based on the scene).
                                    • 24:25 - Still too little movement but definitely better
                                    • 38:31 - Some of the obstructed parts still have very little or no movement

                                    Movement in the fast parts is also pretty erratic, for instance:

                                        rerun119 Yeah, the fast parts definitely are, I think the slow/medium parts are often good though.

                                          fenderwq How are they training the AI if they don't use human scripts? Isn't that how AI learning works?

                                            Aaaand a new update has arrived! Here are new scripted scenes:

                                            We appreciate all the feedback you guys are giving, it helps a lot with the development. Hopefully, you can see the improvement.

                                            Also @Rakly3, care to share your thoughts on the scripts?

                                              Kieshi For objectivity reasons, not at this time.

                                                @Rakly3 just curious if the previously released scripts will be updated with the new AI programming as it improves? Will it only be at a later time? Or will the new programming be added to these scenes continously as new iterations of the AI come out? There are some great scenes that have the AI scripting and I would love to see the improvements added to them as the machine learning gets better.

                                                  Qwertyqwerty1234 They will eventually be updated, but I don't make this decision. Nor am I currently aware of a timeframe. I myself don't find out which scenes have been AI scripted until after it's done.

                                                    Kieshi Nice work, these scripts are definitely improving. I watched VirtualRealPorn House from start to finish and overall the intensity seems much better. The biggest remaining issue is the irregularity in faster parts. This seems to have improved but still needs work (it can be pretty annoying now). There is also still some problem parts that the AI struggles with mostly due to obstruction or not detecting movement from things like jiggling.

                                                    Some notes:
                                                    13:18 & 14:03: irregular fast movement which should be smooth (didn't make a note of the intensity)
                                                    for example 14:03

                                                    18:21 & 19:30 & 28:24 & 30:37: irregular fast movement which should be smooth, the intensity seems pretty good though
                                                    for example 28:24

                                                    22:24: blowjob while holding the base. Position of movement is at the bottom but should be at top, later it corrects itself
                                                    33:45: One of the examples of jiggling not getting detected

                                                    39.12: This part felt all over the place

                                                      I mentioned this another thread, but it is worth repeating: How about a button to filter out AI scripts? You gave us a toggle to see only AI scripts and filter out human ones, but I'm pretty sure that a majority would like to be able to filter out the AI scripts and only see human ones. A few (angry) commenters have said so explicitly.