home.social

Search

86 results for “snoopj”

  1. CW: letters from the dating front

    it's fun that games are a subject we're both interested in. she's just finished the TV show she was watching and was musing about what game she'll play now that she has her evenings free

    NieR:Automata is in the running since I have already gushed about it to her :blobfoxlaughsweat:

    I have dated women who play games before, but I've never dated a Certified Gamer™ who has a relationship to them that I find personally recognizable. The prospects are titillating!

    hachyderm.io/@SnoopJ/117651681

  2. perhaps unsurprisingly, the model seems to be *very* good at segmenting cats

  3. I guess I'm using the updated version of darktable now because updating my database to work with 5.6 means I can't run the older version from the Debian repos

    oops.

  4. oh, you can save the SAM mask as a raster image (in a mysterious location that apparently defaults to my home directory, lol)

    so I guess there's the possibility for a workflow here where you start in darktable, fiddle with the mask generation as much as you're willing to put up with, save as raster, go into your image editor of choice to tidy it up (along with the original photo which I guess you'll have to produce your own copy of if that editor doesn't do RAW?), and then pull the edited raster mask back into darktable to do whatever you wanted in the first place

  5. I dunno, I'm just playing around with it.

  6. here's a more typical example where I might want to use it, but drawing these masks by hand would be too laborious

    it's too slow on *this* system but I can see how it would be something I'd be willing to do on a system with a real GPU that could run this model faster, and provides better results than just stamping a circle mask over everybody's face and calling that good enough

  7. anyway. I *might* use this in cases where the automatic mask works sufficiently well and I want to do some kind of foreground-background separation effect (most likely using the "censorize" module)

  8. there is of course some plugin for GNU/IMP that provides SAM but of course installing it is a huge pain in the ass that involves setting up PyTorch blah blah

  9. like okay yes you can go in there and piss about with every individual vertex and nobody in their right mind is going to do that

    it makes me think that maybe a better workflow for regularly working with really fiddly masks like this challenging scene would be to use SAM in GNU/IMP to get a good starting point and paint the rest of the mask, then export the whole thing as a raster image and bring that back into darktable

  10. as a darktable neophyte it's unclear to me how one might develop a workflow for refining these masks. definitely the 80-20 rule at work here if you want your mask to be *good* but the model cannot one-shot it

  11. I'm not entirely sure how SAM works but darktable's presentation allows you to click regions that should or should NOT be part of the mask. it does something not-unreasonable for this challenging scene with a single inclusion click, and with one more rejection click it creates a pretty darned good mask in a torture-test scene

  12. darktable is no longer chugging on this system, btw.

    might have been a network issue after all

  13. although meta is still a fundamentally evil organization and this model was created using all the same scummy practices as the more broadly discussed ones, &c.

  14. O(100M) is at least a mature-adult number of parameters that has acceptable performance even on this power-sipping system.

    When the standard craziness of the day is O(1T) parameters, O(100M) looks downright compact.

  15. it works pretty well, at any rate. this is a result of SAM-2 with 4 clicks (one for the hull and each bit of the sails)

  16. huh, newer versions of darktable have integration with Segment Anything

    mixed feelings about this

  17. at this point my only criticism of Hell Wait is that it's so good that I haven't been playing Mahjong Soul lol

  18. The joker that boosts fu for the number of enhanced tiles you have in your deck is pretty good. I think it's got the same thing going on as in Balatro where at a certain point in the mid-term, you get pretty good returns from boosting the *left* side of the multiplicative expression.

    but of course, in both games, exponents dominate. you need as much exponentiation as possible to keep up with the geometric growth. Excalibur may not be as powerful as I'd thought since it's "only" 1.25x, but it does work lovely if you can make a lot of your wind/kuro tiles stone since the nuki stack is scored last (and therefore the mult there is always favorable).

  19. the tiger stripe joker (+0.25x on pattern yaku) is pretty powerful as long as you can reliably target pattern yaku. for this run that was pretty much always iipeikou and I probably would have benefited from taking the joker that guarantees your hand is always closed, since iipeikou requires a closed hand

    still, pretty good run. as expected there is an achievement for high round numbers in endless mode, but it's at round 25 not round 20 as I'd expected (the normal game ends after 12 rounds)

  20. two discoveries:

    1) once you get into 100B+ territory the game starts abbreviating the suffixes

    2) you can stack nuki'd tiles high enough to block the target score display lol

  21. ah shoot GODZILLA MINUS ZERO is at the end of its run and I still haven't gotten out to see it

    looks like evening showings are done with, but there are still a few mid-day showings left

  22. thinking about fiscal multipliers again

  23. the projected cost of eRHIC for instance is yet another order of magnitude down

    aip.org/fyi/2020/brookhaven-pi

  24. the total cost of a polarized electron-ion collider would be even less and would yield much more useful physics IMO but that's getting into deep-cur territory so I tried to keep this tied to CERN/LHC since those are more familiar to the general public as "unfathomably large scientific megaproject"

  25. (although comparing an electron machine and a hadron machine in this way is not exactly reasonable, but let's not talk about FCC-hh because I have Some Opinions™ okay)

  26. the projected cost of construction for FCC-ee, the least-unreasonable "let's dream as big as we can" concept for a new CERN accelerator much larger than LHC, is like $20bn

  27. like yes if you throw an unfathomable amount of money at brute force, you *will* see results

    you're probably not getting very good returns on your investment though

  28. RE: mastodon.social/@marginalia/11

    For comparison's sake, the total lifetime cost of the LHC is more like on the order of $10bn

    so

  29. number — and I cannot stress this enough — go up

Share on Mastodon

Enter the server where you have an account.