Open slabodan opened 1 month ago
but the server says I posted anyway, but the post never shows.
This is old. It happens when you solve the captcha incorrectly, sometimes even when you don't enter anything in the captcha field at all. Probably a black hole for bots. It's pretty annoying, I even wrote a userscript that saves my post so I can just post it again without having to write it again.
Sometimes, when they take advanced measures, every post you try to make in a thread gets this treatment. (Probably useragent based.)
By the way, still getting my old 6 character captchas on my home ISP.
Same. It only gives me the strict CAPTCHAs on my proxies, of which even if I manually solve them, it still either bans me for being a proxy, or tells me I'm rangebanned.
tbh I think they just expanded the rangebanned ips to sell more passes, I also have this issue with my IPs but oddly only when posting on /g/, guess everyone has their share of fun over there
Can confirm, there is a new captcha. No problem posting with the new captcha on any board.
No problem posting with the new captcha on any board.
The solver has huge problems with most of them. But after a while it switches back to the old one. (Experienced with old 4chan cookies and routine dynamic IP change.)
No problem posting with the new captcha on any board.
The solver has huge problems with most of them. But after a while it switches back to the old one. (Experienced with old 4chan cookies and routine dynamic IP change.)
I get a few old captchas but no switch.
can confirm here as well. Especially on high activity shitpost thread like /v/
@Yukariin I hope you're still around. Can we do initial fine tune with 500?
Yeah so much better, got 6/8 correct, lets not close this issue for now, I will test tomorrow and see if I need to get a few hundred more but with this accuracy should be easy. Thanks.
Though I suspect that in the next few days or weeks they are gonna go live with this one https://github.com/drunohazarb/4chan-captcha-solver/issues/28
@gir489returns check if you're actually using the new script. For the amount of new data it has been trained on, it is doing surprisingly well for me so far.
Same, I have the newest version and it still behaves like the old one getting spoofed, I guess it needs more time in the oven
My humble harvest of 100 samples. @Yukariin @drunohazarb 100 samples.zip
here's 5k+
here's 5k+
Thanks a lot! @Yukariin should check it out!
@slabodan try this
This works absolutely perfect for me already After like 50 posts only 1 failed
@Yukariin I have a question, though maybe this place is not the most suited but at least I avoid making more posts, I was wondering how the weights64 in the main code are obtained? Do you just code64 the bin files for the weight or do you go through something else? I am kind of missing this step, cheers. Anyway I technically have a model with 100% success, trained on the 6k captchas sent soo far, the only issue is that it's like 45MB
idk works great for me with only 500 captchas
@Dankan37 depends on your model architecture.
For the current model you want to get the prediction variant, without input labels and CTC layer. The resulting file for the current model is 2.8 MB.
Next, you want to convert it to tfjs format, using tfjs_converter. Something like tensorflowjs_converter --input_format=keras model_pred_new_captcha.keras tfjs_model
. The resulting tfjs_model
is a directory that contains model json and binary file (of the same size). You put the json inmodelJSON
.
For the binary file, you want to encode it via base64, something like this base64 -i group1-shard1of1.bin -o model.txt
. And finally, you put the content of the resulting txt into weights64
From your mention of model size and accuracy, I can assume that your model is larger than the current one and may be slower. Also, 100% accuracy can indicate that the model is overfitting, so I'd test it on a different set of data that hasn't been seen during training.
@Dankan37 depends on your model architecture. For the current model you want to get the prediction variant, without input labels and CTC layer. The resulting file for the current model is 2.8 MB. Next, you want to convert it to tfjs format, using tfjs_converter. Something like
tensorflowjs_converter --input_format=keras model_pred_new_captcha.keras tfjs_model
. The resultingtfjs_model
is a directory that contains model json and binary file (of the same size). You put the json inmodelJSON
. For the binary file, you want to encode it via base64, something like thisbase64 -i group1-shard1of1.bin -o model.txt
. And finally, you put the content of the resulting txt intoweights64
From your mention of model size and accuracy, I can assume that your model is larger than the current one and may be slower. Also, 100% accuracy can indicate that the model is overfitting, so I'd test it on a different set of data that hasn't been seen during training.
Are you gonna update the script with the @moffatman samples?
@JonseyJones yeah, but that bundle still contains some old captchas, so I need to fix my classifier first...
I was just about to test the new model and they changed the slider... The autoslider is fucked up, but the solver itself seems to work fine. f5eba491b2526081a17676a5fbeca877e31f302a
I was just about to test the new model and they changed the slider... The autoslider is fucked up, but the solver itself seems to work fine. f5eba49
What board? Old slider still on /mu/
They changed the captchas again, now they have a lot more noise, also that horizontal line, we may need newer samples...
What board?
I tried on a few blue ones (v/vg/g/jp), but it disappeared quickly and now I get a mix of old captchas and new ones (with cropped letters). It's not new, I think I saw it on red boards a few months ago and the scripts have the code to solve it, but it sucks miserably.
@Dankan37 yep, that's the one I got. It's the new slider they used on /trash, I think.
I'm playing around with it and this and it looks like this newest slider is being solved with the nuBestPos
function
async function slideCaptcha(tfgElement, tbgElement, sliderElement) {
// get data uris for captcha back- and foreground
const tbgUri = tbgElement.style.backgroundImage.slice(5, -2);
const tfgUri = tfgElement.style.backgroundImage.slice(5, -2);
// load foreground (image with holes)
const igd = await getImageDataFromURI(tfgUri);
// load background (image that gets slid)
const sigd = await getImageDataFromURI(tbgUri);
const slideWidth = sigd.width - igd.width;
const opqCol = checkOpaquePixels(igd);
let sliderPos;
if (opqCol === null)
{
console.log("old");
// get array with pixels of foreground
// that we compare to background
const chkArray = getBoundries(igd);
// slide, compare and get best matching position
sliderPos = getBestPos(sigd, chkArray, slideWidth);
}
else {
console.log("new");
sliderPos = nuBestPos(igd, sigd, opqCol, slideWidth);
}
// slide in the UI
sliderElement.value = sliderPos;
sliderElement.dispatchEvent(new Event('input'), { bubbles: true });
return 0 - (sliderPos / 2);
}
I set it to always use the classic slider function and it does more or less okay, got it right 4/4 times
Did some testing, the model can solve the new captchas but it slides in the wrong place. On a proxy with no cookies using the code above: Initial state Solution if I press Solve (doesn't autoslide) Solution if I shift the slider until it's on the exact solution and press solve.
This instead is using the code from the latest Github (.11) Again if I slide into position and press solve:
So hopefully it's only a matter of fixing the slider
Has anyone even seen the old "new" /trash/ slider? I can't get it to show up now
Updated the slider solver. Works much better with new one. Might regress with old ones though. a4dfb109d14d95e4e3825c99e6fea90ccf8eb8c4
Solver is still missing sometimes where I have to coax it by removing 1-2 items
Yeah, the noise from the new slider seems to confuse it.
And they update in real-time kek...
The mods are probably watching this, yet they're so retarded and incompetent that they can't even stop known spammers like ACK. They even let a thread archive with his posts undeleted: https://boards.4chan.org/a/thread/271211044 The mods are (I bet Microsoft doesn't like the N word). They will always lose.
@Yukariin The option to auto-post after get captcha option would be cool, as a feature suggestion, like next to save captcha to have that option, off by default ofc.
For me the script works better than ever.
I guess they added that grid to bg so the surrounding pixels can't be compared for the slider? Just playing around with some chatgpt generated denoising, mby that would work.
Seeing some more of the new captchas, I have to say I find it extremely difficult to solve them manually. Praise Allah for AI.
not sure yet if denoising helps feels like a 50/50 but its basically same without it too
Got the bad slider and the noise and the script poops it's pants.
So many bad captcha results from the script. This is a major problem.
I think I have fixed the slider. The noise still fucks up the solver occasionally though...
Tried removing the grid from the background and just reverting the slider functions back to what they were before and I'm getting ok results, but I have to run a python server where I remove the grid and idk what would happen if a captcha without a grid would show up
https://github.com/user-attachments/assets/7eac6411-2216-426e-a0bb-7a3bd28432d7
https://github.com/user-attachments/assets/151e23cf-f8e8-4f99-b909-ad063f530fbf
https://github.com/user-attachments/assets/39d7893c-4201-4aa0-9cf8-b161856b0719
weird case that ive had twice so far - initial guess is incorrect but when I press solve it gets it correctly
https://github.com/user-attachments/assets/421c0fe1-067b-446d-88f6-5ebf79ebd9ff
I think I have fixed the slider. The noise still fucks up the solver occasionally though...
No update found. 1.4.12
Last update, 15h ago.
a4dfb10 18 hours ago
@JonseyJones still testing
Okay, I think this one should work fine eeaae13ee8e569bbd969ef290205f7664cb935d0
I made some changes to the slider scoring value, it's working fine
Just got this on a fresh proxy Second try, same results
Though I suspect that in the next few days or weeks they are gonna go live with this one #28
🙏
chopped off characters in bg