>>28306 cunnyfication is based

>>28284 If you use a less trained version of Noob instead of 1.0, the Danbooru information shines through better. So basically, it's better at Japanese artists, worse at western e621 artists. I was having good results with Epsilon 0.75 but even lower might be better, haven't tried those. The thing to keep in mind is that more training doesn't necessarily mean better, the Danbooru information is getting diluted by the e621 information. Whether that is a good thing or not depends on how much you like e621 shit, most of it is garbage in my opinion.

>>28307 truly

>>28308 I actually was mostly talking about vpred versions. I only tried 1.0 when it comes to eps and even there artist tags felt kinda stronger compared to vpred... I'm satisfied with vpred+lora for now though, I like the benefits it provides despite all the jank. Maybe I should try pure illustrious or eps0.5 some time though.

why are furfags so fucking worthless? why do they need to give out their opinion or even simply reply to EVERY single question ESPECIALLY when they know NOTHING about the subject? both the illustrious and anlatan discords are FULL of them and they all act the fucking same, you can't ask ANYTHING without them butting in with their worthless fucking non-replies AND THEY GO ON AND ON AND ON, they literally can't stop blabbering once they start replying holy FUCK total furfag death can't come soon enough

>>28308 >the Danbooru information is getting diluted by the e621 information is this how it works? shouldnt sdxl be able to handle 12 million images without getting "diluted"

>>28312 It's a 6GB "brain." That's not enough to store 12 million images' worth of data without losing massive amounts of information. Think about it this way, if you have a dataset of exactly 1 picture, the only thing the model will be able to produce is that one picture, but it will reproduce it perfectly. Now extrapolate that to an entire website. A dataset of only Danbooru will be able to reproduce the "soul" of Danbooru perfectly. A dataset of only e621 will reproduce the "soul" of e621 perfectly. A dataset of both will be half as good at both. I don't know what size of "brain" you would need to handle both with no issue, but probably way, way bigger than 6GB. A dataset dump of Danbooru alone is many TB of pictures, not GB. As more evidence of the limits of SDXL, there's not a single checkpoint that can handle both 2D and realism, so western VS eastern art is probably a similar dilemma, although the gap is maybe not as wide.

>>28238 はい. 私はすでにそれをやっています

>>28313 all in how you handle it

v4 preview is out, a chaste model like last time. character segmentation feels like it's about 20% on the mark vs like 3% for all previous models. artist tags are still in. unsurprisingly feels underbaked, probably not worth paying for presently.

>>28316 this is how the segmentation is done, a universal prompt at the top, boxes for character stuff, and a five by five x/y grid for roughly where they should be

Kinda hard to get shit going for v4 considering they probably pruned everything nsfw for the preview

>>28317 So a static grid regional prompter basically?

>>28319 if you choose to, it's optional

>>28316 >v4 preview is out, a chaste model like last time did they have v3 preview? was it considerably worse than release v3? the update blog only mentions v2 and then v3 release after that

>>28321 v1 had a curated version way back in the day that was also bad. pretty much same scenario

>>28322 compared to the full version curated v1 was pretty good (at least in terms of baseline quality) i still unironically think it would've been better to use that as a base model and train the NSFW concepts back in

Haven't prompted in a while, what's the latest thing we're using now? Is it still A111? Forge? Comfy?

>>28327 i'm using reforge, haven't cared to check what forge is up to for a while

>>28328 Where do I get that?

>>28329 https://github.com/Panchovix/stable-diffusion-webui-reForge

>>28330 Thanks

try invoke, it's really good no they don't pay me no i don't use local no i don't mean the paid service

>>28332 Any reasons why you think it's good?

>>28333 it's not gradio, you don't need to switch tabs constantly, a lot of auto extensions are built-in, you've got nodes to waste time with in case you're one of those dunning-kruger types, the inpainting is better than auto and the tiled canvas+regional guidance are much better bandaids than auto/comfy regional extensions i haven't used it in a long while because i don't use local anymore, i shill it just because i think people are really sleeping on it and it could lead to better/faster/less frustrating workflows yet nobody will even give it a try because "muh ironic anti-corpo time is money i can't waste 2 mins clicking through an installer and 3 mins learning how shit works" shitposters think they're switching from 3dsmax to blender

>>28334 telemetry status?

>>28335 disabled via launch arguments

>>28332 why do people even get triggered by invoke suggestions i tried it some time ago and i think its indeed good, im just more used to a1111 shit and went back to reforge, but invoke honestly looks better thought out to me

>>28337 it seems alright but i bounced off it earlier for lack of obvious equivalents for shit i like i'll try it again tomorrow

>>28337 pick one or more a) shitposters b) (see above) ui consolewar faggottry (moot point because invoke switched to nodes quite some time ago) c) "but i dont want to learn a new ui" (also moot for obvious reasons) d) they think people mean the paid service or that you have to pay (heresy) for the ui or something e) it lacks some niche extension they use (fair although the major/important ones are built-in) d) they think it's slower (probably is?, i doubt it's as fast/optimized as reforge) one shitposter in particular really dislikes the emphasis syntax (appending +/– instead of parentheses) but i find it way more convenient, even using the auto shortcut you still have do highlight manually/by double-clicking and then you have to use two more keys, this way you only have to click and smash +/- my only gripe with it is that it still doesn't have NAI and autocomplete extensions, if it did i would delete my old auto install (which i only use for the NAI extension) in a heartbeat

>>28339 (i'm subscribed to their yt channel and every time i watch their update videos i seethe at how bad the gradio uis and NAI site are)

>>28338 whoa nice

I'm back, what did I miss >check /hdg/, picrel Looks like I didn't miss a thing

>>28342 >Looks like I didn't miss a thing Yes, you are correct.

>>28308 >>28313 care to quantify any of these wild speculations? This sounds like another "fur suit eyes" situation

>>28344 >source??? read up on sdxl maybe?